How fast is AI improving?
Two measured quantities describe how fast AI is improving: the scores models reach on tests built to measure what they can do, and the amount of computation used to train them. Of the 18 tests Epoch AI tracks, 7 have been solved, a median of 24 months after each was first run, and the computation used for the largest training run to date has doubled about every 6 months since 2015. Benchmark scores and compute estimates from Epoch AI. The doubling time is fitted by this site to Epoch's compute figures. AI Benchmarking Hub Epoch AI 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/explainers/how-fast-is-ai-improving/. Notable AI Models Epoch AI 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/explainers/how-fast-is-ai-improving/.
The answers below use only figures published on the progress and capability pages, where each is set out with its source. Every figure describes what has already been measured.
How is progress in AI measured?
The usual measure is a benchmark: a fixed set of questions or tasks with answers that can be checked, on which a model's score is the share it gets right. Epoch AI records, for each benchmark and each month, the best score any model it has evaluated has reached. That record is a Running maximum The highest value recorded up to each date. It rises or holds level and never falls, so a line fitted through one cannot slope downward. Glossary entry , so it rises or holds level and never falls, and it shows the best available model, which is not the typical one. It combines Epoch's own evaluations with scores reported by others, which are not strictly comparable. AI Benchmarking Hub Epoch AI 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/explainers/how-fast-is-ai-improving/.
A test stops being useful once the best models all score close to full marks, because it can no longer tell them apart. This site counts a benchmark as 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 , or solved, from the month its best score first reached 90%.
How quickly are AI benchmarks being solved?
Of the 18 benchmarks, 7 have passed 90%, five of them in the last twelve months. Measured from each one's first recorded run to the month it was solved, the median is 24 months. Counted from the month the running best score first reached 0.9. AI Benchmarking Hub Epoch AI 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/explainers/how-fast-is-ai-improving/.
- OTIS Mock AIME 2024-2025, mathematics at the olympiad level, solved in August 2025, 29 months after its first run.
- FrontierMath-Tier-4-2025-07-01-Public, mathematics at the research level, solved in June 2026, 24 months after its first run.
- FrontierMath-2025-02-28-Public, mathematics at the research level, solved in February 2026, 20 months after its first run.
- FrontierMath-Tier-4-v2-Private, mathematics at the research level, solved in June 2026, 17 months after its first run.
- MATH level 5, mathematics at the school level, solved in December 2024, 18 months after its first run.
- GPQA diamond, science knowledge at the expert level, solved in November 2025, 32 months after its first run.
- FrontierMath-Tiers-1-3-v2-Private, mathematics at the research level, solved in September 2026, 32 months after its first run.
The hardest benchmark still open is FrontierMath-Erdos, on which the best score is 4.4%. AI Benchmarking Hub Epoch AI 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/explainers/how-fast-is-ai-improving/.
What can AI models do now?
Grouped by the skill each benchmark tests, the best score reached so far in each area is as follows. The grouping is this site's; the scores are Epoch's. AI Benchmarking Hub Epoch AI 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/explainers/how-fast-is-ai-improving/.
- Mathematics, solving stated mathematical problems with a checkable answer: 100.0% on OTIS Mock AIME 2024-2025, with 6 of 10 benchmarks solved.
- Science knowledge, answering science questions that resist search: 96.4% on GPQA diamond, with 1 of 1 benchmark solved.
- Games and planning, choosing well in a game with rules and a long horizon: 84.0% on Mystery Game Puzzles, with 0 of 3 benchmarks solved.
- Software engineering, making a real codebase pass tests it was failing: 83.5% on SWE-Bench verified, with 0 of 2 benchmarks solved.
- Learning from experience, getting better at an unfamiliar task by repeating it: 78.1% on EBR-bench, with 0 of 1 benchmark solved.
- Factual accuracy, answering short factual questions without inventing an answer: 76.0% on SimpleQA Verified, with 0 of 1 benchmark solved.
Ten of the 18 benchmarks are mathematics, so the set measures what has been built to test as much as what models can do. None of them tests long-horizon work, physical tasks or judgement under uncertainty. AI Benchmarking Hub Epoch AI 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/explainers/how-fast-is-ai-improving/.
How hard are the problems AI can solve?
