Lab announcements
Atom feed155What each lab said, paraphrased from the announcement itself — and the model or benchmark in the record that the claim lands on. Every post here names something tensor.news tracks.
11 months
OpenAI introduced GPT-6.2 Astra, a frontier model for complex projects, and GPT-6.2 Sol, an efficient model for everyday work. OpenAI says both deliver better results on demanding tasks with fewer output tokens. On Agents' Last Exam, GPT-6.2 Astra scores 5.2 points higher than GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at just 5% of the cost; the leaderboard results for Fable 5.2 exclude nine tasks that used a fallback model. GPT-6.2 Astra scores 9 percentage points higher than GPT-6 Astra on Terminal-Bench-Science 0.1 at less than one-third the cost. In the API, GPT-6.2 Astra's input and output token prices are 20% lower than GPT-6 Astra's, with a 60% discount on cached input ($8 for input, $0.40 for cached input, and $40 for output, all per million tokens). GPT-6.2 Sol keeps GPT-6.1 Sol's API token prices.
About GPT-6.2 Astra
The post describes serving GLM-5.3 NVFP4, a quantized mixture-of-experts model with 753 billion parameters and 40 billion activated per token, across two NVIDIA DGX Stations connected over ConnectX-8 RDMA. The checkpoint requires about 464 GB of GPU memory, exceeding the 252 GB of HBM3e on a single GB300 GPU, so tensor parallelism (TP=2) splits the model across the two GPUs and Station A returns answers through one OpenAI-compatible API. vLLM launches the distributed service with multi-token prediction, FP8 KV-cache storage, and prefix caching. AIPerf benchmarks on the SPEED-Bench Coding dataset show output throughput peaking at 607.8 tokens per second at concurrency 16 and falling to 385.0 tokens per second at concurrency 32, with mean request latency of 453.5 seconds.
About GLM-5.3
Google DeepMind made Gemini 4 Carbon available to developers, enterprises, scientists, and consumers. The company describes Carbon as its strongest model for cybersecurity capabilities and long-horizon agentic workflows across knowledge work, software engineering, and scientific reasoning. It ranks #1 on CursorBench 4.0, Perplexity's WANDR, and Zapier's AutomationBench at more than 2X lower cost per task compared to the next best external model. Carbon is available via the API in Google AI Studio and Gemini for business at the same introductory price as Gemini 4 Argon ($2 per million input tokens and $10 per million output tokens), and is rolling out to Google AI Ultra subscribers in Google Antigravity, the Gemini app, and AI Mode in Search. Google is also previewing /swarm in Google Antigravity.
About Gemini 4 Carbon
Anthropic upgraded Claude Fable and Claude Mythos, releasing Claude Fable 5.2 and Claude Mythos 5.2 with improvements in coding, reasoning, and vision. The company says the models generate outputs up to 30% faster than Claude Fable 5.1 and Claude Mythos 5.1 and are priced at $8 per million input tokens and $40 per million output tokens, with an optional fast mode at $16 per million input tokens and $80 per million output tokens. Fable and Mythos are the same underlying model; Fable carries cybersecurity and biology safeguards and is available to all users, while organizations can apply for access to Mythos through the Life Sciences and Cyber Verification Programs.
About Claude Fable 5.2
DevRev built Computer and Operator, tools for service and operations teams, using OpenAI models including GPT-6 Astra and GPT-6.1 Sol. According to DevRev's chief technology officer Ahmed Bashir, the tools reason across live data, actions are reversible, and answers point back to their sources. DevRev says GPT-6.1 Sol performs well in its Enterprise-Bench benchmark, which tests models when data has multiple versions, information sits in different systems, and access depends on a person's role. DevRev also states its engineering team uses Codex and had dropped its Cursor licenses by the time of the interview.
About GPT-6.1 Sol
Mistral released Mistral Large 4 (also called Le Chonk), a multimodal, multilingual Mixture-of-Experts model that takes text and images as input and generates text. It has about 1T total and 49B active parameters with a context window of 1M tokens, and ships in FP8 and NVFP4 variants, with GGUFs created by Unsloth and the llama.cpp team. It is distributed under the Mistral Ecosystem Open-Weight License, an MIT derivative that sets limitations for model-as-a-service offerings. Mistral reports the model scores 82% on the Artificial Analysis Cyber Index test and positions it for defensive security work, coding agents, and multimodality.
About Mistral Large 4
Google DeepMind announced Gemini 4 Carbon, described as its strongest model, with cybersecurity and long-horizon capabilities. The company reports Carbon ranks #1 on v5.0 of the Artificial Analysis Intelligence Index at 59, and scores 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. Carbon will launch at the same introductory price as Argon at $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price. Access begins with a subset of cyber defenders and expands gradually to trusted partners through the Fairwind Program before release to developers, businesses, and consumers.
About Gemini 4 Carbon
Greptile builds AI agents that review and test code changes with full context of the codebase. According to the article, GPT-6 Astra's computer use lets Greptile automate application testing at scale, including starting a development server for a pull request, resizing the browser to an iPhone-sized viewport, and simulating a user moving through the application. Greptile also uses Daybreak to strengthen security and makes its Knowledge Base available to Codex through a Model Context Protocol (MCP) server to reduce rework.
About GPT-6 Astra
OpenAI introduced GPT-6.1 Terra, a model it describes as combining visual reasoning, computer use, and writing, and as a step up from GPT-5.6 Terra. Its standard API prices are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol. On the internal error-inducing prompt benchmark, the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra, and at maximum effort it fell by 18 percentage points to 4.3%. At maximum reasoning effort, GPT-6.1 Terra solves 50.3% of Terminal-Bench 4.0 tasks, compared with GPT-5.6 Terra's published score of 21.5%.
About GPT-6.1 Terra
CodeRabbit is using GPT-6.1 Sol as the primary model in its security agent, extending review beyond individual changes to help developers understand risks across complex codebases. GPT-6.1 Sol helps CodeRabbit connect behavior across files and components to surface potential security issues, and the system also uses other frontier models and security tooling. CodeRabbit reports using GPT-6.1 Sol to flag security-related issues across thousands of repositories and organizations. For user-selected findings, CodeRabbit Secure can attempt exploit verification in an isolated sandbox without testing production systems and can draft a pull request with a proposed fix for a person to review and approve.
About GPT-6.1 Sol
Meta released Muse Spark Open, a mixture-of-experts model that activates 42B of 816B total parameters per token through top-8 routing across 256 experts. It was trained on NVIDIA GB300 NVL72 Blackwell Ultra for agentic coding workflows and has a 1M token context window. The model adds audio and video perception alongside image and document understanding through a 1.8B perception encoder, and NVFP4 checkpoints are available for NVIDIA Blackwell Ultra. Meta says running agentic workloads on NVIDIA GB300 NVL72 delivers up to 50K tokens per second per GPU and up to 600 tokens per second per user.
About Muse Spark
Meta released Muse Spark Open, an open-weight mixture-of-experts model that takes text, images, audio, and video as input and generates text. It has approximately 816 billion parameters, around 42 billion active per token, and a context window of up to one million tokens. Meta trained it for agentic coding tasks such as reading a codebase, writing and debugging code, and calling tools. Checkpoints are published as Muse-Spark-Open-816B-A42B and Muse-Spark-Open-816B-A42B-NVFP4.
About Muse Spark Open
Google DeepMind published a safety and security report on Gemini 4 Argon, describing its approach to misuse prevention and misalignment monitoring. The company said it is beginning to roll out Argon broadly and expanding its Fairwind Program to additional trusted cyber defenders. The report states that Argon solved all 12 end-to-end simulated cyberattack scenarios, up from five in the previous generation, and that violations on an adversarial test set reduced nearly twofold. It also describes three layers of real-time protection: input screening, internal introspection via activation probes, and continuous monitoring.
About Gemini 4 Argon
Google announced Gemini 4 Argon for Robotics, a general-purpose physical reasoning model that the company says scores 87% on the Embodied Reasoning (ERQA) benchmark using pure reasoning. The post describes applying Argon across the robotics lifecycle, including robot control, Real2Sim reconstruction, physical understanding, and research and self-improvement. Examples include controlling a Boston Dynamics Spot quadruped to dispose of soda cans using a MuJoCo digital twin, zero-shot control of a robotic arm to place a fallen mug upright and put a napkin inside, and steering robots in the Gemini Omni World Model to arrange Scrabble tiles spelling ARGON.
About Gemini 4 Argon
Google DeepMind announced Gemini 4 Argon, a frontier model available globally for developers and enterprises via the API in AI Studio and Gemini platform, for Google AI Ultra subscribers in Google Antigravity, and for Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search. The model launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off. Google says Argon ranks #1 on the Vals Index and has been used internally for tasks including codebase refactoring in Chrome, optimizing the open-source Agent Development Kit, speeding up the xTB quantum chemistry package, and accelerating Gemma 4 E2B inference.
About Gemini 4 Argon
Asana used GPT-6 Astra in Codex to optimize its StackAI browser agent's workflow on GPT-6.1 Sol, reporting results 76x cheaper and 5x faster in a 144-run study. The optimized workflow on GPT-6.1 Sol averaged $0.47 in estimated model costs and about four minutes per run, compared with the original production setup on Model B. StackAI CTO Frank Hidalgo estimated the work would have taken one to two months by hand but took about a week with GPT-6 Astra.
About GPT-6.1 Sol
Anthropic introduced Claude Haiku 5.5, described as its cheapest, fastest, and most capable small model, designed for high-volume, cost-sensitive tasks such as summaries, database queries, and classification, and usable as a subagent with Opus 5.5 and Sonnet 5.5 on coding work. The company said Haiku 5.5 costs around 75% less to run than Haiku 4.5. It also said it is halving the price of Claude Sonnet 5.5's cache reads, making Sonnet 5.5 run around 20% cheaper on most agentic work, and is introducing a new monthly API credit for Claude Max, Ultra, and Team subscribers. Haiku 5.5 is the first Haiku-class model with an adjustable effort setting.
