Epoch AI is a research institute focused on quantitative analysis of AI progress. It produces data on training compute, algorithmic efficiency, and AI capability benchmarks. Its "Notable AI Models" dataset and Epoch Capabilities Index are used as references in the field.
Organization and funding
Epoch AI was founded in 2022 as a nonprofit research organization; its co-founder and director is Jaime Sevilla, who describes the organization's mission as "improving society's understanding of the trajectory of AI" so that decisions about AI are informed by evidence (Source: epoch.ai). It funds its public data work through a mix of philanthropic funding and revenue from commissioned research, and has run paid projects for partners including Google, the UK Department for Science, Innovation and Technology, UK ARIA, the Electric Power Research Institute, and Sentinel Bio (Source: epoch.ai). Sevilla characterizes Epoch as neither an AI development company, an AI policy advocate, nor a startup incubator: the organization states it takes no official position on whether advancing AI benefits society and makes no formal policy recommendations, though staff are encouraged to offer personal views (Source: epoch.ai). Two organizations have spun out of Epoch's staff — co-founder Marius Hobbhahn left to start the AI safety organization Apollo Research, and co-founder Tamay Besiroglu, with three other employees, left to found Mechanize, a startup focused on data for automating work tasks (Source: epoch.ai).
Epoch's best-known capability measure is FrontierMath, a private mathematics benchmark commissioned by OpenAI. OpenAI is the only AI company with access to the benchmark (including an upcoming, harder Tier 4), an arrangement Epoch acknowledges has reduced outside confidence in FrontierMath scores reported for OpenAI models. After a transparency dispute over how the OpenAI funding relationship was disclosed to benchmark contributors, Epoch committed to proactively disclosing funding parties and to retaining ownership and equitable access for future benchmarks (Source: epoch.ai).
In its July 31, 2026 brief Epoch announced an expansion of the benchmark, FrontierMath: Open Problems, drawn from unsolved mathematics problems rather than problems with known solutions. The same brief carried items on constraints to parallelizability in AI research and on AI energy demand. The page was retrieved as a summary, so the problem count and the benchmark's composition are not recorded here (Source: epochai.substack.com).
Research activities
Epoch AI maintains a database of AI model training compute and publishes estimates of algorithmic and software efficiency gains, including the work in Ho et al. 2024 and Ho et al. 2025. It collaborates with METR on measuring how long the tasks are that AI can autonomously complete (time horizons). It also develops the GATE model, an economic model of AI automation with an interactive web playground, and publishes the Gradient Updates newsletter, which presents opinionated analysis of AI progress questions.
Compute and power forecasts
Epoch AI's 2025 work projected that a single frontier training run would demand 4–16 GW of power by 2030 (How Much Power Will Frontier AI Training Demand in 2030?). A related 2025 analysis examined whether AI scaling can continue through 2030 across four bottlenecks: power, chips, data, and latency (Can AI Scaling Continue Through 2030?).
In a May 25, 2026 Gradient Updates piece, Luke Emberson and Jaime Sevilla presented a calibrated model of inference capacity. By their estimate, the world's Blackwell GPUs (1.9M GB200 plus 1.5M GB300, roughly 40% of aggregate FLOP/s supply) can currently serve 500M–20B output tokens per second from a Kimi K2.6-class model depending on context length. They compared this against demand proxies (Google's 1.2B tok/s, Exponential View's 5B tok/s, and software-engineer-scaled estimates of 200M–4B tok/s) growing roughly 10×/year versus inference-capacity growth of roughly 3.4×/year, which they argued points to a near-term "compute crunch," particularly for long-context agentic workloads. They noted this is consistent with Anthropic's peak-hour Claude quota reductions and off-peak incentives (Source: epoch.ai).
Automation and capabilities analysis
Epoch AI's Gradient Updates newsletter has examined real-world task automation, including testing of how close AI is to performing specific jobs (Source: epoch.ai). Its analysis of the role of software progress in AI advancement appears in The Least Understood Driver of AI Progress.
In a June 17, 2026 Gradient Updates post, "Toward an O\NET for AI R&D," JS Denain, Joe Kwon and Anson Ho proposed tracking AI research automation through a task taxonomy modelled on the US O\NET occupational database: six categories spanning a frontier AI company's research workflow, broken into more than sixty tasks, each rated from 0 to 5 on how much the authors judge current AI automates it. The authors call it a first attempt, describe the ratings as "quite subjective," and note that Gradient Updates posts represent the authors' views rather than Epoch's as a whole (Source: epochai.substack.com). Both parts of the proposal were used in frontier-lab disclosures within three months. OpenAI classified the growth in its researchers' coding-agent tokens against the Epoch AI taxonomy, with its six phases Decide, Design, Build, Run, Analyze and Communicate, in its September 6, 2026 disclosure (Research acceleration: The view inside OpenAI (OpenAI, September 2026)). Anthropic built its prototype R&D Automation Index, published September 17, 2026, on "an automation rating scale developed by Epoch AI" running from AL0 (no AI involvement) to AL5 (fully autonomous, no human in the loop), with AL4 ("leads") meaning AI completes most of a task end-to-end from a high-level prompt while a human supervises; on that scale Anthropic reported Claude leading 26% of its AI R&D work as of August 2026, up from under 1% in February 2026. Anthropic's ratings were assigned by a Claude judge rather than by Epoch (Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026); see Anthropic R&D Automation Index). See Recursive Self-Improvement (RSI).
On July 8, 2026, Epoch AI published EBR-bench, a benchmark built on the board game Earthborne Rangers that finds little evidence AI systems learn from experience, and a Data Insight reporting that 21 notable organizations disclosed roughly 1,500 high- and critical-severity CVEs in June 2026 — more than 3.5 times the prior monthly record — following Anthropic's April announcement that Claude Mythos Preview could autonomously discover software vulnerabilities (Source: epochai.substack.com). See Autonomous cyber-agents.
