Humanity's Last Exam Diamond (HLE-Diamond)
We show this table for reference; we do not rank on it.
Accuracy on the 1,000-question HLE-Diamond set without tools, using the benchmark owner's high-reasoning evaluation.
Accuracy without tools at high reasoning effort on HLE-Diamond — September 22, 2026 release table
We mirror the published accuracy without tools at high reasoning effort view for HLE-Diamond. GPT-6 Astra leads the public snapshot at 60.6%, followed by Claude Opus 5.5 (55.0%) and Claude Fable 5.1 (51.3%). We do not use these results to rank models overall.
GPT-6 Astra
OpenAI
No tools · high reasoning
Claude Opus 5.5
Anthropic
No tools · high reasoning
Claude Fable 5.1
Anthropic
No tools · high reasoning
9 modelsKnowledgeCurrentDisplay onlyUpdated September 22, 2026 release table
Accuracy without tools at high reasoning effort table (9 models)
ScoreHow to read this leaderboard
HLE-Diamond is a 1,000-question refined subset of Humanity's Last Exam. The published no-tools table evaluates each listed model at high reasoning effort.
Variant check: This is a different question set from original HLE, HLE-Rolling, and HLE-Verified. Highest-effort no-tools and web-and-code runs have separate tables.
Compare rows within this no-tools table. The equal-sized reasoning and knowledge partitions make the overall score an average across two different kinds of question.
Operator receipt: 9 sourced rows are currently displayable on this page; the leading published row is GPT-6 Astra at 60.6%.
Provenance: The scores mirror the Center for AI Safety and Scale AI release table. We have not rerun the gated question set.
Source freshness: The benchmark and these scores were published on September 22, 2026.
Honest limit: The dataset requires access approval, and the release does not publish uncertainty intervals for these model scores. A small score gap needs more evidence before it changes a model choice.
How to read the HLE-Diamond no-tools table
The September 22, 2026 HLE-Diamond release reports nine models without tools at high reasoning effort. The 1,000-question subset contains 500 reasoning and 500 knowledge questions.
We keep these benchmark-owner results separate from full-set HLE, the highest-effort no-tools view, and web-and-code runs. This table is display only and does not move model rankings.
Snapshot
The published HLE-Diamond snapshot places GPT-6 Astra first at 60.6%. The third row is 9.3 points behind. The broader top-10 range is 37.2 points, so the table still separates the published systems.
9 models have been evaluated on HLE-Diamond. The benchmark falls in the Knowledge category. HLE-Diamond is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
About HLE-Diamond
Year
2026
Tasks
1,000 expert questions: 500 reasoning and 500 knowledge
Format
Accuracy without tools, reasoning high
Difficulty
Frontier expert knowledge and reasoning
The September 22, 2026 release reports nine model results. HLE-Diamond contains 500 reasoning and 500 knowledge questions drawn from a refined HLE subset. The no-tools results remain separate from both original HLE and the web-and-code evaluation.
Freshness and provenance
Version
HLE-Diamond, reasoning high without tools
Refresh cadence
Static
Staleness state
Current
Question availability
Gated 1,000-question dataset; public aggregate results
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
What does HLE-Diamond measure?
Accuracy on the 1,000-question HLE-Diamond set without tools, using the benchmark owner's high-reasoning evaluation.
Which model leads the published HLE-Diamond snapshot?
GPT-6 Astra currently leads the published HLE-Diamond snapshot with 60.6% accuracy without tools at high reasoning effort. BenchLM shows this benchmark for display only and does not use it in overall rankings.
How many models are evaluated on HLE-Diamond?
The September 22, 2026 release table snapshot contains 9 AI models.
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 2,000+ readers.
One email each week. Unsubscribe anytime.