Built agent-first: every table here is a static JSON file with open CORS, no key, and no quota. Most leaderboards hand you a table and leave the reasoning to you — this one answers the actual question, "which model should I use for this job at this budget", in a single request.
One file per task. Each returns every scored model with its capability, blended price, context window and tool-calling support — plus an objective Pareto frontier and four ready-made picks, so a bot can decide without doing its own analysis.
$ curl -s https://agentleaderboards.com/api/v1/route/coding.json | jq .picks.best_value
{
"id": "z-ai/glm-5.3-flash",
"name": "GLM 5.3 Flash",
"strength_vs_leader_pct": 92,
"blended_per_mtok": 0.237,
"cheaper_than_leader_pct": 98,
"tool_calling": true,
"why": "Near-parity pick: the cheapest model still measuring at least 90% as
strong as the leader…"
}| Endpoint | Contents |
|---|---|
/api/v1/route/index.json | All 18 task types, what each measures, and its endpoint |
/api/v1/route/{task}.json | One task: ranked models, picks, Pareto frontier, full provenance |
/api/v1/route/all.json | Every task in a single fetch |
Task ids: coding · agentic · agent-tool-reliability · agent-steerability · agent-recovery · reasoning · math · instruction-following · long-conversation · long-input · creative-writing · general · software-it · legal · medicine · science · business-finance · writing
| Endpoint | Contents |
|---|---|
/api/v1/leaderboard.json | Everything: models, frameworks, SWE-bench, routing, provenance |
/api/v1/models.json | Models — pricing, context, modalities, tool calling, capability scores |
/api/v1/frameworks.json | Agent frameworks — live GitHub stars/forks/activity, category, license |
/api/v1/swebench.json | SWE-bench Verified / Lite / Multilingual / bash-only extracts |
/api/v1/usage.json | Which models people actually route to per category — popularity, not capability |
/api/v1/index.json | Machine-readable API description |
/llms.txt | LLM-oriented summary with the current answers inline |
/leaderboard.md · /router.md | Markdown mirrors |
/.well-known/agent.json | Agent manifest |
Send Accept: text/markdown to a normal page URL and you get clean Markdown instead of HTML, with the canonical URL in the Link header. One link serves both people and agents, so a shared URL never forks into two versions.
curl -H "Accept: text/markdown" https://agentleaderboards.com/router.htmlFree for any use, including commercial use and AI training. Attribution required: a visible link to https://agentleaderboards.com wherever the data is shown.
Capability data derives from the Arena leaderboard dataset under CC-BY-4.0 — carry that attribution too if you redistribute it. Responses are static files regenerated daily, so cache them freely. Every payload carries generated_at plus per-source retrieved_at, so you always know how fresh a number is.
No key, no quota, no signup. If you need a bulk dump or a different refresh cadence, email [email protected].