Methodology & sources

Every benchmark figure on the site was published by a third party. Each one is shown with the evaluator's name and links back to the page it came from.

Numbers that a lab published about its own model appear only on that model's page, under a separate heading. They use the lab's own harness and settings, so they are never mixed into rankings, composites or the finder.

41 models from 19 labs: the lineup each lab currently sells or promotes, plus its strongest open-weights release. Superseded versions are left out unless they are still the lab's only option in a tier. Image, video and audio generators are out of scope.

A model marked limited access has been announced and evaluated but cannot yet be used by the general public. A model marked preview is usable but the lab may still change it.

One model, many settings. Most current models have a reasoning-effort dial, and the same model can score very differently along it. The headline figure is the model at its strongest measured setting; the setting is printed beside the number, and each model page shows the full ladder with cost and waiting time at every step.

Tests and votes are different things. An independent test has right answers. A human-preference rating records which reply people liked more in a blind comparison — useful, but it rewards style as well as substance.

Small gaps are noise. Artificial Analysis estimates under ±1 point of uncertainty on its index, Vals reports roughly ±1, and Arena ratings for new models can swing by ±10 or more. Where a margin is published it is shown next to the score.

A blank is not a zero. “Not tested” means the evaluator has not published a run. Rankings always say how many models they cover.

The six area scores (0–100) exist so models can be compared on a whole area at once. For each benchmark, a model is placed between the weakest tracked model (0) and the strongest (100). The area score is the weighted average of those positions across the benchmarks below. A model needs the first benchmark listed, or at least half of the set, to receive a score.

Saturated or thinly covered benchmarks — GPQA Diamond, ARC-AGI-2 and ARC-AGI-3 — are reported on their own pages but kept out of the composites, so a model is not rewarded or penalised for whether it happened to be tested.

List price is the lab's published rate per million input and output tokens. The single blended price assumes three input tokens for every output token.

Cost per task is what Artificial Analysis actually spent running its Intelligence Index, divided by the number of tasks. Because it counts reasoning tokens, it often tells a different story from list price: a cheap-per-token model that thinks at length can cost more per answer than a pricier, terser one.

The efficient frontier is the set of models for which no other tracked model has both a higher Intelligence Index and a lower cost per task.

The finder ranks models by a weighted blend of capability and affordability. Capability mixes the area scores according to the use case; affordability places cost per task on a log scale where the cheapest tracked model scores 100. List price stands in when no measured cost exists, and the card says so.

Priority settings: best quality — 100% capability, 0% affordability; balanced — 65% capability, 35% affordability; lowest cost — 40% capability, 60% affordability. Turning on “speed matters” gives measured output speed 20% of the total. Models with less than half the capability of the best eligible model are dropped before ranking.

• Your task is not a benchmark. Use the rankings to build a shortlist, then test the two or three finalists on your own work.
• Harnesses matter. Agentic scores depend on the scaffolding around the model. The same model scores differently on Terminal-Bench 4.0 at Artificial Analysis, at Vals AI and in a lab's own run.
• Some results include a fallback. Where a provider routes a refused request to another of its models, evaluators score what the product returned. Affected models carry a note.
• It is a snapshot. Leaderboards move weekly. Everything here reflects Oct 8, 2026; follow the source link before relying on a specific figure.

Model Index is independent and not affiliated with any lab or evaluator. Product names and logos are trademarks of their owners, shown only to identify them. 23 benchmarks are tracked in total.