Anthropic·released 2025-09-291 source
Claude Sonnet 4.5 benchmark scores: 27 benchmarks tracked. 30% of its scores are independently reproduced — the rest are self-reported or unverified.
Leads on 0 of 27 benchmarks·best result 97.73 on MATH level 5 (reproduced)·8 of 27 independently reproduced·$3/$15 per M tokens
Consensus: LiteLLM · Cross-check: models.dev, OpenRouter · See every model’s pricing →
| Benchmark | Score | Evidence status | Measured |
|---|---|---|---|
| MATH level 5 | 97.73accuracy (%) | reproduced· optimizedT1 | 2025-09-29Epoch AI |
| OTIS Mock AIME 2024-2025 | 77.76accuracy (%) | reproduced· optimizedT1 | 2025-09-29Epoch AI |
| GPQA diamond | 76.43accuracy (%) | reproduced· optimizedT1 | 2025-09-29Epoch AI |
| DTBench | 72accuracy (%) | unverifiedT1 | 2025-09-29https://conceptualreasoning.ai/dtbench |
| SWE-Bench verified | 71.28accuracy (%) | reproduced· optimizedT1 | 2025-09-29Epoch AI |
| METR Time Horizons | 67.38accuracy (%) | unverifiedT1 | 2025-09-29METR - Measuring AI Ability to Complete Long Tasks |
| ARC-AGI | 63.67accuracy (%) | unverified· optimizedT2 | 2025-09-29 |
| OSWorld | 62.9accuracy (%) | unverified· optimizedT1 | 2025-09-29OS World Website |
| Cybench | 60accuracy (%) | self-reported· optimizedT1 | 2025-09-29Opus 4.5 System Card |
| DeepResearch Bench | 52.6accuracy (%) | unverifiedT1 | 2025-09-29DeepResearchBench Leaderboard |
| WeirdML | 47.71accuracy (%) | unverifiedT1 | 2025-09-29https://htihle.github.io/weirdml.html |
| Terminal Bench | 46.52accuracy (%) | unverifiedT1 | 2025-09-29https://www.tbench.ai/leaderboard/terminal-bench/2.0 |
| LMCA | 45.6accuracy (%) | unverifiedT1 | 2025-09-29https://conceptualreasoning.ai/lmca |
| SimpleBench | 45.16accuracy (%) | unverifiedT1 | 2025-09-29SimpleBench Leaderboard |
| GDPval | 42.5accuracy (%) | unverifiedT2 | 2025-09-29 |
| SimpleQA Verified | 30.7accuracy (%) | unverifiedT2 | 2025-09-29 |
| FrontierMath-Tiers-1-3-v2-Private | 23.86accuracy (%) | reproducedT1 | 2025-09-29Epoch AI |
| ProofBench | 19accuracy (%) | unverifiedT1 | 2025-09-29https://www.vals.ai/benchmarks/proof_bench |
| GSO-Bench | 14.71accuracy (%) | unverifiedT1 | 2025-09-29GSO Leaderboard |
| ARC-AGI-2 | 13.61accuracy (%) | unverifiedT2 | 2025-09-29 |
| VPCT | 9.7accuracy (%) | unverifiedT1 | 2025-09-29VPCT leaderboard |
| HLE | 9.37accuracy (%) | unverifiedT2 | 2025-09-29 |
| Mystery Game Puzzles | 8.57accuracy (%) | reproducedT1 | 2025-09-29Epoch AI |
| Chess Puzzles | 5.3accuracy (%) | unverifiedT2 | 2025-09-29 |
| FrontierMath-Tier-4-v2-Private | 2.44accuracy (%) | reproducedT1 | 2025-09-29Epoch AI |
| EBR-bench | 2.38accuracy (%) | reproducedT1 | 2025-09-29Epoch AI |
| Remote Labor Index | 2.08accuracy (%) | unverifiedT2 | 2025-09-29 |
Claims drawn from cited facts, not live model generation.
This model was developed by Anthropic and released in 2025.
2 cited facts
This model has scores on 27 tracked benchmarks and currently leads 0 of them. The record’s evidence status is independently reproduced.
3 cited facts
Its mean gap to the best recorded score across its tracked benchmarks is 41.95 points.
1 cited fact
6 facts cross-checked across data sources: 2 corroborated, 2 single-source, 2 disputed.
Source facts, citations, and refresh stamp for this record.
Sources: epoch_benchmarks, litellm_prices, modelsdev_models, openrouter_models
Claude Sonnet 4.5 is an AI model developed by Anthropic, released 2025-09-29. Its benchmark record below tags every score as independently reproduced, self-reported, or unverified.
Claude Sonnet 4.5 has recorded scores on 27 benchmarks, each shown with its evidence status.
Claude Sonnet 4.5 has recorded scores on 27 benchmarks — GPQA diamond, MATH level 5, OTIS Mock AIME 2024-2025, DTBench, WeirdML, LMCA, and 21 more. The full table above shows each score with its evidence status.
8 of 27 recorded scores (30%) are independently reproduced rather than self-reported by the lab.
Claude Sonnet 4.5 has 27 tracked claims: 8 independently reproduced, 1 self-reported, 18 unverified.
Listed API pricing: $3 per million input tokens, $15 per million output tokens (prices disputed across sources). See the pricing block for the full breakdown.