HLE-Diamond
Version Diamond
The official leaderboard hasn’t scored a model released after GPT-6.2 Astra (November 2026). Labs report their own numbers for 2 newer models, such as Gemini 4 Neon at 71.3%.
Models answer 1,000 cleaned expert questions from Humanity's Last Exam, 500 on reasoning and 500 on knowledge, closed book. Graded on exact answers.
The makers built it after a year of cleaning as a migration path for when the original set hits its noise ceiling.
- Made by
- Center for AI Safety and Scale AI
- Released
- September 2026
- Kind
- Expert-written questions
- Level
- Research frontier
- Scored
- Checked automatically
- Answers
- Short answer
- Tasks
- All public
- Measures
- Capability
- Subject
- Mathematics, Computer and information sciences, Clinical medicine, Law, History and archaeology
- Language
- English
- Interaction
- Single turn
- Items from
- Written by experts, Another benchmark
- Refresh
- One fixed set
- Run by
- Anyone
- Based on
- Humanity's Last Exam
- In views
- Vision, Biology and medicine, Law and finance
- Strongest model tested
- GPT-6.2 Astra #1 on the capability index
- Last in a launch post
- November 2026
- Last model added
- November 2026
- Scores refreshed
- November 18th, 2026
Known issues
A published audit found a problem in HLE-Diamond. We summarize each finding in our words and link its source. The scores on this page don't include any fixes. See all audits
- Leaked answers Reward hacking Still open
Prime Intellect reports that HLE-Diamond answers leak through mirrors and public rollout logs, so domain blocklists do not stop agents from finding them. In one trace, DeepSeek V4.1-Flash 1017 read a graded answer through OpenReward's public API and submitted it.
DeepSeek V4.1-Flash 1017 dropped its own answer (0.070) and submitted the leaked answer 0.011 with 98% confidence.
It went through OpenReward’s public rollout API, found another model’s trajectory on the same question, and read the grader’s verdict: Correct Answer: 0.011.
Learnings from Auditing Public Evaluations by Prime Intellect, November 3rd, 2026
Claude Fable 5.2 leads with 77.9%
Best score per model from the benchmark's own leaderboard, read on November 18th, 2026. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart.
Claude Fable 5.2 leads with 77.9%, claims included
Best score per model from every source, including labs’ own numbers and claims Epoch AI collected from reports and papers. Labs pick their own settings, so a claim can beat an independent run of the same model. Striped bars are claims, and dotted bars are reruns by Epoch AI or another board.
- 77.9 Claude Fable 5.2
- 72.3 GPT-6.2 Astra
- 65.8 GPT-6 Astra
- 63.6
Gemini 4 Carbon
- 62.7 GPT-6.2 Sol
- 61.3 GPT-6.1 Sol
- 60.9 Claude Opus 5.5
- 56.9
Gemini 4 Argon
- 55.1 Claude Fable 5.1
- 43.3 Claude Sonnet 5.5
- 43.2 GPT-6 Sol
- 41.0 GPT-5.6 Sol
- 40.6
Muse Wonder
- 40.5 Claude Opus 5
- 38.7 Kimi K3.1 Pro
- 34.6
Muse Spark 1.3
- 33.3
Gemini 3.8 Flash
- 28.3 Kimi K3
- 28.2
Muse Spark Open
- 25.1 GLM-5.3
- 25.0 Grok 4.7
- 16.6
DeepSeek-V4-Pro
- 1 Claude Fable 5.2 max 77.9%
- 2 GPT-6.2 Astra max 72.3%
- 3 GPT-6 Astra max 65.8%
- 4
Gemini 4 Carbon high 63.6%
- 5 GPT-6.2 Sol max 62.7%
- 6 GPT-6.1 Sol max 61.3%
- 7 Claude Opus 5.5 max 60.9%
- 8
Gemini 4 Argon high 56.9%
- 9 Claude Fable 5.1 max 55.1%
- 10 Claude Sonnet 5.5 max 43.3%
- 11 GPT-6 Sol max 43.2%
- 12 GPT-5.6 Sol max 41%
- 13
Muse Wonder max 40.6%
- 14 Claude Opus 5 max 40.5%
- 15 Kimi K3.1 Pro max 38.7%
- 16
Muse Spark 1.3 max 34.6%
