AI Safety Tracker

AI progress on benchmarks

18 benchmarks 6 categories to 2026-10

Scores of AI models on 18 capability benchmarks from Epoch AI, grouped by skill and by difficulty, with how long each test stayed unsolved.

How fast is AI improving?

benchmarks solved past 90.0% Counted from the month the running best score first reached 0.9. AI Benchmarking Hub Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Mixes Epoch's own evaluation runs with externally reported scores; these are not strictly comparable. Benchmark questions themselves remain the property of their creators and are not redistributed here. External rows retain their original licensing. Cite Epoch AI, "AI Benchmarking Hub". https://epoch.ai/data/ai-benchmarking-dashboard. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/progress/.

7 of 18

median months from a benchmark's first run to being solved

24

best score on FrontierMath-Erdos, the hardest open test

4.4%

solved in the last twelve months

5

Capability by category

Best score any evaluated model has reached in each area, with the benchmark it came from. Categories are ours; the scores are Epoch's.

  • Mathematics

    100.0%

    Solving stated mathematical problems with a checkable answer.

    OTIS Mock AIME 2024-2025 6 of 10 solved

  • Science knowledge

    96.4%

    Answering science questions that resist search.

    GPQA diamond 1 of 1 solved

  • Games and planning

    84.0%

    Choosing well in a game with rules and a long horizon.

    Mystery Game Puzzles 0 of 3 solved

  • Software engineering

    83.5%

    Making a real codebase pass tests it was failing.

    SWE-Bench verified 0 of 2 solved

  • Learning from experience

    78.1%

    Getting better at an unfamiliar task by repeating it.

    EBR-bench 0 of 1 solved

  • Factual accuracy

    76.0%

    Answering short factual questions without inventing an answer.

    SimpleQA Verified 0 of 1 solved

Ten of the 18 benchmarks Epoch tracks are mathematics, so this grid measures what has been built to test as much as what models can do. There is no benchmark here for long-horizon work, physical tasks or judgement under uncertainty. AI Benchmarking Hub Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Mixes Epoch's own evaluation runs with externally reported scores; these are not strictly comparable. Benchmark questions themselves remain the property of their creators and are not redistributed here. External rows retain their original licensing. Cite Epoch AI, "AI Benchmarking Hub". https://epoch.ai/data/ai-benchmarking-dashboard. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/progress/.

By who else could do it

The same benchmarks ordered by the difficulty of the questions rather than the subject. Difficulty tiers are our own reading of what each test contains.

Tier Meaning Tests Solved Best score
school Problems a strong secondary-school student could do. 1 1 98.1%
general Problems most adults could do, given the time. 4 0 84.0%
olympiad Competition problems at national olympiad standard. 1 1 100.0%
professional Real work from the job, taken from real projects. 2 0 83.5%
expert Problems that need a specialist in the field. 2 1 96.4%
research Problems that need a working researcher, sometimes hours of one. 6 4 100.0%
unsolved Problems nobody has a published answer to. 2 0 44.0%

If the current rate held

Each open benchmark's frontier score over the last 24 months, extended in a straight line to 90.0%. An extrapolation of a fitted line, not a forecast.

Benchmark Category Best now Rate Straight line hits 90.0%
SWE-Bench verified Software engineering 83.5% +2.0 pts/month 2026-07 (0mo)
MirrorCode Software engineering 77.4% +10.6 pts/month 2026-10 (0mo)
Furniture Assembly Games and planning 83.3% +4.7 pts/month 2027-01 (3mo)
Chess Puzzles Games and planning 76.0% +2.8 pts/month 2027-02 (4mo)
SimpleQA Verified Factual accuracy 76.0% +1.9 pts/month 2027-03 (5mo)
Mystery Game Puzzles Games and planning 84.0% +3.0 pts/month 2027-04 (6mo)
EBR-bench Learning from experience 78.1% +4.3 pts/month 2027-04 (6mo)
FrontierMath-2025-02-28-Private Mathematics 52.4% +2.6 pts/month 2027-07 (9mo)
OEIS Open Lite Mathematics 44.0% +2.9 pts/month 2027-10 (12mo)
FrontierMath-Tier-4-2025-07-01-Private Mathematics 47.9% +2.6 pts/month 2028-01 (15mo)
FrontierMath-Erdos Mathematics 4.4% +0.8 pts/month 2035-03 (101mo)

Two things make these dates optimistic in opposite directions, and neither is corrected for. The series fitted is a running maximum, so it can never fall and a fitted rate can never be negative: a benchmark that has stopped improving reads as zero, never as decline. And a straight line is the wrong shape, because benchmark progress is usually flat, then steep, then flat again against the ceiling, so early numbers here run late and late ones run early. A straight line is used anyway because it is the one shape that cannot be tuned to flatter the result. Rows marked flat or falling have a fitted rate of zero, meaning no evaluated model has beaten the existing best in the window. Fitted by this site from Epoch's evaluation runs. Epoch publishes no such projection. AI Benchmarking Hub Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Mixes Epoch's own evaluation runs with externally reported scores; these are not strictly comparable. Benchmark questions themselves remain the property of their creators and are not redistributed here. External rows retain their original licensing. Cite Epoch AI, "AI Benchmarking Hub". https://epoch.ai/data/ai-benchmarking-dashboard. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/progress/.

