Epoch AI
An AI research organization focused on capability evaluation and benchmark design. It maintains frontier benchmarks such as FrontierMath and continuously tracks trends in AI model development.
Core Facts
AI inference prices, measured per unit of compute, have dropped approximately 10-fold over the past two years, a phenomenon the industry calls the AI inference price spiral.
According to estimates by Epoch AI research, high-quality text data available on the internet for training will be largely exhausted by around 2026.
Timeline (last 90 days)
In the whatmodel.sh data, 14,436 results come from independent evaluations (ACL Anthology, arXiv, Epoch AI, LMArena, EvalPlus, EleutherAI harness, etc.) and 7,677 from lab self-reports, with independent evaluations accounting for about 77%.
According to research by institutions such as Epoch AI, the total amount of high-quality text available on the public internet for training (such as books, academic papers, and high-quality web content) is on the order of tens of billions of words.
Epoch AI's InnovationEval is used to evaluate whether a model can independently produce new algorithms—from conception and implementation to experimentation—that are comparable to techniques subsequently discovered by humans.
Claude Opus 5.5, released by Anthropic, scored 167 on Epoch AI's ECI benchmark, ranking first in the world.
GPT-6 Astra peaked at 98% on Frontier Math Tier 4, and still reached 83% without enabling reasoning or using chain-of-thought.
When releasing FrontierMath, Epoch AI noted that the difficulty of its problems corresponds to scenarios that only PhD-level mathematics graduate students or professional researchers can handle.
Epoch AI estimates that high-quality English text data may be exhausted by mainstream LLM training around 2026.
According to Epoch AI's 2026 statistics, the gap between open-source models and the most powerful closed-source models averages only about 4 months.
According to Epoch AI calculations, Google's TPU solution achieves approximately 2,039 H100-equivalent FLOPS/MW, while NVIDIA's full B200 and B300 system solutions achieve approximately 1,175 H100-equivalents/MW, measuring the energy efficiency of the entire data center system.
Epoch AI estimates that Google, as a single vendor, holds approximately one-quarter of the world's total AI compute.
13 more timeline events
All Facts (20)
AI inference prices, measured per unit of compute, have dropped approximately 10-fold over the past two years, a phenomenon the industry calls the AI inference price spiral.
65%VerifiedAccording to estimates by Epoch AI research, high-quality text data available on the internet for training will be largely exhausted by around 2026.
65%UnverifiedThe training data cutoff dates of many open-source models remain at 2023 or even earlier.
70%UnverifiedIn the whatmodel.sh data, 14,436 results come from independent evaluations (ACL Anthology, arXiv, Epoch AI, LMArena, EvalPlus, EleutherAI harness, etc.) and 7,677 from lab self-reports, with independent evaluations accounting for about 77%.
50%UnverifiedAccording to research by institutions such as Epoch AI, the total amount of high-quality text available on the public internet for training (such as books, academic papers, and high-quality web content) is on the order of tens of billions of words.
50%UnverifiedEpoch AI's InnovationEval is used to evaluate whether a model can independently produce new algorithms—from conception and implementation to experimentation—that are comparable to techniques subsequently discovered by humans.
50%UnverifiedClaude Opus 5.5, released by Anthropic, scored 167 on Epoch AI's ECI benchmark, ranking first in the world.
50%UnverifiedGPT-6 Astra peaked at 98% on Frontier Math Tier 4, and still reached 83% without enabling reasoning or using chain-of-thought.
50%UnverifiedWhen releasing FrontierMath, Epoch AI noted that the difficulty of its problems corresponds to scenarios that only PhD-level mathematics graduate students or professional researchers can handle.
50%UnverifiedAccording to Epoch AI's 2026 statistics, the gap between open-source models and the most powerful closed-source models averages only about 4 months.
50%UnverifiedAccording to Epoch AI calculations, Google's TPU solution achieves approximately 2,039 H100-equivalent FLOPS/MW, while NVIDIA's full B200 and B300 system solutions achieve approximately 1,175 H100-equivalents/MW, measuring the energy efficiency of the entire data center system.
50%UnverifiedEpoch AI estimates that Google, as a single vendor, holds approximately one-quarter of the world's total AI compute.
50%UnverifiedEpoch AI has announced that all problems in Frontier Math Tier 4 have been solved by AI, with GPT-6 Astra solving the last unsolved problem.
50%UnverifiedFrontier Math Tier 4 went from a top score of under 5% at launch to a top score of 98% in under 14 months, leading Epoch to conclude that the benchmark has been saturated.
50%UnverifiedFrontier Math is an evaluation benchmark maintained by Epoch AI, with problems designed by professional mathematicians that deliberately avoid solution paths already available on the internet to reduce the possibility of models answering from memorization.
50%UnverifiedFrontierMath is a benchmark designed specifically to measure AI's advanced mathematical reasoning capabilities.
50%UnverifiedAn Epoch AI analysis shows that based on actual chip purchase data from each quarter between 2023 and 2025, the AI compute purchasable per $100 million per year has increased by approximately 49%, equivalent to doubling every 21 months.
50%UnverifiedAccording to Epoch AI data, between 2023 and 2024, startups that included 'AI' in their company name or product description had a median seed-round funding approximately 40% higher than comparable non-AI companies.
50%UnverifiedECI (Effective Capability Index) is a method that converts model capabilities into a temporal progress measure, with its design philosophy drawing on the concept of purchasing power parity in economics.
50%UnverifiedResearch data from Epoch AI shows that the computing power used for AI training has grown by more than 300,000 times from 2012 to the present.
50%