Methodology & sources

Every score is tagged with who produced it. An unmarked number was run by the benchmark’s own maintainer or an independent evaluator, on the same harness for every model. A number marked with a dagger (†) was reported by the model’s own lab and has not been reproduced independently. Lab-reported numbers are only used where no independent one exists, and never replace one.

Where no trustworthy number exists for a model on a benchmark, the cell is empty. A missing score is never shown as zero, and models without a score are listed separately rather than ranked last.

Most current models can think for longer when asked to, and evaluators test each setting separately. The catalogue shows one row per model, using its best-scoring configuration for each benchmark; the setting is noted beside the score on the model’s page.

Because that best score is usually the slowest and most expensive way to run the model, each model page also has a table of every effort setting with its score, cost per task and time to first answer.

List price is the lab’s own API price in US dollars per million tokens, input and output, at the base tier. Blended price weights three parts input to one part output, the common convention for a single comparable figure.

Cost per task is what Artificial Analysis actually spent, on average, for the model to complete one task of its Intelligence Index at the stated effort, reasoning tokens included. Two models with the same list price can differ several-fold here, which is why the value chart uses it by default.

Each pick is the top model on one metric, computed from the same data as the tables. Nothing is hand-picked.

• Most capable overall: highest Artificial Analysis Intelligence Index.
• Coding agents: highest Terminal-Bench 4.0 score.
• Professional work: highest GDPval-AA rating.
• Novel reasoning: highest verified ARC-AGI-2 score.
• People's choice: highest Arena text rating.
• Value near the frontier: highest Intelligence Index points per dollar of cost per task, among models scoring within 15% of the best Intelligence Index.
• Fastest capable model: highest measured output speed, among models scoring at least 70% of the best Intelligence Index.
• Best open weights: highest Intelligence Index among models with openly released weights.

Bars always start at zero and share one axis within a chart, so their lengths can be compared directly. Ratings that have no meaningful zero, such as Arena scores, are drawn as dots on a scale instead: a bar would turn a 30-point gap into what looks like a landslide.

Colour identifies a model only on the comparison page, where each of up to four models keeps its colour in every chart. Everywhere else a single colour is used and models are named in text; with 12 labs, colour-coding by lab would not be distinguishable, least of all for colour-blind readers. Lab logos identify labs.

• Benchmarks saturate and leak. Several well-known tests are now near their ceiling or known to be partly in training data; each benchmark page says so where it applies.
• Agent benchmarks measure a model together with its harness. A score achieved in one coding agent may not transfer to another.
• Close scores are often statistical ties. Arena ratings a few points apart overlap within their confidence intervals.
• A model marked New has been out for under 60 days as of the data date; independent results for it are usually still arriving.
• None of this replaces trying a model on your own task. Use the numbers to build a shortlist, then test.

The current line-up of each major lab: the models a developer would choose between today, from flagship to fast tier, plus notable open-weight releases. Superseded generations are left out so that the lists stay readable. Artificial Analysis Intelligence Index is the default ranking because it is independent and covers every model here.

A lab is listed once it has a current-generation model with independent results. Amazon (Nova 2 Lite, December 2025) and Microsoft (MAI-Thinking-1, in preview with no published price) are not listed yet for that reason.

Specs and prices come from each lab’s own model and pricing pages. Release dates, context windows and prices for the few models whose official pages could not be read fall back to the Artificial Analysis and Arena listings.

The numbers are not typed by hand. Dated snapshots of every source live in the repository, and a script joins them to the hand-written model list to produce the catalogue. Refreshing the site means replacing the snapshots and running one command; the project README has the details.