Methodology

How we rank

Every tier on the map comes from public benchmark results by one fixed rule. Nobody hand-picks a winner, and the rule is the same for every model and tool.

The map

There are 27 tasks, from chat and coding to transcription, music and 3D, grouped by what they work on. Each task has its own map. Its columns are the ways to do the task: models, apps, workflow platforms, agents, editors, builders, runtimes and APIs. Its two rows split cloud options (hosted by the vendor, closed models) from local ones (open-weights models and tools that can run entirely on your own hardware).

1. Public benchmarks only

Each task has a priority list of public benchmarks. A model is rated on the first one it has a result for; the first is the task's main benchmark, the rest are fallbacks so a model missing from one leaderboard still gets a tier. Fallbacks are used only when their own leader is a top model on the main benchmark, and they can award at most an A. Every result stores the exact leaderboard entry, the snapshot date and the source link.

2. From score to quality

A result becomes a quality score from 0 to 100 by its distance from the best result on the same benchmark (its #1 at the snapshot, even when that is an older version we no longer list):quality = 100 − 100 × gap ÷ scale. Three scales cover every board, so a tier means the same thing everywhere:

• Arenas (blind human votes, Elo-style ratings) use 400. On the Elo scale a gap is a win rate: an S-tier option still wins at least ~44% of head-to-head comparisons against the #1, A ~36%, B ~26%.
• Point and percentage benchmarks use 40: S is within 4 points of the best, A within 10, B within 18.
• Error rates (word error rate for transcription, lower is better) use 10: S is within 1 point of the best, A within 2.5, B within 4.5. Best results sit near 5%, so one point already means about a fifth more mistakes.

3. Tiers

• S: quality 90–100 (within noise of the best)
• A: quality 75–89.9
• B: quality 55–74.9
• C: quality 0–54.9

Tiers are deliberately coarse: a few points on a leaderboard are often within its margin of error.

4. Tools

Few apps and platforms have benchmarks of their own. Where one exists (for example coding agents on Terminal-Bench, which tests an agent and a model together), the tool is rated by it (filled badge). Each such board ranks one kind of product, so tools are compared only on the board for their own kind (search APIs on a search-API board, deep-research apps on a deep-research board). Otherwise a tool inherits the tier of the best model it gives you (outlined badge). That is a ceiling, not a measurement: it says what quality you can reach with the tool, not how pleasant the tool is.

In the local row, only open-weights models count, because that is what you can run yourself. A local tool that also takes cloud models shows that separately (“with cloud models: tier S”).

5. Best value and the picks

The Best value lens re-orders every cell by how many options in it are both better and cheaper; options that nothing beats on both get a $ mark. “Cheaper” means the API price for cloud models in the task's unit (per million tokens blended at three input tokens per output token, per image, per second of video, per million characters of speech, per minute of audio or per generation), the GPU memory needed for local models, and the cheapest paid plan for tools (zero when free).

• Best overall: the model with the highest quality (where only tools have results, the best measured tool).
• Best value: the cheapest model (by API price) in tier A or better.
• Best free start: the highest-rated cloud tool you can use without paying.
• Best local: the highest-rated open-weights model that fits a 24 GB GPU.

6. Staying current

Only the current version of each model family is ranked; older versions are kept on the model's page. Each record carries a “last verified” date, and anything not re-checked within 90 days is flagged for review. See the changelog for what's actually changed recently.

Limits

• Arena leaderboards measure blind human preference, which rewards style as well as substance. Automated tests measure one narrow skill well.
• Some numbers are reported by the model's vendor rather than measured independently; their source says so.
• A benchmark can only rank what it tests. A tool with no rated model shows a dashed “–” rather than a guess.
• Some tasks (for example workflow automation or meeting notes) have no public leaderboard yet. Their options are mapped but not ranked.