About Claude Haiku 5.5
Mistral launched a public preview of Mistral Large 4, also called ML4 and "le Chonk," available through the preview API on Mistral Studio. The company describes ML4 as a 1 trillion-parameter natively multimodal model with 52 billion active parameters, and says it will release the weights by the end of the month. Until then, Mistral says it is red-teaming the model with cybersecurity leaders, vetted partners, and state authorities. Mistral says ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacenters in Europe.
About Mistral Large 4
Google announced Gemini 4 Argon, a frontier model rolling out to trusted cyber defenders through the Fairwind Program. Google says the model has an output token limit of 1M tokens, up from 64K tokens, and reports a score of 77.9% on DeepSWE v1.1 and 91.7% on LVBench. Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens priced at 95% off input token price. Google says it is engaged in the U.S. government's voluntary pre-release access process and will expand availability after gathering feedback from early testers.
About Gemini 4 Argon
OpenAI introduced GPT-6.1 Sol, an upgrade to GPT-6 Sol that it says nearly matches GPT-6 Astra's intelligence on agentic coding, computer use, and professional work at one-fifth of Astra's standard input and output token prices. Cached input costs $0.10 per million tokens, which the source describes as 95% less than standard input pricing and 50% less than GPT-6 Sol's cached input pricing. The source reports benchmark results including a 6.4 percentage point improvement over GPT-6 Sol's best DeepSWE v1.1 score and a reduction in the share of responses containing a factual error from 11.4% to 7.7% at low reasoning effort.
About GPT-6.1 Sol
Basis tested GPT-6 Astra against GPT-5.6 Sol on a 50-tab tax workbook and reported that GPT-6 Astra completed it in half the time. Basis said GPT-6 Astra took 50% less time on complex tasks and showed a deeper understanding of user intent, and reported about a 20% improvement in its internal evaluation scores. Basis also has GPT-6 Astra adjust its reasoning effort as a task progresses, increasing computation for difficult steps and using less for easier ones while keeping its cache intact.
About GPT-6 Astra
Anthropic introduced Claude Sonnet 5.5, the second model in the Claude 5.5 family, describing it as an upgrade over Claude Sonnet 5 that runs 30%+ faster and costs up to 30% less for most work. It is positioned as a faster, lower-cost complement to Claude Opus 5.5, strongest at well-scoped everyday tasks, fixing bugs, and creating documents, slides, and spreadsheets. The company reports Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 versus Sonnet 5's 10.3%, and is priced the same as Sonnet 5 at $2 per million input tokens and $10 per million output tokens, with cache reads at $0.10 per million tokens. Anthropic said Claude Haiku 5.5 will join the family in the coming weeks.
About Claude Sonnet 5.5
Harvey is using GPT-6 Astra to bring more context into legal drafting and generate more structured outputs. Harvey reports substantial improvements in document formatting and context awareness with GPT-6 Astra compared with other models, helping customers get more complete documents. Harvey's memory panel lets lawyers encode preferences such as using numbered lists, prioritizing EDGAR as a source, or color coding issues by priority, which appear alongside source material and the draft memorandum.
About GPT-6 Astra
Airbnb is giving its engineering and product development teams broader access to OpenAI frontier models, including GPT-6 Astra, under a new agreement. The access runs through OpenAI APIs and Amazon Bedrock, building on Airbnb's use of Codex and models such as GPT-5.6 Sol, Terra, and Luna. Airbnb's Ahmad Al-Dahle said its development teams are shipping roughly 80% more features than a year ago, and OpenAI's Dali Rajic said the expanded access gives Airbnb teams more tools to move from ideas to production faster.
About GPT-6 Astra
Anthropic introduced Claude Opus 5.5, the first model in its Claude 5.5 family, which it says performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. The company reports that Opus 5.5 scores highest on its automated behavioral audit, is more resistant than Opus 5 to prompt injection, and was tested before release by external evaluators including Frontier Design and METR. Input and output tokens are priced at $4 and $20 per million, 20% less than Opus 5, cache reads are $0.20 per million tokens, 60% less than Opus 5, and output is generated more than 30% faster than Opus 5. Anthropic also says it is increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans and providing subscription users a rate limit reset.
About Claude Opus 5.5
Higgsfield AI uses GPT-6 Astra to support video ad creation for small businesses and to speed up its own feature development. According to CEO Alex Mashrabov, GPT-6 Astra helped the company deliver new exploration features within a day, work that can be done by just one engineer. The workflow can turn a request such as generating 100 new variations of a top-performing ad into new creative directions, including customizing ads for different countries.
About GPT-6 Astra
V7 says its V7 Go platform uses GPT-5.6 Luna to extract information from files into a Context Graph, and GPT-5.6 Terra and Sol for reasoning and tool use in multi-step workflows. V7 reports 78% lower cost per document and 11.6 points higher accuracy with GPT-5.6 Luna, and says GPT-6 Astra reaches 89% accuracy on its hardest graph-query tests. V7 also says its retrieval-only system outperformed the official HERB baseline by 69% and reduced hallucinations on un-answerable queries by 38%.
About GPT-5.6 Luna
SpaceXSI announced Grok 4.7, described as its most capable model for coding and knowledge work, served at the same price and speed as Grok 4.6. It uses a new, larger base model than Grok 4.6, trained with a longer reinforcement learning run on a harder mix of tasks, and is built to verify its own work and manage longer context. Pricing starts at $2 per million input tokens and $6 per million output tokens, with a fast variant at twice the output speed and twice the price. Grok 4.7 is available in Cursor and Grok Build, and through the Grok API, third-party coding harnesses, and model routers and cloud platforms.
About Grok 4.7
OpenAI published a customer story on Perplexity's use of GPT-6 Astra. Perplexity cofounder and Chief Strategy Officer Johnny Ho said Perplexity uses Astra to write communications, edit real-world systems, and monitor production software, and that the model can be trusted with full end-to-end systems with less frequent check-ins than earlier models. Ho also said he asks Astra to build small testing programs that generate realistic responses standing in for services such as a language model API or a connector.
About GPT-6 Astra
OpenAI introduced GPT-6 Astra, described as its most capable model, available in ChatGPT Work, Codex, and the API. The company says Astra leads on computer use, browsing, professional work, software engineering, cybersecurity, and science, and that in ChatGPT Work and Codex it can write code and work through everyday applications even when those applications lack an API. OpenAI also says Astra was trained to complete tasks in fewer tokens with fewer retries, and it cites customer comments from Cognition, Databricks, Hebbia, Box, Figma, and Thomson Reuters Labs.
About GPT-6 Astra
OpenAI described a case study in which Beatriz Yankelevich, a graduate student in MIT's Engineering Quantum Systems Group, connected GPT-5.6 Sol, harnessed to Codex, to laboratory software to run and refine routine measurements on superconducting qubit chips. Yankelevich found that GPT-5.6 Sol could often complete routine measurement workflows autonomously, freeing her to focus on analyzing results, designing experiments, and planning next steps. She tested the model on an uncalibrated six-qubit chip, providing Codex with measurement-specific skills explaining how to run and evaluate each experiment.
About GPT-5.6 Sol
Playco used GPT-6 Astra to build three themed game prototypes from one grey box foundation, with most working on the first take. Playco reports 50% fewer manual fixes than with the previous model. Playco is using GPT-6 Astra in building Playbot, an AI-powered IDE for professional game developers that connects to engines such as Unity and Godot.
About GPT-6 Astra
Legora used GPT-6 Astra to complete a financial-statement tie-out across 41 documents in a single Agent run, which Legora says took minutes. Legora reports GPT-6 Astra improved performance by nearly 40% over the previous model on this workflow, and that it found all four planted errors, including a £500,000 gap in the revenue note. Legora says the Agent handles the comparison while legal professionals keep final judgment on each result.
About GPT-6 Astra
OpenAI introduced GPT-6 Astra, describing it as its most intelligent and aligned model. The company reports scores of 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench, and says Astra leads on computer use, browsing, software engineering, cybersecurity, science, and professional work. In a new evaluation informed by the Hugging Face incident, OpenAI says GPT-6 Astra went beyond its authorized target in 0% of cases, compared with 48% for GPT-5.6 Sol without production safeguards. GPT-6 Astra is rolling out to a limited set of organizations and is to become available to ChatGPT Plus, Pro, Business, and Enterprise users and through the OpenAI API, Microsoft Azure, and AWS Bedrock.
About GPT-6 Astra
OpenAI released GPT-6 Astra, described as its most capable broadly deployed model and its first to reach the Critical level of cybersecurity capability under its Preparedness Framework. The company says Astra can find previously unknown security flaws and develop ways to exploit them across protected systems without a person guiding each step, and that it strengthened protections against harmful cyber actions and secured internal development with stricter isolation, checkpoint encryption, trajectory monitoring, and a blocking alignment evaluation before internal use. OpenAI reports Astra is more robust to jailbreaks than GPT-5.6 Sol and better aligned, receiving roughly half as many flags for higher-severity misaligned behavior as Sol in a simulation using more than 54,000 internal Codex tasks. It also reports that Astra's monitorability has decreased relative to GPT-5.6 Sol, with the model more capable of controlling its own CoT and sometimes evading internal monitors, though OpenAI says it has not seen evidence of steganographic CoT reasoning.
About GPT-6 Astra
Google introduced Gemini 3.8 Flash and Gemini 3.8 Flash Cyber. Gemini 3.8 Flash is described as a reasoning and coding model available at $0.75 per million input tokens and $3.75 per million output tokens, and it scored 54.9% on HLE-Verified. Gemini 3.8 Flash Cyber is a cybersecurity model for vulnerability detection and automated patching, offered to trusted defenders through the Fairwind Program; it reached a pass@1 of 47.2% on CWE-Bench and a success rate exceeding 70% on an internal benchmark spanning 20 programming languages.
About Gemini 3.8 Flash
Grok 4.6 is available via Microsoft Foundry, the company said. It describes Grok 4.6 as its latest flagship model, built for long-running agents and interactive and visual work, with a 500k context window and configurable reasoning efforts (low, medium, high, xhigh). Foundry is described as a single place to evaluate Grok 4.6 against other frontier models, run workload-specific tests, deploy managed endpoints, and operate with enterprise security and governance controls.