The frontier on every benchmark

Running best score by month. The dashed line is 90.0%, the Benchmark saturation The point at which the best recorded score on a test stops separating leading models. This site uses 90 per cent of the maximum score as the threshold and reports the month it was first passed. Glossary entry point, past which a test stops separating the leading models.

  • OTIS Mock AIME 2024-2025

    100.0 %

    2023-03 to 2026-10

    Solved 2025-08, 29 months

  • FrontierMath-Tier-4-2025-07-01-Public

    100.0 %

    2024-06 to 2026-06

    Solved 2026-06, 24 months

  • FrontierMath-2025-02-28-Public

    100.0 %

    2024-06 to 2026-06

    Solved 2026-02, 20 months

  • FrontierMath-Tier-4-v2-Private

    100.0 %

    2025-01 to 2026-10

    Solved 2026-06, 17 months

  • MATH level 5

    98.1 %

    2023-06 to 2025-10

    Solved 2024-12, 18 months

  • GPQA diamond

    96.4 %

    2023-03 to 2026-10

    Solved 2025-11, 32 months

  • FrontierMath-Tiers-1-3-v2-Private

    93.7 %

    2024-01 to 2026-10

    Solved 2026-09, 32 months

  • Mystery Game Puzzles

    84.0 %

    2023-06 to 2026-10

    Games and planning, general

  • SWE-Bench verified

    83.5 %

    2024-11 to 2026-06

    Software engineering, professional

  • Furniture Assembly

    83.3 %

    2025-11 to 2026-10

    Games and planning, general

  • EBR-bench

    78.1 %

    2025-08 to 2026-10

    Learning from experience, general

  • MirrorCode

    77.4 %

    2026-02 to 2026-09

    Software engineering, professional

  • Chess Puzzles

    76.0 %

    2024-02 to 2026-10

    Games and planning, expert

  • SimpleQA Verified

    76.0 %

    2024-02 to 2026-10

    Factual accuracy, general

  • FrontierMath-2025-02-28-Private

    52.4 %

    2024-06 to 2026-05

    Mathematics, research

  • FrontierMath-Tier-4-2025-07-01-Private

    47.9 %

    2024-06 to 2026-05

    Mathematics, research

  • OEIS Open Lite

    44.0 %

    2026-04 to 2026-07

    Mathematics, unsolved

  • FrontierMath-Erdos

    4.4 %

    2026-04 to 2026-09

    Mathematics, unsolved

Mixes evaluation runs performed by Epoch AI with externally reported scores, which are not strictly comparable. The line is the best available capability at each point in time, not the typical one, and one unusually strong model moves it permanently. AI Benchmarking Hub Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Mixes Epoch's own evaluation runs with externally reported scores; these are not strictly comparable. Benchmark questions themselves remain the property of their creators and are not redistributed here. External rows retain their original licensing. Cite Epoch AI, "AI Benchmarking Hub". https://epoch.ai/data/ai-benchmarking-dashboard. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/progress/.

Most recent notable models

The last 12 entries added to Epoch's model database, newest first, to 2026-10-12.

Published Model Organisation Domain Released as
2026-10-12 Muse Wonder Meta AI Language API only
2026-10-12 Muse Wonder Full Meta AI Language API only
2026-09-30 Gemini 4 Argon Google DeepMind Language Other
2026-09-29 GPT-6.1 Sol OpenAI Multimodal API only
2026-09-28 Claude Sonnet 5.5 Anthropic Language API only
2026-09-22 Claude Opus 5.5 Anthropic Language API only
2026-09-22 GPT-6 Luna OpenAI Multimodal API only
2026-09-22 GPT-6 Luna Decisions OpenAI Language API only
2026-09-22 GPT-6 Sol OpenAI Multimodal API only
2026-09-21 Grok 4.7 xAI Language API only
2026-09-18 Step 5 Preview StepFun Language API only
2026-09-09 DeepSeek V4.1 Flash DeepSeek Language Open weights

Notability is Epoch's editorial judgement, not a census, and entries are added some days after a model appears, so the very newest releases are missing rather than absent. A training-compute figure usually arrives later still and for closed models often never does, which is why this table does not carry one. Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/progress/.

Notable models released per quarter

Models meeting Epoch's notability criteria, by publication quarter since 2019.

complete quarter 2026-Q4, still in progress Notable AI Models Epoch AI Licence CC-BY-4.0 Updated weekly Retrieved , today Inclusion is editorial ("notability criteria"), not exhaustive. Compute figures are frequently estimates; check the per-row confidence and notes columns before treating any value as measured. Cite Epoch AI, "Notable AI Models". https://epoch.ai/data/ai-models. Retrieved 28 October 2026 via AI Safety Tracker, https://aisafetytracker.org/progress/.

Counts models Epoch judged notable, which is a curation decision rather than a census. A rise partly reflects more models and partly reflects more attention being paid to them. The final bar covers a quarter that has not finished.

Related

  • Capability Training compute, cost and release type for the same models.
  • Release disclosures What each developer published about a model's safety when it shipped.
  • Developers The eight frontier developers, model by model.