What the AI Capability Score Means (and What It Doesn't)
One number for how capable a model is, from an independent research group. Here's what it measures, what the range means, and how to use it without being misled.
Short answer: the capability score on our pages is the Epoch Capabilities Index, from Epoch AI, an independent research group. It combines results from over 50 benchmarks into one scale, where higher means more capable on average. Claude Fable 5.2 scores 174 (range 170–179), Claude Opus 5.5 169 (range 166–172) and Claude Haiku 5.5 160 (range 158–163). Read the range, not just the number: two models are only level when each score sits inside the other's range.
What the score measures
No single benchmark covers everything, and the best models max out old benchmarks within months. Epoch's index gets round both problems. It works out how hard each benchmark is by comparing models that took several of them, then places every model on one scale: a model scores higher when it does well on harder benchmarks.
The number itself is abstract. A score of 169 isn't "169% of something", and the gap between 150 and 160 doesn't mean a model is 7% better. What matters is the order and the size of the gap compared with the uncertainty. Among the models we price today, scores run from 135 (gpt-oss-20b) to 174 (Claude Fable 5.2).
Why there's a range, and what "level" means
Each score comes with a range that shows Epoch's uncertainty. Claude Opus 5.5's is 169 (range 166–172). A model scored on fewer benchmarks usually has a wider range.
On this site, two models are level only when each one's score falls inside the other's range. Ranges that merely touch aren't enough:
- Level: Claude Sonnet 5.5 at 166 (range 164–169) and Claude Opus 5.5 at 169 (range 166–172). Each score sits inside the other's range, so on general capability they're a match, and Sonnet costs half as much per token.
- Not level: Claude Haiku 5.5 at 160 (range 158–163) and Claude Sonnet 5.5 at 166 (range 164–169). The ranges come close, but Sonnet scores clearly higher.
We use the same rule everywhere: on comparison pages, in the "level or higher for less" notes on model pages, and on the value chart on all models.
What the score doesn't tell you
- How a model does on your task. It's an average across many kinds of work. A model can lead on coding and trail on writing; Epoch notes that a model can top some benchmarks without having the highest overall score.
- What a task costs. Models use very different numbers of tokens to finish the same job, so a cheaper price per token doesn't always mean a cheaper task. Our agent cost calculator shows how much input piles up on a long task.
- Speed, context handling and reliability with tools. None of these are in the score.
- Anything about brand-new models. A model only gets a score once Epoch has enough results for it.
How to weigh capability against price
- Pick the level you need. Start from a model you know does the job, and look for models that score level with it or higher.
- Compare cost on your workload, not price per token. Caching, long prompts and output length change the order. In our coding-agent example, GPT-6.1 Terra scores level with Claude Opus 5.5 and costs far less to run.
- Test the shortlist on your own work. The score narrows the field; your own prompts decide.
Every comparison page now puts the two together: price per token, cost on a coding agent, the capability verdict, and the cheapest model at the stronger one's level. Try Kimi K3 vs Claude Opus 5.5, or build your own on the comparison tool.
Why some models aren't scored
We record a score only when we're sure Epoch's model is the one we price. Names that differ only in punctuation or spacing match automatically. Anything else, such as a snapshot date, needs evidence from the provider's own page or the model's API ID. When Epoch lists two versions and we can't tell which one the API serves, we leave the model unscored rather than guess. Small models and very new releases are often not on the index yet. Unscored models say "Not yet scored", and we never estimate one.
Why Epoch AI's index
It's independent rather than published by a model maker, it shows its uncertainty, and its data is free to reuse with credit under the CC BY 4.0 licence. We refresh the scores from Epoch's data at every check-in. Model makers' own benchmark tables appear in our launch articles, always labelled as theirs, but never in our comparisons, because each maker chooses the tests that suit it.
Questions
What is a good capability score?
It's relative. Among current models we price, scores run from 135 to 174. Most everyday work doesn't need the top of the scale: Claude Haiku 5.5 and GPT-6 Luna score around 160 at a fraction of the price of the leaders.
Is the highest-scoring model always worth it?
No. Claude Fable 5.2 scores highest, at $8.00 / $40.00 per 1M tokens. For most tasks a model that scores a little lower costs several times less. Pay for the top score when a cheaper model can't do the job.
How often do scores change?
Epoch updates the index as new benchmark results come in, and scores often move by a point or so. We refresh them at each check-in.
Source: Epoch Capabilities Index by Epoch AI, CC BY 4.0, from Epoch's data download (11 November 2026). Scores and prices in this guide update automatically.