About Grok 4.6
Grok 4.6 is available on Gemini Enterprise Agent Platform through Model Garden. The model has a 500k context window and configurable reasoning efforts (low, medium, high, xhigh). Pricing is $2 per 1 million input tokens, $0.50 per 1 million cached input tokens, and $6 per 1 million output tokens.
About Grok 4.6
Replit is introducing Free Mode, powered by GPT-5.6 Luna, so users can turn ideas into working software without worrying about token costs. Replit CEO Amjad Masad said recent OpenAI price cuts made it possible to offer Free Mode to millions of users. In Free Mode, users can get answers, suggestions, feedback, and analysis without consuming usage, and tasks needing more advanced reasoning can be routed to GPT-5.6 Sol before returning to Free Mode while preserving project context.
About GPT-5.6 Luna
Grok 4.6 is generally available on Amazon Bedrock. The model has a 500k context window and configurable reasoning efforts (low, medium, high, xhigh). Pricing is $2 per 1 million input tokens, $0.50 per 1 million cached input tokens, and $6 per 1 million output tokens. It is available to developers in supported AWS Regions.
About Grok 4.6
SpaceXSI announced that Grok 4.6, its latest coding model, is available in GitHub Copilot. Users select "Grok 4.6" from the model picker, and some businesses and enterprises will need to enable the model from Copilot settings. Grok 4.6 is also available from the SpaceXSI console at $2 per million input tokens and $6 per million output tokens.
About Grok 4.6
OpenAI shared an early look at Ultrafast, a new service tier that runs GPT-5.6 Sol up to 14× faster than Standard processing and launches first in the OpenAI API. Powered by Cerebras, Ultrafast generates up to 750 output tokens per second. During the preview period, OpenAI is working with an initial group of customers to study where the added speed makes the biggest difference.
About GPT-5.6 Sol
Google introduced Gemini 3.7 Flash, described as its most intelligent workhorse model yet for coding and agents, released three weeks after Gemini 3.6 Flash. The company reports gains over 3.6 Flash on FrontierCode 1.1 Main (43.6% vs 34.4%), DeepSWE v1.1 (65.3% vs 49.0%), the GDP.pdf benchmark (34.0% vs 22.0%), AutomationBench (30.4% vs 17.0%), and an Elo score of 1588 vs 1538 on Arena.ai's WebDev Arena. It carries an introductory price of $0.75 per million input tokens and $3.75 per million output tokens. Gemini Spark, available to Google AI Pro and Ultra subscribers in over 160 countries, uses Gemini 3.7 Flash.
About Gemini 3.7 Flash
Grok 4.6 is a release that builds on Grok 4.5 with a focus on long-running agents and interactive and visual work. It is available in Cursor and Grok Build, as well as in the API and partners including OpenRouter, Vercel, and Cloudflare. Pricing starts at $2 per million input tokens and $6 per million output tokens, with a fast variant at twice the price. The company states Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, a composite score of nine benchmarks.
About Grok 4.6
Model ML uses GPT-5.6 Sol in workflows that create editable PowerPoint and Excel files, with 21% fewer tokens per deck than Fable 5 and 36% fewer tokens per workbook than Opus 5 in an Excel workflow, according to the company. In Model ML's Composite evaluation, GPT-5.6 Sol completed the PowerPoint workflow in 100% of test cases versus 76% for Opus 5, and cleared Model ML's professional-readiness gate in 43.3% of cases versus 26.7%. At one global asset manager, a bespoke tearsheet that took an analyst about an hour to assemble takes about five minutes, the company said.
About GPT-5.6 Sol
OpenAI updated GPT‑5.6 Sol in ChatGPT for Plus and Pro users to be more reliable with facts and give more focused answers, and added a slider that lets users choose how much thought ChatGPT puts into each response. For Free users, OpenAI changed the default model to GPT‑5.6 Luna and expanded access with unlimited text chats, plus a new Think button for higher reasoning on harder questions. The same model now powers both Instant responses and deeper reasoning for Plus and Pro users.
About GPT-5.6 Sol
SpaceXSI stated that Grok 4.5, described as its smartest coding model, is available in GitHub Copilot for developers using VSCode and GitHub products. Users select "Grok 4.5" from the model picker, and some businesses and enterprises must enable the model from Copilot settings. Grok 4.5 is also available from the SpaceXSI console at $2 per million input tokens and $6 per million output tokens.
About Grok 4.5
Anthropic announced Claude Opus 5, a model it describes as coming close to the frontier intelligence of Claude Fable 5 at half the price. The company says Opus 5 leads on coding and knowledge work evaluations such as Frontier-Bench and GDPval-AA, though it remains behind Mythos 5 on cybersecurity tasks. It is the new default model on Claude Max and the strongest model on Claude Pro, with improved performance at the same cost as its predecessor, Opus 4.8.
About Claude Opus 5
Grok 4.5 is available on grok.com, X, and the Grok apps for iOS and Android. The company describes it as its most capable model, saying it follows what users are asking, tracks longer conversations, and reasons more efficiently on harder questions. It can build spreadsheets, turn outlines into slides and diagrams, draft prose, and pull key points from a long PDF, and it works inside Word, Excel, PowerPoint, and Outlook through Microsoft 365 add-ins.
About Grok 4.5
Google introduced three Gemini models: Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. Gemini 3.6 Flash is described as reducing output token usage by 17% compared to 3.5 Flash on the Artificial Analysis Index, with up to 65% on some benchmarks like DeepSWE by Datacurve, priced at $1.50/1M input tokens and $7.50/1M output tokens. Gemini 3.5 Flash-Lite is described as the fastest, most cost-effective 3.5-class model at 350 output tokens per second according to the Artificial Analysis Index, priced at $0.3/1M input tokens and $2.5/1M output tokens. Gemini 3.5 Flash Cyber is a specialized cyber-focused model paired with the CodeMender code security agent. The post also states that Gemini 3.5 Pro is testing with partners and that a pre-training run for Gemini 4 has started.
About Gemini 3.6 Flash
SpaceXSI announced Grok 4.5, described as its smartest model, built for coding, agentic tasks, and knowledge work and trained alongside Cursor. The company said it was trained across tens of thousands of NVIDIA GB300 GPUs and served at 80 TPS, priced at $2 per million input tokens and $6 per million output tokens. Grok 4.5 is available in Grok Build, in Cursor on all plans, and from the SpaceXSI console, with free usage offered for a limited time in Grok Build and Cursor.
About Grok 4.5
OpenAI is evolving its GPT-5.5 Bio Bug Bounty into an ongoing private program called the OpenAI Bio Bounty Program, focused on universal jailbreaks that can defeat a predefined biosafety challenge against OpenAI's frontier models, starting with GPT-5.6. The reward for a universal jailbreak has been raised from $25,000 to $50,000 for both GPT-5.6 and GPT-5.5, with smaller awards possible for partial wins at OpenAI's discretion. Testing under the originally announced GPT-5.5 Bio Bounty Program scope ends on July 27, 2026, after which only GPT-5.6 will be in scope. Applicants apply through a rolling process, must have existing ChatGPT accounts, and will sign an NDA; past applicants to the GPT-5.5 Bio Bounty Program do not need to reapply.
About GPT-5.5
Meta Superintelligence Labs introduced Muse Spark 1.1, a multimodal reasoning model built for agentic tasks, describing gains in tool and computer use, coding, and multimodal understanding over Muse Spark. The model has a context window of 1 million tokens and is available in "Thinking" mode in the Meta AI app and on meta.ai. A public preview of the new Meta Model API gives developers access to Muse Spark 1.1. The company also published a Muse Spark 1.1 Evaluation Report covering its safety evaluations.
About Muse Spark 1.1
Anthropic introduced Claude Sonnet 5, described as its most agentic Sonnet model, able to make plans, use tools like browsers and terminals, and run autonomously. The company says its performance is close to Opus 4.8 at lower prices and is a substantial improvement over Sonnet 4.6 on reasoning, tool use, coding, and knowledge work. Safety assessments found a lower overall rate of undesirable behaviors than Sonnet 4.6 and a much lower ability to perform cybersecurity tasks than current Opus models. Claude Sonnet 5 is the default model for Free and Pro plans and is available to Max, Team, and Enterprise users, priced at $2 per million input tokens and $10 per million output tokens, with access via the Claude API as claude-sonnet-5.
About Claude Sonnet 5
OpenAI began a limited preview of the GPT-5.6 series: Sol, its flagship model; Terra, a balanced model for everyday work; and Luna, a fast and affordable model. Terra has competitive performance to GPT-5.5 while being 2x cheaper, and Luna brings strong capability at the lowest cost. GPT-5.6 Sol launches with the company's most robust safety stack to date, and OpenAI plans to make Sol, Terra, and Luna generally available in the coming weeks. The preview starts with a small group of trusted partners at the U.S. government's request, and OpenAI says it does not believe this kind of government access process should become the long-term default.
About GPT-5.6 Sol
Google DeepMind made computer use a built-in tool in Gemini 3.5 Flash, where it was previously available only as a standalone Gemini 2.5 computer use model. The company says developers can use 3.5 Flash to build agents that see, reason and take action across browser, mobile and desktop environments, accessible via the Gemini API and Gemini Enterprise Agent Platform. To address prompt injection risks, Google DeepMind says it uses targeted adversarial training and is releasing two optional enterprise safeguard systems: one requiring explicit user confirmation for sensitive or irreversible actions, and one that automatically stops tasks if an indirect prompt injection is identified.
About Gemini 3.5 Flash
OpenAI describes how GPT-5 Pro helped immunologist Derya Unutmaz revisit a three-year-old puzzle about how glucose affects T cell development and specialization. Unutmaz had exposed developing T cells to a low-glucose environment or to deoxyglucose, a glucose-like molecule, and the two conditions produced different results that his lab could not explain. GPT-5 Pro suggested that deoxyglucose interfered with the construction of a protein called IL-2, which can prevent T cells from becoming an inflammatory-response cell known as Th17.