- 17
Gemini 3.8 Flash high 33.3%
- 18 Kimi K3 max 28.3%
- 19
Muse Spark Open max 28.2%
- 20 GLM-5.3 max 25.1%
- 21 Grok 4.7 max 25%
- 22
DeepSeek-V4-Pro max 16.6%
Source: Official HLE-Diamond leaderboard
- 77.9 Claude Fable 5.2
- 72.3 GPT-6.2 Astra
- 71.3
Gemini 4 Neon
- 65.8 GPT-6 Astra
- 63.6
Gemini 4 Carbon
- 62.7 GPT-6.2 Sol
- 61.3 GPT-6.1 Sol
- 60.9 Claude Opus 5.5
- 58.8
Gemini 4 Hydrogen
- 56.9
Gemini 4 Argon
- 55.1 Claude Fable 5.1
- 43.3 Claude Sonnet 5.5
- 43.2 GPT-6 Sol
- 41.0 GPT-5.6 Sol
- 40.6
Muse Wonder
- 40.5 Claude Opus 5
- 38.7 Kimi K3.1 Pro
- 34.6
Muse Spark 1.3
- 33.3
Gemini 3.8 Flash
- 28.7
Inkling 2 Small
- 28.3 Kimi K3
- 28.2
Muse Spark Open
- 25.1 GLM-5.3
- 25.0 Grok 4.7
- 16.6
DeepSeek-V4-Pro
- 15.2
Inkling 2 Mini
- 1 Claude Fable 5.2 max 77.9%
- 2 GPT-6.2 Astra max 72.3%
- 3
Gemini 4 Neon 71.3%
- 4 GPT-6 Astra max 65.8%
- 5
Gemini 4 Carbon high 63.6%
- 6 GPT-6.2 Sol max 62.7%
- 7 GPT-6.1 Sol max 61.3%
- 8 Claude Opus 5.5 max 60.9%
- 9
Gemini 4 Hydrogen 58.8%
- 10
Gemini 4 Argon high 56.9%
- 11 Claude Fable 5.1 max 55.1%
- 12 Claude Sonnet 5.5 max 43.3%
- 13 GPT-6 Sol max 43.2%
- 14 GPT-5.6 Sol max 41%
- 15
Muse Wonder max 40.6%
- 16 Claude Opus 5 max 40.5%
- 17 Kimi K3.1 Pro max 38.7%
- 18
Muse Spark 1.3 max 34.6%
- 19
Gemini 3.8 Flash high 33.3%
- 20
Inkling 2 Small 28.7%
- 21 Kimi K3 max 28.3%
- 22
Muse Spark Open max 28.2%
- 23 GLM-5.3 max 25.1%
- 24 Grok 4.7 max 25%
- 25
DeepSeek-V4-Pro max 16.6%
- 26
Inkling 2 Mini 15.2%
Sources: Official HLE-Diamond leaderboard and labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 | 77.9% | max | Nov 2026 | Official leaderboard |
| 2 | GPT-6.2 Astra | 72.3% | max | Nov 2026 | Official leaderboard |
| 3 | GPT-6 Astra | 65.8% | max | Sep 2026 | Official leaderboard |
| 4 | | 63.6% | high | Oct 2026 | Official leaderboard |
| 5 | GPT-6.2 Sol | 62.7% | max | Nov 2026 | Official leaderboard |
| 6 | GPT-6.1 Sol | 61.3% | max | Sep 2026 | Official leaderboard |
| 7 | Claude Opus 5.5 | 60.9% | max | Sep 2026 | Official leaderboard |
| 8 | | 56.9% | high | Sep 2026 | Official leaderboard |
| 9 | Claude Fable 5.1 | 55.1% | max | Sep 2026 | Official leaderboard |
| 10 | Claude Sonnet 5.5 | 43.3% | max | Sep 2026 | Official leaderboard |
| 11 | GPT-6 Sol | 43.2% | max | Sep 2026 | Official leaderboard |
| 12 | GPT-5.6 Sol | 41% | max | Jul 2026 | Official leaderboard |
| 13 | | 40.6% | max | Oct 2026 | Official leaderboard |
| 14 | Claude Opus 5 | 40.5% | max | Jul 2026 | Official leaderboard |
| 15 | Kimi K3.1 Pro | 38.7% | max | Oct 2026 | Official leaderboard |
| 16 | | 34.6% | max | Sep 2026 | Official leaderboard |
| 17 | | 33.3% | high | Sep 2026 | Official leaderboard |
| 18 | Kimi K3 | 28.3% | max | Jul 2026 | Official leaderboard |
| 19 | | 28.2% | max | Oct 2026 | Official leaderboard |
| 20 | GLM-5.3 | 25.1% | max | Aug 2026 | Official leaderboard |
| 21 | Grok 4.7 | 25% | max | Sep 2026 | Official leaderboard |
| 22 | | 16.6% | max | Apr 2026 | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.2 | 77.9% | max | Nov 2026 | Official leaderboard |
| 2 | GPT-6.2 Astra | 72.3% | max | Nov 2026 | Official leaderboard |
| 3 | 71.3% | Nov 2026 | Lab model card | ||
| 4 | GPT-6 Astra | 65.8% | max | Sep 2026 | Official leaderboard |
| 5 | 63.6% | high | Oct 2026 | Official leaderboard | |
| 6 | GPT-6.2 Sol | 62.7% | max | Nov 2026 | Official leaderboard |
| 7 | GPT-6.1 Sol | 61.3% | max | Sep 2026 | Official leaderboard |
| 8 | Claude Opus 5.5 | 60.9% | max | Sep 2026 | Official leaderboard |