The same benchmarks can be ordered by who else could answer their questions, from problems a strong secondary-school student could do to problems nobody has a published answer to. The difficulty levels are this site's reading of what each test contains. AI Benchmarking Hub Epoch AI 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/explainers/how-fast-is-ai-improving/.
- School, problems a strong secondary-school student could do: best score 98.1%, 1 of 1 solved.
- General, problems most adults could do, given the time: best score 84.0%, 0 of 4 solved.
- Olympiad, competition problems at national olympiad standard: best score 100.0%, 1 of 1 solved.
- Professional, real work from the job, taken from real projects: best score 83.5%, 0 of 2 solved.
- Expert, problems that need a specialist in the field: best score 96.4%, 1 of 2 solved.
- Research, problems that need a working researcher, sometimes hours of one: best score 100.0%, 4 of 6 solved.
- Unsolved, problems nobody has a published answer to: best score 44.0%, 0 of 2 solved.
At the easiest level the best score is 98.1%, and at the hardest it is 44.0%. Of the six research-level tests, which need a working researcher, four have been solved. AI Benchmarking Hub Epoch AI 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/explainers/how-fast-is-ai-improving/.
How fast is the computing power behind AI growing?
The computation used to train a model is measured in FLOP Floating-point operation. Training compute is reported as the total number of operations used in the final training run, on a logarithmic scale, so each axis step is ten times the one below it. Glossary entry , the number of arithmetic operations performed, and written in powers of ten: 1e25 is a one followed by 25 zeros, and each step from one power to the next is ten times as much. Epoch AI estimates it for 540 notable models. The largest training run it records is GPT-6 Astra, published by OpenAI in September 2026, at 9.4e26 FLOP. Compute is Epoch AI's estimate, not a disclosure. Notable AI Models Epoch AI 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/explainers/how-fast-is-ai-improving/.
A straight line fitted to the largest run to date, from 2015 onward, doubles about every 6 months. It describes what has already happened, and only the models that set a new record enter it, which is a small number of points. Epoch records 24 models at or above 1e25 FLOP, the line above which the EU AI Act presumes a model carries systemic risk. Fitted by this site to Epoch's compute figures, not a figure Epoch publishes. Notable AI Models Epoch AI 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/explainers/how-fast-is-ai-improving/.
Many of these figures are Epoch's estimates rather than disclosures, and Epoch marks its confidence in each. A model whose compute is unknown does not appear at all, which leans the record toward developers that publish.
How far behind are openly released AI models?
A model released with Open weights A release in which the trained parameters are distributed and can be run by a third party. It does not imply an open-source licence, and the licence attached to a given release may restrict use. Glossary entry can be downloaded and run by anyone, and once published the release cannot be withdrawn. Measured by training compute, the largest open-weight model is Llama 3.1-405B, at 3.8e25 FLOP. The largest model of any kind first reached that level 8 months before Llama 3.1-405B was published, and the open-weight record has not moved in the 27 months since. GPT-6 Astra used about 25 times as much compute. Measured horizontally: the date the open-weight record was set, minus the date the first model of any kind reached that compute level. Computed by this site from Epoch's figures. Notable AI Models Epoch AI 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/explainers/how-fast-is-ai-improving/. Notable AI Models Epoch AI 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/explainers/how-fast-is-ai-improving/.
How many new AI models are released?
Epoch AI counted 35 Notable model A model meeting Epoch AI's notability criteria, which is the population used for every model count on this site. Glossary entry published in the third quarter of 2026, against 19 in the third quarter of 2021. Notability is Epoch's editorial judgement, so the count reflects how many models there are and how much attention is paid to them. Notable AI Models Epoch AI 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/explainers/how-fast-is-ai-improving/.
What these measures cannot show
- How fast AI will improve. Every figure on this page describes what has already been measured.
- How the typical model performs. Each score is the best any evaluated model has reached, and one unusually strong model moves it permanently.
- Skills no benchmark tests, among them long-horizon work, physical tasks and judgement under uncertainty.
- The compute of models whose developers publish nothing, which Epoch can estimate only from what is known about them, or not at all.
The benchmark-by-benchmark record, with a frontier line for every test, is on the progress page. AI Benchmarking Hub Epoch AI 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/explainers/how-fast-is-ai-improving/. Notable AI Models Epoch AI 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/explainers/how-fast-is-ai-improving/.