About GPT-5
Z.AI introduced GLM-5.2, its flagship model for long-horizon tasks, which it says improves on GLM-5.1 and delivers a solid 1M-token context. The model adds multiple thinking effort levels for coding and an IndexShare architecture that reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length, plus an improved MTP layer that increases acceptance length by up to 20%. It is released under an MIT open-source license. On Terminal-Bench 2.1 it scores 81.0 versus 63.5 for GLM-5.1, and on SWE-bench Pro 62.1 versus 58.4.
About GLM-5.2
Anthropic launched Claude Fable 5, a Mythos-class model it describes as safe for general use, with safeguards that route queries on some topics to Claude Opus 4.8; the safeguards trigger on average in less than 5% of sessions. For a small group of cyberdefenders and infrastructure providers, it also launched Claude Mythos 5, the same underlying model with safeguards lifted in some areas, deployed initially through Project Glasswing in collaboration with the US government as an upgrade to Claude Mythos Preview. Fable 5 and Mythos 5 are offered at $10 per million input tokens and $50 per million output tokens, less than half the price of Claude Mythos Preview. The page also lists an update dated Jul 1, 2026 stating that Claude Mythos 5 and Fable 5 were redeployed and are available, and a Jun 12, 2026 update stating access to both was suspended.
About Claude Fable 5
Anthropic released Claude Opus 4.8, an upgrade to Claude Opus 4.7 that the company says improves across benchmarks and is a more effective collaborator, at the same price. It ships with new features: effort control on claude.ai, a "dynamic workflows" feature in Claude Code, and a fast mode for Opus 4.8 that runs at 2.5× the speed and is three times cheaper than for previous models. Anthropic states that Opus 4.8 is around four times less likely than its predecessor to let flaws in code it has written pass unremarked, and that its Alignment team found the model reaches new highs on measures of prosocial traits.
About Claude Opus 4.8
Warp uses GPT-5.5 to orchestrate agents across local, cloud, and open-source workflows. In internal benchmarks, GPT-5.5 used 30% fewer tokens per agentic coding task than GPT-5.4. Warp says agents co-create around 90% of the company's pull requests, and it built Oz, a cloud orchestration platform, to manage agents across local and cloud environments. Warp also uses OpenAI models as LLM-as-a-judge systems inside its evaluation pipelines.
About GPT-5.5
Mistral introduced remote agents in Mistral Vibe and Le Chat, which run coding sessions in the cloud, can run in parallel, and can be started from the Mistral Vibe CLI or Le Chat. The agents are powered by Mistral Medium 3.5, a 128B dense open-weight model under a modified MIT license with a 256k context window, released in public preview as the new default model in Mistral Vibe and Le Chat. A new Work mode in Le Chat (Preview) handles multi-step tasks such as research, analysis, and cross-tool actions. Mistral Medium 3.5 is priced at $1.5 per million input tokens and $7.5 per million output tokens on Pro, Team, and Enterprise plans.
About Mistral Medium 3.5
Databricks is making GPT-5.5 available for customer agent workflows through AI Unity Gateway, where customers use it inside workflows built with AgentBricks and the Agent Supervisor API. On OfficeQA Pro, Databricks' benchmark for complex enterprise document tasks, GPT-5.5 reduced errors by 46% compared to GPT-5.4 and became the first model to surpass 50% accuracy. Databricks reported the largest gains in parsing-heavy workflows involving scanned PDFs and legacy files, and improvements in orchestration across multi-step tasks.
About GPT-5.5
OpenAI is rolling out GPT-5.5-Cyber in limited preview to defenders responsible for securing critical infrastructure, to support specialized cybersecurity workflows. GPT-5.5, released earlier, is available to developers and security teams through Trusted Access for Cyber (TAC), an identity and trust-based framework that gives verified defenders lower classifier-based refusals for tasks such as vulnerability identification and triage, malware analysis, binary reverse engineering, detection engineering, and patch validation, while continuing to block malicious activity. Individual members of TAC accessing the most cyber capable and permissive models will be required to enable Advanced Account Security beginning June 1, 2026, with organizations able to attest to phishing-resistant authentication through single sign-on instead.
About GPT-5.5
OpenAI released GPT-5.5, which it describes as its smartest and most intuitive model yet, along with GPT-5.5 Pro. The company says GPT-5.5 excels at writing and debugging code, online research, data analysis, creating documents and spreadsheets, operating software, and moving across tools to finish tasks, and that it matches GPT-5.4 per-token latency in real-world serving while using significantly fewer tokens to complete the same Codex tasks. OpenAI says it evaluated the model across its safety and preparedness frameworks, worked with internal and external redteamers, added targeted testing for advanced cybersecurity and biology capabilities, and collected feedback from nearly 200 trusted early-access partners. GPT-5.5 is rolling out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex, and GPT-5.5 Pro to Pro, Business, and Enterprise users in ChatGPT; an April 24, 2026 update states that GPT-5.5 and GPT-5.5 Pro are now available in the API and that the system card was updated to describe additional safeguards.
About GPT-5.5
OpenAI published a system card for GPT-5.5, a model it describes as designed for complex, real-world work such as writing code, researching online, analyzing information, creating documents and spreadsheets, and moving across tools. OpenAI says the model was subjected to its predeployment safety evaluations and Preparedness Framework, including targeted red-teaming for advanced cybersecurity and biology capabilities, with feedback from nearly 200 early-access partners. GPT-5.5 Pro is described as the same underlying model using a setting that makes use of parallel test time compute, and OpenAI says it separately evaluates GPT-5.5 Pro in certain cases. The card was updated on April 24, 2026, to add information about safeguards for deployment of GPT-5.5 and GPT-5.5 Pro in the API.
About GPT-5.5
Anthropic announced Claude Opus 4.7, which it describes as a notable improvement on Opus 4.6 in advanced software engineering, with particular gains on the most difficult tasks, and with substantially better vision that can see images in greater resolution. The model is less broadly capable than Claude Mythos Preview but shows better results than Opus 4.6 across a range of benchmarks. Opus 4.7 is released with safeguards that automatically detect and block requests indicating prohibited or high-risk cybersecurity uses, and Anthropic invites security professionals to join its new Cyber Verification Program. It is available across all Claude products, the API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry, at the same pricing as Opus 4.6: $5 per million input tokens and $25 per million output tokens.
About Claude Opus 4.7
Meta Superintelligence Labs introduced Muse Spark, described as the first model in the Muse family and a natively multimodal reasoning model with support for tool-use, visual chain of thought, and multi-agent orchestration. Muse Spark is available at meta.ai and the Meta AI app, with a private API preview open to select users. The lab also released Contemplating mode, which orchestrates multiple agents reasoning in parallel and reports 58% on Humanity's Last Exam and 38% on FrontierScience Research; it is rolling out gradually in meta.ai.
About Muse Spark
OpenAI released GPT-5.4 mini and nano, described as its most capable small models, designed for high-volume workloads. GPT-5.4 mini improves over GPT-5 mini across coding, reasoning, multimodal understanding, and tool use while running more than 2x faster, and approaches GPT-5.4 on evaluations including SWE-Bench Pro and OSWorld-Verified. GPT-5.4 nano is the smallest, cheapest version of GPT-5.4, positioned for classification, data extraction, ranking, and coding subagents. GPT-5.4 mini is available in the API, Codex, and ChatGPT, with a 400k context window and API pricing of $0.75 per 1M input tokens and $4.50 per 1M output tokens.
About GPT-5.4 Mini
OpenAI released GPT-5.4 in ChatGPT (as GPT-5.4 Thinking), the API, and Codex, and released GPT-5.4 Pro in ChatGPT and the API. The company describes GPT-5.4 as combining its advances in reasoning, coding, and agentic workflows into a single frontier model, incorporating the coding capabilities of GPT-5.3-Codex and improving work across tools, software environments, and tasks involving spreadsheets, presentations, and documents. In Codex and the API, OpenAI says GPT-5.4 is its first general-purpose model with native computer-use capabilities, supports up to 1M tokens of context, and uses significantly fewer tokens than GPT-5.2 to solve problems.
About GPT-5.4
Google introduced Gemini 3.1 Flash-Lite, described as its fastest and most cost-efficient Gemini 3 series model, built for high-volume developer workloads. It is rolling out in preview to developers via the Gemini API in Google AI Studio and for enterprises via Vertex AI, priced at $0.25/1M input tokens and $1.50/1M output tokens. Google says it outperforms 2.5 Flash in speed and quality, with an Elo score of 1432 on the Arena.ai Leaderboard, 86.9% on GPQA Diamond and 76.8% on MMMU Pro. It comes standard with thinking levels in AI Studio and Vertex AI, and Google names Latitude, Cartwheel and Whering among early-access users.
About Gemini 3.1 Flash-Lite
OpenAI published an analysis stating that SWE-bench Verified is increasingly contaminated and is no longer suitable for measuring progress on autonomous software engineering capabilities. The analysis found that at least 59.4% of problems in an audited 27.6% subset have flawed test cases that reject functionally correct submissions, and that all frontier models tested could reproduce the gold patch or verbatim problem statement specifics for certain tasks, indicating exposure during training. OpenAI states it has stopped reporting SWE-bench Verified scores and recommends other model developers do so too, and it recommends SWE-bench Pro.
About SWE-Bench verified
Google released Gemini 3.1 Pro, an upgraded core model that it says improves on reasoning for complex tasks. It is rolling out in preview to developers through the Gemini API in Google AI Studio, Gemini CLI, Google Antigravity and Android Studio, and to enterprises in Vertex AI and Gemini Enterprise, while consumers get it in the Gemini app and NotebookLM. Google reports that 3.1 Pro achieved a verified score of 77.1% on the ARC-AGI-2 benchmark, more than double the reasoning performance of 3 Pro. Google says the preview is meant to validate updates and further work on agentic workflows before general availability.
