Science benchmarks
Quantitative science and reasoning performance on graduate-level evaluations. Every score is sourced and citable — see the provenance notes below before comparing across rows or variants.
Frequently asked questions
- What is HLE (no tools)?
- Humanity's Last Exam (HLE) is a 2,500-question benchmark spanning dozens of academic subjects — mathematics, humanities, and the natural sciences — created because earlier benchmarks like MMLU had become saturated (frontier models were scoring over 90%). Each question has a single, unambiguous, expert-verified answer that can't be found via a quick web search; 80% are exact-match, the rest multiple-choice. This page tracks two variants on the same question set — 'no tools' (closed-book, the anchor benchmark here) and 'with tools' (web search, calculator, and code execution enabled) — plus ARC-AGI-2, a separate abstract visual-reasoning benchmark. See the sourcing notes below for how the variants and ARC-AGI-2 compare.
- Which models are currently tracked?
- 16 models: Claude Opus 5.5, Claude Sonnet 5.5, Claude Opus 5, GPT-5.5, Claude Sonnet 5, Claude Opus 4.8, GPT-5.4 mini, Gemini 2.5 Pro, Claude Sonnet 4.6, o3, DeepSeek R1, Gemini 2.5 Flash, Claude Sonnet 4, Claude 3.5 Sonnet, GPT-4o, Gemini 3.5 Flash.
- Which model scores highest on HLE (no tools)?
- Claude Opus 5.5, at 64.4% as of 2026-09.
- How often is the data updated?
- This page was last rebuilt October 7, 2026 from Modelglass's science benchmark registry. Scores are appended as new evaluations are verified — an existing result is never edited or deleted, only superseded by a newer entry.
- Where does the score come from?
- Every score below links to its source (the ↗ icon next to each result) and shows an icon for how it was obtained: 🏢 vendor-reported (the model developer's own announcement), 📊 public leaderboard (a public, independently-maintained evaluation run — for HLE, typically the CAIS AI Safety Dashboard), 🔬 independent evaluation (a credible third party ran the benchmark), or 📄 published paper (a peer-reviewed or arXiv result). See the sourcing notes below the table for the HLE tool-variant comparison, the vendor-vs-leaderboard score gap, and ARC-AGI-2's compute sensitivity.
Who this is for
This page is built for researchers, model evaluators, and technically-minded readers who want a domain-agnostic read on frontier reasoning and knowledge capability, not a specific task outcome. Humanity's Last Exam was created explicitly to give researchers and policymakers a benchmark that hasn't saturated the way MMLU and similar exams have; ARC-AGI-2 serves a related but distinct audience — those studying whether a model reasons and generalizes rather than pattern-matches its way to an answer.
What an HLE score means
HLE (no tools) is the percentage of 2,500 expert-written, deliberately search-resistant academic questions — across 100+ subjects — a model answers correctly closed-book (HLE paper ↗). Subject-matter experts score around 90%; at the benchmark's 2025 release the best models were near 3%. The bands below are our reading of what a given score implies — HLE has no official interpretation guide, and its authors are explicit that it measures closed-ended academic knowledge, not open-ended research, creativity, or autonomy.
| Score range | What it means | Best used for |
|---|---|---|
| ≥ 50% | Answers more than half of HLE's closed-book expert questions — a level first reached in 2026, still far below the ~90% experts achieve. | The current ceiling for broad, tool-free graduate-level knowledge and reasoning. Treat as a strong generalist, not a verified domain expert — it is still wrong on roughly half. |
| 30–50% | Answers between 1 in 3 and 1 in 2 of the closed-book questions. | Wide knowledge across fields, but the failure rate is high enough that any specialist-domain claim needs checking against a source before you rely on it. |
| 15–30% | Answers between 1 in 7 and 1 in 3. | A capable generalist that is unreliable at the graduate/research level — useful for orientation and framing questions, not for answers you will act on unverified. |
| 5–15% | Answers fewer than 1 in 7. | Below the level where a closed-book answer to an expert question carries weight; the model's value on hard technical material is in explaining and structuring it, not in being right. |
| < 5% | At or near the benchmark's floor — roughly 1 in 20 or fewer. | No signal of tool-free graduate-level capability; earlier-generation or smaller models sit here. |
Before you rely on a band
- This is the closed-book variant — the same models score 10–20 points higher with web search and code execution enabled (the HLE with-tools column). A low no-tools score is a recall limit, not necessarily a reasoning one.
- The questions are adversarial by design — each was filtered to be unanswerable by a quick web search and to stump then-frontier models. HLE accuracy is not a proxy for everyday factual reliability, which is far higher.
- Models are poorly calibrated on HLE — its authors report low accuracy and low calibration, i.e. confident wrong answers. A model in any band will still state incorrect answers with full confidence.
- It is a young, fast-moving benchmark — best scores went from ~3% at release (2025) to the mid-50s. Treat a band as a snapshot and re-check where a model sits rather than trusting a months-old figure.