In an analysis circulated July 22, 2026, Epoch AI researchers argued the OpenAI–Hugging Face evaluation-security breach was foreseeable: UK AISI evaluations had shown GPT-5.6 Sol and Claude Mythos 5 consistently compromising simulated corporate networks, and the evaluator Irregular had found GPT-5.6 Sol discovering real zero-days (Source: epochai.substack.com). See AI Autonomy Risk.
Epoch published pre-release benchmark results for GPT-6 Astra on September 12, 2026, using access OpenAI supplied before release. It placed the model first on the Epoch Capabilities Index among 247 tracked models, recorded the first solutions any model has produced on FrontierMath Erdős — 2 of 68 unsolved Erdős problems formalized in Lean — and treated FrontierMath: Tier 4 as saturated after Astra scored 97.6% (±2.4%) on Tier 4 (v2), given as 98% in the newsletter, and solved its last open problem (Source: epochai.substack.com; Source: epoch.ai).
FrontierMath Erdős was introduced the same day in a separate announcement by Tom Adamczewski and Greg Burnham, which sets out the benchmark's construction and Epoch's reason for building it (Announcing FrontierMath Erdős (Epoch AI, September 2026)). The stated complaint is that the most impressive mathematical results have come from inside AI companies that disclose little about which problems were attempted, with which models, under which scaffolds and at what cost. Epoch's answer is a fixed protocol and full disclosure: 68 problems chosen from erdosproblems.com by its curator Thomas Bloom for significance and difficulty and open as of August 2026, about 10% of the 652 then unsolved; a Lean proof or disproof checked by the Lean FRO's Comparator proof checker; one attempt per problem at $300 and 72 hours; and an open-sourced harness. Fifty of the 68 formalizations come from Google's Formal Conjectures project and Epoch directed AI to produce the other 18, stating that they passed its initial review but that "we are not Lean experts, and so errors may exist." Alongside the benchmark Epoch reported non-benchmark runs at larger budgets in which Astra solved 5 of the 68 at least once across 269 further failed attempts and over $220,000 of compute, stating three times that this is not a benchmark score.
On the same date Epoch published an analysis by Venkat Somala of Huawei's AI chip roadmap through 2031, finding Huawei's most powerful chip delivers roughly half the arithmetic performance of Nvidia's H100, which shipped from 2022, and estimating Huawei production at 0.9 million H100-equivalents in 2026 against Nvidia's 23 million — under 4% of Nvidia's AI compute (Will Huawei catch up to Nvidia by 2030? (Epoch AI, September 2026)). The report concludes that export controls constrain both levers Huawei could use to close the gap, and that although Huawei "may eventually catch up," the available information suggests it is "unlikely to happen by 2030." Its published model, confidence intervals and input parameters accompany the report. See Huawei — Ascend AI Accelerators and Export Controls (AI).
Frontier-lab revenue tracking
Josh You and Lynette Bye published an analysis on August 27, 2026 putting OpenAI's revenue run rate above $40 billion against $13 billion in August 2025, and recording that Anthropic grew from $1 billion to $9 billion across 2025 and reached a reported $65 billion run rate by the end of July 2026. Taken together the two labs tripled in 2024, grew more than fourfold in 2025, and have grown 3.5× so far in 2026, from $30 billion to $105 billion combined. The authors record a comparability caveat: where tokens are bought through cloud platforms, OpenAI books only the cut it receives while Anthropic books the full cost. They note the within-year pace may be easing, the pair having stood at roughly $55 billion combined at the end of the first quarter and taken about four months to double again. Their framing is that either frontier AI growth slows within roughly six years or the industry will have reshaped the economy, and that the open question is whether growth is driven by capability progress or by diffusion, which slows after a few years (Source: epochai.substack.com). See AI Bubble Debate.
Economic measurement
On August 24, 2026 Epoch published "The Nvidia-sized hole in US GDP statistics" by Isabel Juniewicz, Daniel Carey, Phil Trammell and Anson Ho. The report argues that US national accounts miss most of the value-add of fabless chipmakers, chiefly Nvidia, because chips designed in the US but manufactured and sold abroad are recorded as imports when they arrive in US data centers but are never recorded as US exports. Using Nvidia's US-booked operating income as a proxy, it estimates that annualized US GDP growth over the prior year was underestimated by about 0.3 percentage points, and that the gap could widen to almost two percentage points of growth per year by 2028 if Nvidia's operating income keeps growing at its current pace. Epoch states it confirmed the analysis with the Bureau of Economic Analysis; the report proposes recording such sales as goods exports, as the IMF's 2025 BPM-7 guidelines recommend, or recording fabless chipmakers' markups as IP exports (The Nvidia-sized hole in US GDP statistics (Epoch AI, August 2026)).
Relationships
- supports: The Nvidia-sized hole in US GDP statistics (Epoch AI, August 2026) — Epoch report on fabless chipmakers' value-add missing from US GDP; authors Anson Ho and Phil Trammell.
- related: Lennart Heim — compute-governance scholar whose training-compute methodology overlaps with Epoch's Notable AI Models dataset.
- related: RAND — AI's Power Requirements Under Exponential Growth (2025) — Epoch's power forecasts and RAND's converge in the same order of magnitude.
- related: Recursive Self-Improvement (RSI) — Epoch's AI R&D automation scale underlies Anthropic's R&D Automation Index and OpenAI's September 2026 classification of agent use.
- related: Measurements for understanding the pace of AI development inside frontier labs (Anthropic Institute, September 2026), Anthropic R&D Automation Index — Anthropic's September 2026 index built on Epoch's AL0–AL5 automation scale.