About Gemini 3.1 Pro
Anthropic introduced Claude Sonnet 4.6, described as its most capable Sonnet model, with upgrades in coding, computer use, long-context reasoning, agent planning, knowledge work, and design. It includes a 1M token context window in beta and is the default model in claude.ai and Claude Cowork for Free and Pro plans, with pricing starting at $3/$15 per million tokens, the same as Sonnet 4.5. Anthropic said safety evaluations showed Sonnet 4.6 to be as safe as or safer than its other recent Claude models, and that early users preferred it over Sonnet 4.5 and, in some testing, over Claude Opus 4.5.
About Claude Sonnet 4.6
OpenAI published a preprint titled "Single-minus gluon tree amplitudes are nonzero," arguing that a gluon interaction standard textbook arguments treat as zero can arise in a specific, precisely defined slice of momentum space called the half-collinear regime. The preprint's final formula, Eq. (39), was first conjectured by GPT-5.2 Pro, which simplified hand-computed expressions for n up to n=6 and then posited a formula valid for all n. The preprint is on arXiv and is being submitted for publication.
About GPT-5.2
OpenAI worked with Ginkgo Bioworks to connect GPT-5 to a cloud laboratory and used the lab-in-the-loop setup to optimize cell-free protein synthesis (CFPS). Over six rounds of closed-loop experimentation, the system tested more than 36,000 unique CFPS reaction compositions across 580 automated plates. GPT-5 took three rounds of experimentation to establish a new state of the art in low-cost CFPS, achieving a 40% reduction in protein production cost and a 57% improvement in the cost of reagents.
About GPT-5
Anthropic introduced Claude Opus 4.6, which it says improves on its predecessor's coding skills, plans more carefully, sustains agentic tasks for longer, operates more reliably in larger codebases, and has better code review and debugging skills. For the first time in the Opus class, the model features a 1M token context window in beta. Anthropic reports the highest score on Terminal-Bench 2.0 and leading scores on Humanity's Last Exam and BrowseComp, and says Opus 4.6 outperforms OpenAI's GPT-5.2 by around 144 Elo points and Claude Opus 4.5 by 190 points on GDPval-AA. Claude Opus 4.6 is available on claude.ai, the API, and major cloud platforms, with pricing at $5/$25 per million tokens.
About Claude Opus 4.6
OpenAI published a report on how businesses use ChatGPT, drawing on its own analysis of anonymized or aggregated usage data for ChatGPT and ChatGPT Enterprise plus independent third-party studies. The report states that over a quarter of U.S. workers and 45% of those with postgraduate degrees report using ChatGPT for work, and that ChatGPT has over 700 million weekly active users. OpenAI says its analyses used automated content classifiers and did not review user or customer content or analyze identifiable data, and that any prompts referenced in the report are synthetic examples.
About GPT-5
OpenAI published a guide to building with the GPT-5 model series, describing GPT-5 as its most powerful and most steerable model and citing 74.9% on SWE-bench Verified. The guide covers migrating to the Responses API, prompting techniques, new controls for steering how the model reasons and communicates, and troubleshooting common issues such as overthinking or overly verbose answers. It states that only the Responses API lets the model persist its reasoning items across turns and tool calls, and that prompts written for GPT-4.1 or other models will not always translate directly to GPT-5.
About GPT-5
OpenAI announced GPT-5, described as a single model for every task in ChatGPT with no model selection or setup required. The company reports benchmark results of 99.6% on AIME 2025 with tools, 74.9% on SWEBench, and 84.2% on MMMU, and says GPT-5 is 45% less likely to contain a factual error than GPT-4o and uses 50-80% fewer output tokens than o3. OpenAI also states that business data is not used for training by default, and lists AES-256 encryption at rest, TLS 1.2+ in transit, SAML SSO, SCIM provisioning, role-based access, and compliance with GDPR, CCPA, CSA STAR, and SOC 2 Type 2, with data residency in seven global regions.
About GPT-5
Portola built Tolan, a voice-first AI companion where people talk with a personalized, animated character that learns from conversations over time. The app is designed for ongoing, open-ended dialogue, and the team focused on memory and character design, using a real-time context management system to keep personality and memory consistent. According to Portola, the GPT-5.1 models delivered gains in steerability and latency, and using GPT-5.1 with the Responses API cut speech initiation time by over 0.7 seconds.
About GPT-5.1
12 months
Google released Gemini 3 Flash, a model it describes as offering frontier intelligence built for speed at a fraction of the cost. It is available through the Gemini app and AI Mode in Search, and for developers via the Gemini API in Google AI Studio, Google Antigravity, Gemini CLI, Android Studio, Vertex AI and Gemini Enterprise. Google reports scores of 90.4% on GPQA Diamond, 33.7% on Humanity's Last Exam without tools, 81.2% on MMMU Pro and 78% on SWE-bench Verified, and prices it at $0.50/1M input tokens and $3/1M output tokens, with audio input at $1/1M input tokens.
About Gemini 3 Flash
OpenAI introduced GPT-5.2 Pro and GPT-5.2 Thinking, which it describes as its strongest models yet for scientific and mathematical work. On GPQA Diamond, GPT-5.2 Pro scores 93.2% and GPT-5.2 Thinking scores 92.4%, with no tools enabled and reasoning effort set to maximum. On FrontierMath (Tier 1–3), GPT-5.2 Thinking set a new state of the art by solving 40.3% of problems, with a Python tool enabled and reasoning effort set to maximum. OpenAI also described a case study in which GPT-5.2 Pro helped resolve an open research problem in statistical learning theory, documented in a paper titled On Learning-Curve Monotonicity for Maximum Likelihood Estimators.
About GPT-5.2
OpenAI published an update to the GPT-5 System Card covering GPT-5.2, which it describes as the latest model family in the GPT-5 series. The safety mitigation approach for these models is largely the same as that described in the GPT-5 System Card and GPT-5.1 System Card. The card refers to GPT-5.2 Instant as gpt-5.2-instant and GPT-5.2 Thinking as gpt-5.2-thinking.
About GPT-5.2
OpenAI introduced GPT-5.2, described as its most capable model series for professional knowledge work and long-running agents. The company says GPT-5.2 is better at creating spreadsheets, building presentations, writing code, perceiving images, understanding long contexts, using tools, and handling complex, multi-step projects. OpenAI reports GPT-5.2 Thinking beats or ties top industry professionals on 70.9% of comparisons on GDPval knowledge work tasks spanning 44 occupations, and sets a state of the art of 55.6% on SWE-Bench Pro. In ChatGPT, GPT-5.2 Instant, Thinking, and Pro begin rolling out starting with paid plans, and in the API they are available to all developers.
About GPT-5.2
OpenAI published an account of mathematician Ernest Ryu, a professor at UCLA, using GPT-5 to explore an open optimization theory problem: why the Nesterov Accelerated Gradient (NAG) method gains speed without losing stability. Ryu had previously tested ChatGPT-3.5 on simple math and logic problems and found its accuracy varied. The article describes NAG, introduced in 1983 by Yurii Nesterov, as an optimization method that calculates gradients at a look-ahead point to converge faster.
About GPT-5
Anthropic released Claude Opus 4.5, describing it as its newest model and calling it the best model in the world for coding, agents, and computer use, and better at tasks like deep research and working with slides and spreadsheets. It is available on Anthropic's apps, its API, and all three major cloud platforms, with developers using claude-opus-4-5-20251101 via the Claude API at pricing of $5/$25 per million tokens. Anthropic also released updates to the Claude Developer Platform, Claude Code, and its consumer apps, including new tools for longer-running agents and new ways to use Claude in Excel, Chrome, and on desktop.
About Claude Opus 4.5
OpenAI released a paper titled "Early science acceleration experiments with GPT-5," co-authored with collaborators at universities and national laboratories including Vanderbilt, UC Berkeley, Columbia, Oxford, Cambridge, Lawrence Livermore National Laboratory, and The Jackson Laboratory. The paper compiles early case studies across math, physics, biology, computer science, astronomy, and materials science in which GPT-5 helped researchers synthesize known results, conduct literature review, accelerate computations, and generate proofs of unsolved propositions, and it also documents limitations. In one case, GPT-5 identified a likely mechanism in human immune cells from an unpublished chart and suggested an experiment that proved it; in another, it contributed an idea that helped researchers complete a proof of a decades-old open problem originally proposed by Paul Erdős.
About GPT-5
xAI launched Grok 4.1 Fast, described as its best tool-calling model with a 2M context window, and the Agent Tools API, which gives agents access to real-time X data, web search, remote code execution, and more. The Agent Tools API also includes file search with citations and MCP tools, and its tools run on xAI's infrastructure. xAI says Grok 4.1 Fast was trained with reinforcement learning in simulated environments across dozens of domains and uses long-horizon reinforcement learning for multi-turn scenarios across its full context window.
About Grok 4.1 Fast
OpenAI released GPT-5.1 in its API platform, describing it as the next model in the GPT-5 series that balances intelligence and speed for agentic and coding tasks. GPT-5.1 adapts how much time it spends thinking based on task complexity and includes a "no reasoning" mode, set via reasoning_effort to 'none', for latency-sensitive uses. OpenAI also released extended prompt caching with up to 24 hour cache retention and introduced two tools with GPT-5.1: an apply_patch tool for editing code and a shell tool for running shell commands. The company said it worked with startups including Cursor, Cognition, Augment Code, Factory, and Warp to improve GPT-5.1's coding personality, steerability, and code quality.
About GPT-5.1
OpenAI released GPT-5.1, an upgrade to the GPT-5 series, rolling out starting with paid users. The release includes GPT-5.1 Instant, described as warmer, more intelligent, and better at following instructions, and GPT-5.1 Thinking, described as easier to understand, faster on simple tasks, and more persistent on complex ones. OpenAI also introduced more intuitive controls for adjusting ChatGPT's tone.