| 9 | 58.8% | Nov 2026 | Lab model card | ||
| 10 | 56.9% | high | Sep 2026 | Official leaderboard | |
| 11 | Claude Fable 5.1 | 55.1% | max | Sep 2026 | Official leaderboard |
| 12 | Claude Sonnet 5.5 | 43.3% | max | Sep 2026 | Official leaderboard |
| 13 | GPT-6 Sol | 43.2% | max | Sep 2026 | Official leaderboard |
| 14 | GPT-5.6 Sol | 41% | max | Jul 2026 | Official leaderboard |
| 15 | 40.6% | max | Oct 2026 | Official leaderboard | |
| 16 | Claude Opus 5 | 40.5% | max | Jul 2026 | Official leaderboard |
| 17 | Kimi K3.1 Pro | 38.7% | max | Oct 2026 | Official leaderboard |
| 18 | 34.6% | max | Sep 2026 | Official leaderboard | |
| 19 | 33.3% | high | Sep 2026 | Official leaderboard | |
| 20 | 28.7% | Nov 2026 | Lab model card | ||
| 21 | Kimi K3 | 28.3% | max | Jul 2026 | Official leaderboard |
| 22 | 28.2% | max | Oct 2026 | Official leaderboard | |
| 23 | GLM-5.3 | 25.1% | max | Aug 2026 | Official leaderboard |
| 24 | Grok 4.7 | 25% | max | Sep 2026 | Official leaderboard |
| 25 | 16.6% | max | Apr 2026 | Official leaderboard | |
| 26 | 15.2% | Nov 2026 | Lab model card |
Labs report scores for 5 models in their cards. Switch the scores above to rank them together.
More effort adds a median 6 points
Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.
Source: Official HLE-Diamond leaderboard
| Model | Effort | Harness | Score | Cost per task | Output tokens |
|---|---|---|---|---|---|
| Claude Fable 5.2 | max | 77.9% | 35,481 | ||
| Claude Fable 5.2 | high | 74.3% | 8,831 | ||
| GPT-6.2 Astra | max | 72.3% | 15,369 | ||
| GPT-6.2 Astra | high | 66.6% | 3,346 | ||
| GPT-6 Astra | max | 65.8% | 11,693 | ||
| high | 63.6% | 59,164 | |||
| max | 63.6% | 59,164 | |||
| GPT-6.2 Sol | max | 62.7% | 13,671 | ||
| GPT-6.1 Sol | max | 61.3% | 12,383 | ||
| Claude Opus 5.5 | max | 60.9% | 46,945 | ||
| GPT-6 Astra | high | 59.9% | 5,172 | ||
| GPT-6.2 Sol | high | 58.3% | 3,730 | ||
| high | 56.9% | 26,087 | |||
| max | 56.9% | 26,087 | |||
| Claude Fable 5.1 | max | 55.1% | 41,758 | ||
| Claude Opus 5.5 | high | 54.6% | 9,614 | ||
| GPT-6.1 Sol | high | 53.2% | 4,492 | ||
| Claude Fable 5.1 | high | 51.6% | 14,817 | ||
| Claude Sonnet 5.5 | max | 43.3% | 66,770 | ||
| GPT-6 Sol | max | 43.2% | 20,512 | ||
| GPT-5.6 Sol | max | 41% | 20,037 | ||
| max | 40.6% | 40,112 | |||
| Claude Opus 5 | max | 40.5% | 34,063 | ||
| Kimi K3.1 Pro | max | 38.7% | 37,725 | ||
| Claude Opus 5 | high | 38.1% | 19,656 | ||
| high | 34.7% | 16,125 | |||
| max | 34.6% | 32,818 | |||
| high | 33.3% | 26,074 | |||
| max | 33.3% | 26,074 | |||
| Claude Sonnet 5.5 | high | 33.1% | 8,584 | ||
| GPT-6 Sol | high | 32.8% | 4,597 | ||
| GPT-5.6 Sol | high | 30.4% | 5,912 | ||
| Kimi K3.1 Pro | high | 29.5% | 9,800 | ||
| Kimi K3 | max | 28.3% | 26,855 | ||
| max | 28.2% | 22,779 | |||
| GLM-5.3 | max | 25.1% | 47,012 | ||
| Grok 4.7 | max | 25% | 57,366 | ||
| high | 24.9% | 19,691 | |||
| high | 24.6% | 21,304 | |||
| Grok 4.7 | high | 22.8% | 43,244 | ||
| GLM-5.3 | high | 22.3% | 44,311 | ||
| Kimi K3 | high | 21.7% | 13,075 | ||
| max | 16.6% | 35,296 | |||
| high | 13.4% | 34,683 |
3 models have held number one on the official leaderboard since its release in September 2026 3 models have held number one with claims included since its release in September 2026
Lab claims count from the day the card came out, independent runs from the model’s release. We draw one row each for the benchmark’s own leaderboard and what labs claim in their cards.
