API reference

Overall grades and the benchmarks API

Every live model has one overall grade, gray to green, graded on public benchmarks: general chat, web development and coding. See it on the model rankings, in /v1/models, or as JSON from one public endpoint.

How grades work#

Each benchmark score is placed on 0–100: a percentage as it is, a rating (LMArena's are Elo-style, relative and drifting) between the lowest and highest rating among our models. The grade's composite is the weighted average of a model's scores, preferring a leaderboard's own run over a vendor's reported number. LMArena Text, real people's everyday questions, carries the most weight, then web development and coding. Only benchmarks whose terms allow showing their results count publicly.

Tier (colour code)	Label and meaning
`green`	Excellent: composite 70 or more
`lime`	Strong: 55 to 70
`yellow`	Good: under 55
`gray`	Not rated: its scores cover under 40% of the benchmarks' weight

The lines may move; the response always carries the ones in use.

GET /api/public/benchmarks#

No key needed, open to any origin (CORS), and cached for ten minutes. Live models only, best grade first.

curl https://fullsendai.com/api/public/benchmarks

Response

{  "generated_at": "2026-10-15T12:00:00.000Z",  "thresholds": { "green": 70, "lime": 55, "coverage": 40 },  "benchmarks": [    {      "key": "lmarena-webdev",      "name": "LMArena WebDev",      "url": "https://lmarena.ai/leaderboard/webdev",      "unit": "rating",      "scale": "rating",      "range": [1226, 1811.1],      "weight": 20,      "credit": "LMArena (CC BY 4.0)",      "terms": "Dataset licensed CC BY 4.0: …"    }  ],  "recommended": { "best_for_coding": "…", "best_value": "…", "cheapest_good": "…" },  "models": [    {      "id": "model-id",      "name": "Model name",      "icon": "fullsend",      "grade": { "tier": "green", "label": "Excellent", "composite": 78.4, "rank": 1 },      "scores": [        {          "benchmark": "lmarena-webdev",          "name": "LMArena WebDev",          "score": 1655.3,          "unit": "rating",          "scale": "rating",          "kind": "leaderboard",          "measured_on": "2026-10-13",          "source_url": "https://huggingface.co/datasets/lmarena-ai/leaderboard-dataset",          "credit": "LMArena (CC BY 4.0)"        }      ],      "price": { "input": 3, "output": 15, "blended": 6, "unit": "USD per 1M tokens" }    }  ]}

Fields#

Field	What it is
`models[].grade.tier`	One of `gray`, `yellow`, `lime`, `green`.
`models[].grade.composite`	0–100, one decimal; null when not rated.
`models[].grade.rank`	Position among rated live models; ties share a rank.
`models[].scores[].kind`	`leaderboard` (the benchmark ran it) or `vendor` (the model's maker reported it).
`models[].scores[].measured_on`	When the score was published. Ratings change daily; always show them with this date.
`models[].price.blended`	US dollars per million tokens at three input tokens to one output.
`recommended`	The best overall grade (best_for_coding, a name kept for compatibility), the most grade per dollar, and the cheapest light-green-or-better model.
`benchmarks[].range`	What 0 and 100 map to. For a rating benchmark, the current spread among our models.

Grades in /v1/models#

GET /v1/models (OpenAI shape) carries the same public grade on every model, as "grade": { "tier": "green", "composite": 78.4 }.

Credits and terms#

Each benchmark in the response has its credit and terms. Show the credit wherever you show its numbers: LMArena's data is CC BY 4.0, so “LMArena (CC BY 4.0)” goes with any LMArena rating.

Grades describe public benchmark results for the underlying models, not a guarantee of how a model does on your code.

Stuck? Email [email protected]System status Changelog