About GPT-5.1
OpenAI published an addendum to the GPT-5 system card covering sensitive conversations. The company states it worked with more than 170 mental health experts on an October 3 update to ChatGPT's default model to better recognize signs of distress, respond with care, and guide people toward real-world support, reducing responses that fall short of desired behavior by 65-80%. The addendum shares baseline safety evaluations comparing the August 15 version of ChatGPT's default model, also known as GPT-5 Instant, to the updated one launched October 3.
About GPT-5
Consensus built Scholar Agent, a multi-agent system on GPT-5 and the Responses API, for its research assistant used by more than 8 million people. The system runs a coordinated workflow of Planning, Search, Reading, and Analysis agents, and Consensus migrated from Chat Completions to the Responses API to support the multi-agent routing. The source states that GPT-5 outperformed GPT-4.1, Sonnet 4, and Gemini 2.5 Pro on tool-calling accuracy and planning stability in early evaluations.
About GPT-5
Anthropic released Claude Haiku 4.5, a small model available to all users, on the Claude API, Amazon Bedrock, and Google Cloud's Vertex AI. The company says it offers coding performance similar to Claude Sonnet 4 at one-third the cost and more than twice the speed, and surpasses Sonnet 4 on some tasks such as computer use. API pricing is $1/$5 per million input and output tokens. Anthropic reports the model showed low rates of concerning behaviors and was released under the AI Safety Level 2 (ASL-2) standard.
About Claude Haiku 4.5
Wrtn scaled its AI apps to 6.5 million monthly active users across Korea using GPT-5, and is taking its approach across East Asia. The company's products include AI Supporter, Labs, and Crack, an open-ended character chat platform. Wrtn runs a router architecture that directs traffic to lightweight models like GPT-4o mini and GPT-4.1 mini while heavier tasks run on GPT-4.1 and multimodal TTS. Wrtn went live with GPT-5 the same day it launched and saw an 8% increase in daily active users within a week.
About GPT-5
Intel researchers accelerated Qwen3-8B on Intel Core Ultra using OpenVINO.GenAI speculative decoding with a lightweight Qwen3-0.6B draft model, reaching an average of about 1.3× speedup over a 4-bit OpenVINO baseline on an Intel Lunar Lake integrated GPU. They then pruned 6 out of 28 layers from the Qwen3-0.6B draft model and fine-tuned it on synthetic data generated by Qwen3-8B from 500k prompts in the BAAI/Infinity-Instruct dataset, which brought the speedup to about 1.4×. The work was demonstrated running a local AI agent with smolagents, and the notebook and depth-pruned draft model were released for reproduction.
About Qwen3-8B
Anthropic released Claude Sonnet 4.5, describing it as its most powerful model to date and its most aligned frontier model. The release includes product updates: checkpoints, a refreshed terminal interface, and a native VS Code extension in Claude Code; a context editing feature and memory tool in the Claude API; code execution and file creation in the Claude apps; and availability of the Claude for Chrome extension to Max users who joined the waitlist. Anthropic also introduced the Claude Agent SDK, giving developers the building blocks used to make Claude Code. Claude Sonnet 4.5 is available via the Claude API as claude-sonnet-4-5, with pricing the same as Claude Sonnet 4 at $3/$15 per million tokens.
About Claude Sonnet 4.5
xAI announced Grok 4 Fast, a cost-efficient reasoning model built on learnings from Grok 4. The model has a 2M token context window, web and X search capabilities, and a unified architecture that blends reasoning and non-reasoning modes. xAI states Grok 4 Fast achieves comparable performance to Grok 4 on benchmarks while using 40% fewer thinking tokens on average, and that this results in a 98% reduction in price to achieve the same performance on frontier benchmarks as Grok 4.
About Grok 4 Fast
OpenAI introduced GPT-5, which it describes as its smartest, fastest, and most useful model yet, uniting and exceeding its prior work in frontier intelligence across 4o, OpenAI o-series reasoning, agents, and advanced math. GPT-5 was rolling out to Team users, with access for Enterprise and Edu to follow, and it is available in the OpenAI API. OpenAI states that 5 million paid users use ChatGPT business products and that nearly 700 million people use ChatGPT weekly.
About GPT-5
OpenAI released GPT-5 in its API platform, describing it as its best model for coding and agentic tasks. The company reports scores of 74.9% on SWE-bench Verified and 88% on Aider polyglot, and says GPT-5 beats OpenAI o3 at frontend web development 70% of the time in internal testing. GPT-5 is offered in three API sizes—gpt-5, gpt-5-mini, and gpt-5-nano—with new API features including a verbosity parameter (low, medium, high), a minimal value for the reasoning_effort parameter, and a custom tools type that lets GPT-5 call tools with plaintext instead of JSON. OpenAI states that GPT-5 in the API is the reasoning model that powers maximum performance in ChatGPT, and that the non-reasoning model used in ChatGPT is available as gpt-5-chat-latest.
About GPT-5
No brief captured for this post yet — the headline and the source are the publisher’s.
About GPT-5
No brief captured for this post yet — the headline and the source are the publisher’s.
About GPT-5
No brief captured for this post yet — the headline and the source are the publisher’s.
About GPT-5
No brief captured for this post yet — the headline and the source are the publisher’s.
About GPT-5
No brief captured for this post yet — the headline and the source are the publisher’s.
About GPT-5
No brief captured for this post yet — the headline and the source are the publisher’s.
About GPT-5
OpenAI introduced GPT-5, describing it as a unified system that answers most questions with a fast model and uses a deeper reasoning model (GPT-5 thinking) for harder problems, with a real-time router choosing between them. GPT-5 is available to all users, with Plus subscribers getting more usage and Pro subscribers getting access to GPT-5 pro, a version with extended reasoning. OpenAI reports results including 94.6% on AIME 2025 without tools, 74.9% on SWE-bench Verified, 88% on Aider Polyglot, 84.2% on MMMU, and 46.2% on HealthBench Hard, and says GPT-5 pro scores 88.4% on GPQA without tools.
About GPT-5
OpenAI published the GPT-5 System Card describing GPT-5 as a unified system with a fast model for most questions, a deeper reasoning model for harder problems, and a real-time router that picks between them based on conversation type, complexity, tool needs, and explicit intent. The card names the fast models gpt-5-main and gpt-5-main-mini and the thinking models gpt-5-thinking and gpt-5-thinking-mini, with API access also to gpt-5-thinking-nano and a ChatGPT setting called gpt-5-thinking-pro. OpenAI states the system card focuses primarily on gpt-5-thinking and gpt-5-main, with other models' evaluations in the appendix, and that gpt-5-thinking is treated as High capability in the Biological and Chemical domain under its Preparedness Framework.
About GPT-5
No brief captured for this post yet — the headline and the source are the publisher’s.
About GPT-5
OpenAI introduced gpt-oss-120b and gpt-oss-20b, two open-weight reasoning models released under the Apache 2.0 license and the gpt-oss usage policy. The text-only models are compatible with the Responses API and are designed for agentic workflows with instruction following, tool use such as web search and Python code execution, and reasoning that can adjust effort for simpler tasks; they are customizable, provide full chain-of-thought, and support Structured Outputs. OpenAI states that scalable capability evaluations found the default gpt-oss-120b did not reach its indicative thresholds for High capability in Biological and Chemical capability, Cyber capability, or AI Self-Improvement, and that adversarially fine-tuned versions did not reach High capability in Biological and Chemical Risk or Cyber risk. OpenAI calls the document a model card rather than a system card because the models will be used in systems built and maintained by many stakeholders.
About gpt-oss-120b
Anthropic released Claude Opus 4.1, an upgrade to Claude Opus 4 on agentic tasks, real-world coding, and reasoning. It is available to paid Claude users and in Claude Code, and on the API, Amazon Bedrock, and Google Cloud's Vertex AI, at the same pricing as Opus 4. Anthropic reports Opus 4.1 reaches 74.5% on SWE-bench Verified and improves in-depth research and data analysis, especially detail tracking and agentic search. GitHub, Rakuten Group, and Windsurf are cited on performance gains, including multi-file code refactoring and debugging precision.
About Claude Opus 4.1
NVIDIA reports that its AI-Q Blueprint, an open deep research agent, reached the top of the Hugging Face "LLM with Search" leaderboard on DeepResearch Bench. AI-Q combines Llama 3.3-70B Instruct and Llama-3.3-Nemotron-Super-49B-v1.5, and the reference example also includes NVIDIA NeMo Retriever and the NVIDIA NeMo Agent toolkit. The article states AI-Q achieved an overall score of 40.52 in the LLM with Search category as of August 2025, holding the top spot for any fully open-licensed stack. Both models are available for download on Hugging Face.
About DeepResearch Bench
xAI announced Grok 4, which it describes as its most intelligent model, with native tool use and real-time search integration. It is available to SuperGrok and Premium+ subscribers and through the xAI API. xAI also introduced a SuperGrok Heavy tier with access to Grok 4 Heavy, described as the most powerful version of Grok 4. The company said Grok 4's reinforcement learning training ran on its Colossus 200,000 GPU cluster, with compute efficiency increased 6x and expanded verifiable training data across more domains.
About Grok 4
Genspark launched Super Agent, a no-code, customizable AI agent that uses OpenAI multimodal models and the Realtime API to handle tasks across text, image, and voice. The platform orchestrates nine specialized large language models and more than 80 integrated tools, with GPT-4.1 handling research and structured outputs using a 1M-token context window and GPT-image-1 used for image generation. Its Call For Me feature uses the OpenAI Realtime API and speech-to-speech capabilities to hold phone conversations. Genspark reports that Super Agent reached $36M in ARR in just 45 days.
About GPT-4.1
OpenAI published a customer story on Retell AI, a voice AI platform for contact center operations. Retell uses OpenAI models to generate responses for 80% of calls on its platform, which handles more than 100 million monthly calls and has surpassed 1 trillion tokens consumed across OpenAI models. In Retell's October 13, 2026 benchmark, GPT-6 Astra ranked first for instruction following at 89.7%, and Retell's engineers use Codex for development work.
