Epoch Capabilities Index
In plain wordsA single number summarizing an AI model's general capability, calculated by the research group Epoch AI by combining its scores on dozens of tests onto one common scale.
A composite score of general AI capability that Epoch AI computes by fitting results from more than 50 benchmarksbenchmark, n.: A fixed, shared test — a set of questions or tasks with an agreed way of scoring — used to compare AI systems with one another and to track progress over time. onto one shared scale, so that models released years apart can still be compared after individual benchmarks have saturated. Each benchmark is modeled by a logistic curve with its own difficulty and slope, and benchmark difficulties and model capabilities are estimated jointly from overlapping results rather than set by hand — an approach akin to item response theory in psychometrics. A model needs results on at least four benchmarks to receive a score. The scale has no maximum and is anchored so that Claude 3.5 Sonnet scores 130 and GPT-5 scores 150. Epoch uses it to track the pace of frontier progress and the gap between open-weight and closed models; as of 1 October 2026 the top-ranked model, Claude Opus 5.5, scored 167.
On the Epoch Capabilities Index, the best open-weight model trails the frontier by about four points.
Origin
Introduced by Epoch AI in late 2025; its statistical method is set out in 'A Rosetta Stone for AI Benchmarks' (Anson Ho and colleagues, November 2025), written with Google DeepMind researchers and funded by Google DeepMind.
Usage note
Like an IQ score, the index compresses a jagged profile into one number: models with equal scores can differ sharply on particular skills, and results depend on which benchmarks are included.
Related
- benchmarkbenchmark, n.: A fixed, shared test — a set of questions or tasks with an agreed way of scoring — used to compare AI systems with one another and to track progress over time.
- benchmark saturationbenchmark saturation, n.: When the best AI systems score so close to the maximum on a test that it can no longer tell them apart or show further progress.
- construct validityconstruct validity, n.: Whether high marks on a test actually reflect the skill the test is named after — for example, whether a 'reasoning' test rewards reasoning rather than memory or test-taking tricks.
- jagged intelligencejagged intelligence, n.: The uneven pattern of AI abilities: today's systems can solve very hard problems yet fail at some a child could manage, so their skills form a spiky profile rather than a smooth one.
- algorithmic progressalgorithmic progress, n.: Improvements in AI methods — not faster chips — that let a model reach the same ability with less computing power. Researchers estimate this has cut the computing needed by half every eight to sixteen months.
- AI timelinesAI timelines, n.: Forecasts of when AI will reach major milestones, such as matching humans at most work. People with short timelines expect this within years; others expect decades, or never.
Sources
- Epoch AI. Epoch Capabilities Index. epoch.ai/benchmarks/eci
- Ho, A., Denain, J.-S., Atanasov, D., Albanie, S. & Shah, R. (2025). A Rosetta Stone for AI Benchmarks. arXiv:2512.00193. arxiv.org/abs/2512.00193