LLM benchmark: model rankings, scores and prices

How to use it

The table lists the 164 language models for which Epoch AI publishes at least one score, with their public price taken from OpenRouter on October 25, 2026. Start by picking the measure that matches your use: the overall capabilities index for a broad view, SWE-bench Verified or Terminal-Bench if you need a model for a coding agent, GPQA Diamond for scientific reasoning, ARC-AGI for the ability to adapt to never-seen problems.

• Then filter by vendor, by licence type (open weights you can run yourself, or proprietary models available through an API), by minimum context size and by maximum blended price.
• The chart redraws at every change. Hover or tap a dot to read the name, price and score; click to open the model's page.
• Click a table header to sort (click again to reverse). The last column always shows the chosen measure.

Everything runs in your browser: nothing is sent anywhere, and the page keeps working offline once loaded.

What the selected benchmarks measure

The Epoch AI Capabilities Index (ECI) combines many benchmarks into a single number, on a scale that makes it possible to compare models that did not all take the same tests. 125 models in the table have one; the highest on October 25, 2026 belongs to Claude Opus 5.5 (168.8). Individual benchmarks say more specific things:

• GPQA Diamond : PhD-level science questions (124 models scored)
• SWE-bench Verified : fixing real GitHub issues (28 models scored)
• Terminal-Bench : real tasks in a terminal (37 models scored)
• ARC-AGI-1 : visual abstract-reasoning puzzles (73 models scored)
• ARC-AGI-2 : harder version of the ARC puzzles (73 models scored)
• FrontierMath : research-level maths problems (tiers 1-3) (85 models scored)
• OTIS Mock AIME : maths competition problems (124 models scored)
• SimpleQA Verified : accuracy on simple facts (80 models scored)
• Aider Polyglot : coding exercises in several languages (25 models scored)
• OSWorld : operating a desktop computer (6 models scored)

Scores are success rates. A gap of two or three points between two models is often within the test's margin of error: look instead at clear gaps and at consistency across several benchmarks close to your use. To understand why ARC-AGI remains hard for models, see the glossary entry.

Reading the score / price chart

The blended price weighs three input tokens for one output token, which matches the typical use of an assistant or an agent: you send a lot of context (files, history, tool results) and receive a shorter answer. The horizontal scale is logarithmic, because prices range from a few cents to several tens of dollars per million tokens.

Orange dots form the Pareto frontier: for each of them, no other displayed model is both cheaper and better scored on the chosen measure. That is where to look for the best value. A model far below the frontier costs more than another that does better; it may still be worth it for a reason the chart does not show (longer context, open weights, hosting in a given region, built-in tool).

To go from a price per million tokens to a real monthly budget, use the Lab's price calculator, and the guide LLM API pricing to understand billing.

Where the data comes from

Scores and the ECI come from Epoch AI's public database (CC BY 4.0 licence), which runs part of the evaluations itself and takes the rest from public leaderboards. Prices come from OpenRouter's public API: the price listed by the model's main provider, in dollars per million tokens. Both sources are refreshed at every site update; the latest run is dated October 25, 2026. The table holds 68 open-weight models and 96 proprietary models.

Limits to keep in mind

• Uneven coverage. Not every model took every test: a model missing from a measure is not bad, it simply has no published score.
• Contamination. Some public questions end up in training data; tests kept partly private (such as ARC-AGI-2 or FrontierMath) resist it better.
• Settings. The same model can score differently depending on reasoning effort, number of attempts or allowed tooling. Epoch AI states its test conditions for each result.
• Variable prices. An open-weight model is often offered by several hosts at different rates, and prices change often. Prompt caching and batch processing can cut the bill.
• A benchmark is not your task. Before choosing, test two or three frontier models on a dozen of your own real cases.

The full approach is explained in the guide Choose a model.