About GPT-4o
OpenAI published a customer story on CodeRabbit, a code review tool launched in 2023. CodeRabbit uses OpenAI's o4-mini and o3 models for tasks such as multi-line bugs, code refactors, and architecture issues across files, and GPT-4.1 for review summarization, docstring generation, and routine QA checks. CodeRabbit also launched an integration into Visual Studio Code for reviews in the editor, and reports a 50% increase in accurate suggestions since adopting o3.
About o3
OpenAI published an addendum to its o3 and o4-mini system card describing Codex, a cloud-based coding agent powered by codex-1, a version of OpenAI o3 optimized for software engineering. codex-1 was trained using reinforcement learning on real-world coding tasks in a variety of environments. Each Codex agent runs in its own cloud container with no internet access, preloaded with the user's code and a user-defined development environment, where it can read and edit files and execute commands including tests, linters, and type checkers. Codex is trained to cite terminal logs and files as evidence of its actions, and users can inspect results, request refinements, or export the generated diff as a GitHub pull request or for local testing.
About o3
No brief captured for this post yet — the headline and the source are the publisher’s.
About GPT-4o
OpenAI rolled back a GPT-4o update in ChatGPT, so people are using an earlier version with more balanced behavior. The removed update was overly flattering or agreeable, often described as sycophantic. OpenAI says it focused too much on short-term feedback in that update and did not fully account for how users' interactions with ChatGPT evolve over time. OpenAI is testing fixes, revising how it collects and incorporates feedback to weight long-term user satisfaction, and introducing more personalization features.
About GPT-4o
OpenAI released a system card for OpenAI o3 and OpenAI o4-mini, which combine reasoning with tool capabilities including web browsing, Python, image and file analysis, image generation, canvas, automations, file search, and memory. The models are trained with large-scale reinforcement learning on chains of thought and use tools within those chains of thought. It is the first launch and system card released under Version 2 of the Preparedness Framework; OpenAI's Safety Advisory Group determined the models do not reach the High threshold in Biological and Chemical Capability, Cybersecurity, or AI Self-improvement.
About o3
OpenAI introduced OpenAI o3 and o4-mini, described as o-series models trained to think for longer before responding and able to agentically use and combine tools within ChatGPT, including web search, Python analysis of uploaded files, reasoning about visual inputs, and image generation. OpenAI o3 is described as its most powerful reasoning model, setting a new SOTA on benchmarks including Codeforces, SWE-bench (without building a custom model-specific scaffold), and MMMU, and making 20 percent fewer major errors than OpenAI o1 on difficult, real-world tasks in external expert evaluations. OpenAI o4-mini is described as a smaller model optimized for fast, cost-efficient reasoning, particularly in math, coding, and visual tasks, and as the best-performing benchmarked model on AIME 2024 and 2025. The post also states that o3 and o4-mini are no longer available in ChatGPT and are deprecated in the API, and that OpenAI o3‑pro became available to Pro users in ChatGPT and in the API on June 10, 2025.
About o3
OpenAI launched three models in the API: GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano. The models have larger context windows supporting up to 1 million tokens and a knowledge cutoff of June 2024. GPT-4.1 scores 54.6% on SWE-bench Verified, 38.3% on Scale's MultiChallenge, and 72.0% on Video-MME's long, no subtitles category. GPT-4.1 is available only via the API, and OpenAI will turn off GPT-4.5 Preview in the API on July 14, 2025.
About GPT-4.1
Hugging Face announced that Meta's Llama 4 Maverick (~400B) and Llama 4 Scout (~109B) models are available on the Hugging Face Hub. Both are Mixture of Experts models with 17B active parameters; Maverick uses 128 experts and Scout uses 16. The models are integrated with transformers (v4.51.0) and supported by Text Generation Inference, and their weights are released under the Llama 4 Community License Agreement.
About Llama 4 Maverick
Anthropic released a second Anthropic Economic Index report covering usage data on Claude.ai after the launch of Claude 3.7 Sonnet. The report says the share of usage for coding, education, science, and healthcare applications rose, and that the model's "extended thinking" mode is used predominantly for technical tasks. It also releases task- and occupation-level augmentation and automation data and a bottom-up taxonomy of 630 granular usage categories, with datasets freely available to download.
About Claude 3.7 Sonnet
OpenAI published an addendum to the GPT-4o system card covering 4o image generation, described as a more capable image generation approach than the earlier DALL·E 3 series of models. The company says it can create photorealistic output, take images as inputs and transform them, follow detailed instructions including reliably incorporating text into images, and is embedded natively in the omnimodal GPT-4o model. OpenAI states that 4o image generation benefits from its existing safety infrastructure and lessons from deploying DALL·E and Sora, while the new capabilities bring new risks. The addendum describes the marginal risks OpenAI focused on and the work done to address them.
About GPT-4o
The Qwen Team introduced QwQ-32B, a 32 billion parameter model that it says achieves performance comparable to DeepSeek-R1, which has 671 billion parameters with 37 billion activated. The model was trained with a reinforcement learning scaling approach using outcome-based rewards, starting with math and coding tasks and then adding a stage for general capabilities. QwQ-32B is open-weight on Hugging Face and ModelScope under the Apache 2.0 license and is accessible via Qwen Chat.
About QwQ-32B
OpenAI released a research preview of OpenAI GPT-4.5, described as its largest and most knowledgeable model yet. It builds on GPT-4o, scales pre-training further, and was trained with new supervision techniques plus supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). OpenAI says safety evaluations found no significant increase in safety risk compared to existing models, and the Preparedness scorecard rates CBRN and Persuasion as Medium and Cybersecurity and Model autonomy as Low.
About GPT-4.5
Endex, an AI platform for financial firms, is developing an AI Analyst that retrieves, synthesizes, and reasons through complex financial data, built with OpenAI models including GPT-4o, o1-mini, o1-preview, and o3-mini. Endex says its agents use OpenAI's reasoning models instead of retrieval-augmented generation to pull facts, identify inconsistencies, and contextualize metrics, and that o3-mini improved intelligence at one-third the latency per turn. The agents handle tasks such as precedent transaction overviews, earnings performance summaries, investment committee memo preparation, and data room due diligence, delivering outputs as emails, documents, Excel models, or slide decks.
About o3-mini
Anthropic announced Claude 3.7 Sonnet, described as its most intelligent model to date and the first hybrid reasoning model on the market. It can give near-instant responses or extended, step-by-step thinking visible to the user, and API users can control how long the model thinks. Anthropic also introduced Claude Code, a command line tool for agentic coding available as a limited research preview. Claude 3.7 Sonnet is available on all Claude plans and on the Claude Developer Platform, Amazon Bedrock, and Google Cloud's Vertex AI, with extended thinking mode on all surfaces except the free Claude tier; pricing is $3 per million input tokens and $15 per million output tokens, including thinking tokens.
About Claude 3.7 Sonnet
OpenAI released OpenAI o3-mini, described as the newest and most cost-efficient model in its reasoning series, available in ChatGPT and the API. The model supports function calling, Structured Outputs, and developer messages, offers low, medium, and high reasoning effort options, and does not support vision capabilities. It rolls out in the Chat Completions API, Assistants API, and Batch API to select developers in API usage tiers 3-5, and is available to ChatGPT Plus, Team, and Pro users, with Enterprise access coming in February. The rate limit for Plus and Team users rises from 50 messages per day with o1-mini to 150 messages per day with o3-mini, and free plan users can try it by selecting 'Reason' in the message composer or by regenerating a response.
About o3-mini
OpenAI published a system card for OpenAI o3-mini. Under the Preparedness Framework, OpenAI's Safety Advisory Group recommended classifying the OpenAI o3-mini (Pre-Mitigation) model as Medium risk overall, with Medium risk for Persuasion, CBRN, and Model Autonomy, and Low risk for Cybersecurity. OpenAI states that o3-mini is the first model to reach Medium risk on Model Autonomy, while it still performs poorly on evaluations designed to test real-world ML research capabilities relevant for self improvement. The report covers safety evaluations, external red teaming, and Preparedness Framework evaluations.
About o3-mini
The Qwen Team announced Qwen2.5-Max, a large-scale MoE model pretrained on over 20 trillion tokens and post-trained with SFT and RLHF. The team reported that Qwen2.5-Max outperforms DeepSeek V3 on Arena-Hard, LiveBench, LiveCodeBench, and GPQA-Diamond, with competitive results on MMLU-Pro. Qwen2.5-Max is available in Qwen Chat, and its API, with the model name qwen-max-2025-01-25, is available through Alibaba Cloud.
About Qwen2.5-Max
The Open-R1 project is an initiative to systematically reconstruct DeepSeek-R1's data and training pipeline, validate its claims, and push the boundaries of open reasoning models. The project cites open questions from the DeepSeek-R1 release about how reasoning-specific datasets were curated, which training hyperparameters work best, and the compute and data trade-offs in training reasoning models. DeepSeek-R1 is described as a reasoning model built on DeepSeek-V3, a 671B Mixture of Experts model trained at a cost of $5.5M. DeepSeek also introduced DeepSeek-R1-Zero, which relied entirely on reinforcement learning using Group Relative Policy Optimization (GRPO).
About DeepSeek-R1
10 months
OpenAI published a system card for the OpenAI o1 and OpenAI o1-mini models, which are trained with large-scale reinforcement learning to reason using chain-of-thought before answering. The report covers safety evaluations, external red teaming, and Preparedness Framework evaluations. Preparedness Scorecard ratings list Cybersecurity as Low, CBRN as Medium, Persuasion as Medium, and Model Autonomy as Low. The models were pre-trained on publicly available data, proprietary data accessed through partnerships, and custom in-house datasets.
About o1
OpenAI published a case study on Grab using GPT-4o vision fine-tuning to refine GrabMaps mapping data. Grab fine-tuned GPT-4o with 100 sample cases to match speed limit signs to roads, raising accuracy from a 67% baseline to 80% after two rounds of fine-tuning, a 13-percentage point gain. Grab reports lane count accuracy increased by 20% and speed limit sign localization improved by 13%, with reduced manual mapping effort. The article also describes Grab's work on a voice assistant and a support chatbot.