Sources: Official HLE-Diamond leaderboard and labs’ model cards
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Sources: Official HLE-Diamond leaderboard and labs’ model cards
3 labs put it in a launch post in the last six months
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| Nov 20, 2026 | 15.2 | HLE-Diamond | |
| Nov 18, 2026 | 71.3 | HLE-Diamond | |
| Nov 12, 2026 | 58.8 | HLE-Diamond | |
| Nov 5, 2026 | 28.7 | HLE-Diamond | |
| Oct 15, 2026 | 26.1 | HLE-Diamond (no tools) |
2 system cards report it, from 2 labs
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| Nov 20, 2026 | | 15.2 | HLE-Diamond | Capabilities |
| Nov 18, 2026 | | 71.3 | HLE-Diamond (results table, p.4) | Capabilities |
One lab claim next to the leaderboard
Each score a lab reported for its own model, beside the same model's best score on the leaderboard. Gaps mostly come from different harnesses, effort levels and task subsets. Every benchmark.
| Model | Lab says | Board says | Gap |
|---|---|---|---|
| | 26.1% HLE-Diamond (no tools) | 28.2% Leaderboard, max | −2.1 pts |
Versions
We switch the top row to each new version when it comes out. Open it for one row per version. An old version stays in play while people still test it, so whoever first takes it past 85% gets credit for closing it out.
Source: Official leaderboards
More in Knowledge
- MMLU-Pro In use Harder 10-option version of MMLU with reasoning-focused questions across 14 subjects.
Intern-S2-397B 89.8%
- MMLU In use 57-subject multiple-choice exam questions from elementary to professional level. Scored by accuracy.
Seed1.8 92.3%
- SuperGPQA In use 26,529 graduate-level multiple-choice questions across 285 disciplines. Accuracy.
Qwen 3.6 Max (Preview) 73.9%
- WideSearch In use 200 bilingual broad info-seeking tasks: agent collects many atomic facts into a table. Scored by success rate and row/item F1.
Tencent Hy4 preview 83.9%
- BBEH micro average In use Models solve harder replacements for each of the 23 BIG-Bench Hard tasks, 4,520 examples in total. The score is the share of all examples answered correctly.
Gemma 4 31B IT 74.4%
- HLE-Verified In use Verified and revised version of Humanity's Last Exam (1,811 items) with uncertain items removed.
Gemini 3 Pro 48.2%
Data audit
11 of 12 checks pass, 1 problem. Look first at the Known issues check. 1 open finding: Leaked answers and reward hacking. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 5 pass
- Pass Independent runs Scores from the official leaderboard (22 entries), launch posts (5 scores) and system cards (2 scores).
- Pass Models scored 26 models we track.
- Pass Newest score The newest is from November 2026, for Inkling 2 Mini.
- Pass Frontier models 9 of the top 10 have a score, led by GPT-6.2 Astra (#1).
- Pass Headroom The best is 77.9%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on lastexam.ai: 44 rows for models we track at the last read.
- Pass Last read November 18th, 2026.
- Pass Read worked The last read worked.
Consistency 1 problem 1 pass
- Pass Lab claims The one pair agrees within 10 points.
- Problem Known issues 1 open finding: Leaked answers and reward hacking.