- Narrow question shape — short academic questions with one verifiable answer. Says little about long-form work, open-ended research, or applied judgement.
| Model | |||
|---|---|---|---|
| Claude Opus 5.5 anthropic/claude-opus-5-5 | 64.4% 🏢 ↗ 2026-09 | 67.7% 🏢 ↗ 2026-09 | — |
| Claude Sonnet 5.5 anthropic/claude-sonnet-5-5 | 56.9% 🏢 ↗ 2026-09 | 64.5% 🏢 ↗ 2026-09 | — |
| Claude Opus 5 anthropic/claude-opus-5 | 56.3% 🏢 ↗ 2026-07 | 64.7% 🏢 ↗ 2026-07 | 90.4% 🔬 ↗ 2026-07 |
| GPT-5.5 openai/gpt-5.5 | 43.8% 📊 ↗ 2026-09 | — | — |
| Claude Sonnet 5 anthropic/claude-sonnet-5 | 43.2% 🏢 ↗ 2026-06 | 57.4% 🏢 ↗ 2026-06 | — |
| Claude Opus 4.8 anthropic/claude-opus-4-8 | 42.2% 📊 ↗ 2026-07 | 57.9% 🏢 ↗ 2026-06 | — |
| GPT-5.4 mini openai/gpt-5.4-mini | 23.5% 📊 ↗ 2026-07 | — | — |
| Gemini 2.5 Pro google-deepmind/gemini-2.5-pro | 21.6% 📊 ↗ 2026-07 | — | 3.8% 🔬 ↗ 2025-05 |
| Claude Sonnet 4.6 anthropic/claude-sonnet-4-6 | 21.1% 📊 ↗ 2026-07 | 46.8% 🏢 ↗ 2026-06 | — |
| o3 openai/o3 | 20.3% 📊 ↗ 2026-07 | — | 6.5% 🔬 ↗ 2025-05 |
| DeepSeek R1 deepseek/deepseek-r1 | 17.7% 🏢 ↗ 2025-05 | — | 1.1% 🔬 ↗ 2025-05 |
| Gemini 2.5 Flash google-deepmind/gemini-2.5-flash | 12.1% 📊 ↗ 2026-07 | — | — |
| Claude Sonnet 4 anthropic/claude-sonnet-4 | 9.6% 📊 ↗ 2026-07 | — | 5.9% 🔬 ↗ 2025-05 |
| Claude 3.5 Sonnet anthropic/claude-3-5-sonnet | 8.9% 📄 ↗ 2025-01 | — | — |
| GPT-4o openai/gpt-4o | 2.7% 📊 ↗ 2026-07 | — | — |
| Gemini 3.5 Flash google-deepmind/gemini-3.5-flash | — | — | 72.1% 🔬 ↗ 2026-05 |
See how benchmark data like this can power a cost-aware routing setup
See how →Why we track this
We track HLE (no tools) because it's a genuine differentiator between models — not a vanity metric. Used alongside the other benchmarks on this page and pricing data elsewhere in the registry, it's designed to help you find the most appropriate model for your specific use case, not just the highest-scoring one.
Sourcing and comparability
Source types: 🏢 vendor-reported scores come directly from the model developer and may use proprietary evaluation setups. 📊 leaderboard scores are from public, independently-maintained evaluation runs. 🔬 independent scores are from credible third-party evaluations. 📄 paper scores are from peer-reviewed or arXiv publications.
HLE variant warning: HLE (no tools) and HLE (with tools) are separate evaluations on the same question corpus and must not be compared directly. With-tools scores are systematically 10–20 percentage points higher than no-tools scores for the same model. GPT-4o and Claude 3.5 Sonnet entries show only no-tools scores from the original January 2025 benchmark paper — they are historical baselines, not current performance figures.
Vendor vs. leaderboard gap: Some models have both a vendor-reported score and a CAIS leaderboard score for the same benchmark. These differ because vendor announcements often use extended thinking, optimised prompting, or internal test sets, while the CAIS AI Dashboard uses standardised conditions across all models. The table shows the most recent entry per benchmark — check the source badge and click the ↗ link to see which evaluation applies. For Anthropic models the typical gap is 7–14 percentage points.
ARC-AGI-2 compute sensitivity: ARC-AGI-2 scores are far more compute-sensitive than HLE. The same model can score 2–4× higher with extended thinking tokens or a higher-compute configuration vs. the default API. The scores shown here are from the ARC Prize Foundation's "We tested every major AI reasoning system" evaluation at ARC-AGI-2's May 2025 launch, which used matched compute settings across models. Only 4 of the 12 tracked models were included in that evaluation — the remaining cells show "—". Human average on ARC-AGI-2 is approximately 66%.
Data is growing: This table is sourced from Modelglass's science benchmark registry, a registry of science and reasoning benchmark results. Scores are append-only — once recorded, a result is never edited or deleted.