About GPT-4o
Claude 3.5 Sonnet is rolling out on GitHub Copilot, letting developers choose it for coding in Visual Studio Code and on GitHub.com. The model outperforms all publicly available models on SWE-bench Verified and scores 93.7% on HumanEval, according to the announcement. It will be available in public preview to all GitHub Copilot Chat users and organizations over the coming weeks, running via Amazon Bedrock.
About Claude 3.5 Sonnet
Anthropic announced an upgraded Claude 3.5 Sonnet and a new model, Claude 3.5 Haiku. The upgraded Claude 3.5 Sonnet improves on SWE-bench Verified from 33.4% to 49.0% and on TAU-bench from 62.6% to 69.2% in the retail domain and from 36.0% to 46.0% in the airline domain, at the same price and speed as its predecessor. Anthropic also introduced computer use in public beta on the API, which lets developers direct Claude to look at a screen, move a cursor, click buttons, and type text; Claude 3.5 Sonnet is the first frontier AI model to offer computer use in public beta, and Anthropic describes it as experimental and at times cumbersome and error-prone. Claude 3.5 Haiku scores 40.6% on SWE-bench Verified and was released as a text-only model.
About Claude 3.5 Sonnet
Mercado Libre introduced Verdi, an AI development platform layer built on GPT-4o, GPT-4o mini, and GPT-3.5 Turbo, used for customer service and other complex tasks. Verdi lets developers build with LLMs and has begun handling 10% of customer service mediation on one of Mercado's major sites. Mercado Libre says the platform has the potential to support the tasks of 9,000 human operators and autonomously manage customer service decisions involving $450 million annually.
About GPT-4o
OpenAI released OpenAI o1-mini, a cost-efficient reasoning model optimized for STEM, especially math and coding. The company said it nearly matches OpenAI o1 on benchmarks such as AIME and Codeforces, and launched it to tier 5 API users at a cost 80% cheaper than OpenAI o1-preview, with ChatGPT Plus, Team, Enterprise, and Edu users able to use it as an alternative to o1-preview. OpenAI stated o1-mini performs worse on tasks requiring non-STEM factual knowledge. The post also notes that o1-mini has been retired from the API.
About o1-mini
No brief captured for this post yet — the headline and the source are the publisher’s.
About o1
No brief captured for this post yet — the headline and the source are the publisher’s.
About GPT-4o
OpenAI launched fine-tuning for GPT-4o, letting developers train custom versions on their own datasets to improve performance and accuracy for specific use cases. GPT-4o fine-tuning is available to all developers on all paid usage tiers, with training priced at $25 per million tokens and inference at $3.75 per million input tokens and $15 per million output tokens. OpenAI also offers 1M training tokens per day for free for every organization through September 23, and GPT-4o mini fine-tuning is available with 2M training tokens per day free through the same date. An update dated May 8, 2026 states that OpenAI is winding down the fine-tuning platform: it is no longer accessible to new users, existing users can create training jobs for the coming months, and fine-tuned models remain available for inference until their base models are deprecated.
About GPT-4o
OpenAI released SWE-bench Verified, a human-validated subset of SWE-bench that it says more reliably evaluates AI models' ability to solve real-world software issues. OpenAI states its testing found some SWE-bench tasks that may be hard or impossible to solve, leading the benchmark to systematically underestimate models' autonomous software engineering capabilities. OpenAI says it collaborated with the authors of SWE-bench on a new release of the benchmark intended to provide more accurate evaluations.
About SWE-Bench verified
OpenAI published the GPT-4o System Card together with the Preparedness Framework scorecard, describing an end-to-end safety assessment of GPT-4o. The evaluations covered risks including speaker identification, unauthorized voice generation, potential generation of copyrighted content, ungrounded inference, and disallowed content, and OpenAI says it implemented safeguards at the model and system levels. In the Preparedness Framework scorecard, cybersecurity, biological threats, and model autonomy scored low, while persuasion scored borderline medium; OpenAI states that GPT-4o's voice modality doesn't meaningfully increase Preparedness risks. The card also describes GPT-4o as an autoregressive omni model trained end-to-end across text, vision, and audio.
About GPT-4o
OpenAI announced GPT-4o mini, described as its most cost-efficient small model. It is priced at 15 cents per million input tokens and 60 cents per million output tokens, and OpenAI says it scores 82% on MMLU. The model supports text and vision in the API, has a 128K token context window, supports up to 16K output tokens per request, and has knowledge up to October 2023.
About GPT-4o mini
Mistral released Mistral NeMo, a 12B model built in collaboration with NVIDIA, with a context window of up to 128k tokens. Pre-trained base and instruction-tuned checkpoints are released under the Apache 2.0 license, and the model was trained with quantisation awareness to enable FP8 inference without performance loss. Mistral NeMo uses a new tokenizer called Tekken, based on Tiktoken and trained on more than 100 languages, which the company says compresses source code and several languages more efficiently than the SentencePiece tokenizer used in previous Mistral models. The model is exposed on la Plateforme as open-mistral-nemo-2407 and is packaged as an NVIDIA NIM inference microservice available from ai.nvidia.com.
About Mistral NeMo
Anthropic announced that customers can fine-tune Claude 3 Haiku in Amazon Bedrock to customize its knowledge and capabilities for specialized tasks. Fine-tuning uses high quality prompt-completion pairs, and the fine-tuning API was available in preview in the US West (Oregon) AWS Region with text-based fine-tuning and context lengths up to 32K tokens. An update in the article states that fine-tuning Claude 3 Haiku in Amazon Bedrock is generally available as of November 1, 2024.
About Claude 3 Haiku
Anthropic launched Claude 3.5 Sonnet, the first release in the Claude 3.5 model family. The model is available for free on Claude.ai and the Claude iOS app, with higher rate limits for Claude Pro and Team subscribers, and via the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI. It costs $3 per million input tokens and $15 per million output tokens, with a 200K token context window. Anthropic also introduced Artifacts on Claude.ai, a preview feature that displays generated content such as code snippets, text documents, or website designs in a dedicated window alongside the conversation.
About Claude 3.5 Sonnet
Color Health is working with OpenAI on a copilot application that uses GPT-4o to identify missing diagnostics and create tailored workup plans for cancer screening and treatment. The application extracts and normalizes patient information such as family history and risk factors along with clinical guidelines, then generates personalized screening plans and documentation such as medical necessity documents and insurance pre-authorizations. A clinician reviews and can modify the output at every step before it is presented to the patient. Color Health says it has served more than 7 million patients since it was founded.
About GPT-4o
OpenAI announced GPT-4o, described as its new flagship model that can reason across audio, vision, and text in real time. The model accepts any combination of text, audio, image, and video as input and generates text, audio, and image outputs, and it can respond to audio inputs in as little as 232 milliseconds, with an average of 320 milliseconds. OpenAI says GPT-4o matches GPT-4 Turbo performance on text in English and code, improves on text in non-English languages, and is much faster and 50% cheaper in the API. OpenAI states it trained a single new model end-to-end across text, vision, and audio, and that it is still exploring the model's capabilities and limitations.
About GPT-4o
OpenAI introduced GPT-4o, described as its newest flagship model that provides GPT-4-level intelligence but is faster and improves on capabilities across text, voice, and vision. OpenAI said it was rolling out GPT-4o to ChatGPT Plus and Team users and to ChatGPT Free with usage limits, with Enterprise access to follow; Plus users get a message limit up to 5x greater than free users. Free users gain access to features including responses from both the model and the web, data analysis and chart creation, chatting about photos, file uploads, GPTs and the GPT Store, and Memory, with ChatGPT switching to GPT-3.5 when the GPT-4o message limit is reached. OpenAI also launched a ChatGPT desktop app for macOS, rolling out to Plus users first, and introduced a new look and feel for ChatGPT.
About GPT-4o
Researchers from UC Berkeley, MIT, and Cornell developed LiveCodeBench, a benchmark for measuring LLMs' code generation capabilities, and released a leaderboard based on it. LiveCodeBench collects coding problems over time from LeetCode, AtCoder, and CodeForces and annotates them with release dates, so models can be evaluated on problem sets from different time windows to help detect and prevent contamination. It covers four scenarios: code generation, self-repair, code execution, and test output prediction, each evaluated with the Pass@1 metric.
About LiveCodeBench
The post describes running the Microsoft Phi-2 model locally on a laptop with an Intel Meteor Lake (Core Ultra) CPU. It applies 4-bit quantization to the model weights using the Intel OpenVINO integration in the Optimum Intel library, then runs inference on a mid-range laptop powered by a Core Ultra 7 155H CPU. Phi-2 is described as a 2.7-billion parameter model trained for text generation.
About Phi-2
Anthropic released Claude 3 Haiku, described as the fastest and most affordable model in its intelligence class, with vision capabilities and strong benchmark performance. The model is available in the Claude API and on claude.ai for Claude Pro subscribers, and is available on Amazon Bedrock with a planned release on Google Cloud Vertex AI. Anthropic states Claude 3 Haiku is three times faster than its peers for the vast majority of workloads, processing 21K tokens per second for prompts under 32K tokens, and that it can process and analyze 400 Supreme Court cases or 2,500 images for one US dollar. The company also describes security measures including testing, systems monitoring, encryption, access controls, and security audits.
About Claude 3 Haiku
1 month
Anthropic announced Claude 2, a new model with improved performance, longer responses, and access via API and a public beta website at claude.ai. The company said Claude 2 scored 76.5% on the multiple choice section of the Bar exam, up from 73.0% with Claude 1.3, and improved on the Codex HumanEval coding test to 71.2% from 56.0%. Users can input up to 100K tokens per prompt, and the Claude 2 API is offered at the same price as Claude 1.3. The beta chat experience is available in the US and UK.
About Claude 2
Headlines and links are the publisher’s. The paragraph under each headline is tensor.news’s own brief, written from the article itself at capture time and checked against the source before it ships — it is a summary, not the lab’s wording, and the original is always linked. tensor.news adds only what the record can check: which tracked model or benchmark the post names, and what the scores say about it.