tensor.news
Integrity IndexModelsCompareLabsPapersMethodologyAbout
tensor.news

© 2026

Evidence

  • Integrity Index
  • Model Claim Ledger
  • Compare

Entities

  • Models
  • Benchmarks
  • Labs
  • Papers

Method

  • Methodology
  • About
  • Privacy

Machine

  • Feed (Atom)
  • llms.txt
  • llms-full.txt
  • Sitemap

tensor.news grades benchmarks and tags model claims by evidence status. Every fact carries a source record and provenance. Scores are framed as task performance under a disclosed harness, never as deployed capability.

← Integrity Index

HLE

Humanity's Last Exam

Rank #3 of 51 · 37 models scored · top score 43.74

AINTEGRITY 100 / 100

Frontier expert knowledge/reasoning across 100+ subjects

Strongest on discrimination, weakest on saturation headroom. Caveat: the public test set leaves contamination on the table.

consistent harnesspublic test set

Headroom to ceiling

56.3 headroomtop 43.7ceiling 100

Score distribution

median 13.7Gemini 3.1 Pro 43.702346

State of the art over time

0234643.720242026

Integrity breakdown

Weighted composite. Each signal is computed from the source-backed leaderboard below.

Discrimination30%100

Does it still separate models?

Saturation headroom30%99

How far from ceiling / clustered at the top?

Contamination resistance15%100

Public vs held-out; training-leak risk.

Harness comparability15%100

Apples-to-apples, or mixed / vendor-optimized?

Freshness10%100

How old is the benchmark?

Leaderboard

#ModelScoreEvidenceMeasured
1Gemini 3.1 Pro43.74unverified2026-02-19
2GPT-5.4 Pro41.51unverified2026-03-05
3Muse Spark37.56unverified2026-04-08
4Gemini 3 Pro34.37unverified2025-11-18
5GPT-5.433.03unverified2026-03-05
6Claude Opus 4.732.98unverified2026-04-16
7Claude Opus 4.631.13unverified2026-02-05
8GPT-5 Pro28.19unverified2025-10-07
9GPT-5.224.16unverified2025-12-11
10GPT-521.55unverified2025-08-07
11Claude Opus 4.521.43unverified2025-11-24
12Kimi K2.520.56unverified2026-02-02
13GPT-5.119.83unverified2025-11-13
14Gemini 2.5 Pro (Jun 2025)17.69unverified2025-06-05
15o316.3unverified2024-12-20
16GPT-5 mini15.38unverified2025-08-07
17Gemini 2.5 Pro (Mar 2025)14.03unverified2025-03-25
18o4-mini13.95unverified2025-04-16
19Gemini 2.5 Pro (May 2025)13.66unverified2025-05-06
20Claude Sonnet 4.59.37unverified2025-09-29
21Gemini 2.5 Flash (Apr 2025)7.65unverified2025-04-17
22Claude Opus 4.17.06unverified2025-08-05
23Gemini 2.5 Flash (May 2025)6.47unverified2025-05-20
24Claude Opus 46.22unverified2025-05-22
25Gemini 3.1 Flash-Lite4.03unverified2026-03-03
26Claude 3.7 Sonnet3.4unverified2025-02-24
27o13.32unverified2024-12-05
28Claude Sonnet 43.11unverified2025-05-22
29Gemini 2.0 Flash Thinking (Jan 2025)1.85unverified2025-01-21
30Llama 4 Maverick0.92unverified2025-04-05
31GPT-4.50.67unverified2025-02-27
32GPT-4.10.63unverified2025-04-14
33Amazon Nova Pro0unverified2024-12-03
34Claude 3.5 Sonnet (October 2024)0unverified2024-10-22
35Mistral Medium 30unverified2025-05-07
36GPT-4o (Nov 2024)0unverified2024-05-13
37Gemini 1.5 Pro (Sept 2024)0unverified2024-09-24

At a glance

Models scored
37
Top model
Gemini 3.1 Pro
Top score
43.74
Score spread
43.74
Headroom
56.26

Benchmark context

Task type
reasoning
Metric
Accuracy + calibration/confidence-error
Score ceiling
100
Human baseline
No aggregate (at/beyond individual expert frontier)
Test-set privacy
public_plus_private_holdout
Construction
~2500 Qs from ~1000 expert contributors, multi-stage review; ~10-14% multimodal
Key assumptions
Qs that stump current models measure the frontier; confidence exposes overclaiming
Limitations
Short-answer/MC lucky hits; obscure recall; public portion contaminates; not agentic
Contamination history
Public + private holdout to detect gaming; public leakage expected
Harness notes
Single-shot; big swings text-only vs browsing-enabled; SOTA ~mid-50s% 2026
Authors
CAIS & Scale AI
Released
2025-01
License
Public dataset + withheld holdout
Source
agi.safe.ai/

Provenance

37 source records · scores from epoch_benchmarks. Every fact on this page is evidence-linked.

Follow the record

Integrity neighbors

  • APEX-AgentsA · 100
  • Chess PuzzlesA · 100
  • SimpleQA VerifiedA · 100
  • ARC-AGI-2A · 99
  • FrontierMath-2025-02-28-PrivateA · 99

Top models here

  • Gemini 3.1 Pro43.74
  • GPT-5.4 Pro41.51
  • Muse Spark37.56
  • Gemini 3 Pro34.37
  • GPT-5.433.03

Head-to-head

  • Gemini 3.1 Pro vs GPT-5.4 Pro
  • Gemini 3.1 Pro vs Muse Spark
  • GPT-5.4 Pro vs Muse Spark