Token leaderboards
These boards rank real agent deployments by burn, efficiency and framework overhead. The three boards below are recomputed in your browser from the published usage profiles, so switching the time window moves them together; the live telemetry section switches to real ingest data when it is connected.
Loading dataset…
Token burn
Derived as runs per day × tokens per task × 1 day.
| # | Agent | Tokens (24h) | Share of window | Δ 7d |
|---|---|---|---|---|
| Loading the burn board… | ||||
Ranked by derived token volume.
Cost per solved task
The price of a successful outcome rather than a single attempt — retries and failed runs are folded into the number.
| # | Agent | Cost / solved task | Success rate | Solved / $1,000 | Tokens / solved task |
|---|---|---|---|---|---|
| Loading the efficiency board… | |||||
Sorted cheapest first.
Framework overhead
The same task, run through different orchestration layers. Overhead is the extra token volume a framework adds on top of the leanest baseline measured on matched tasks — system prompts, tool schemas, serialised state and multi-agent message history.
| Framework | Overhead vs baseline | What drives the extra tokens | Source |
|---|---|---|---|
| Loading framework data… | |||
Overhead measured against the leanest framework on matched tasks.
Telemetry from instrumented agents
Real token counts reported through the ingest API by agents running in production — no estimates, no prompt content.
Checking the ingest API for aggregated token counts…
Benchmark sheets
What each agent scores on a public benchmark, next to the tokens and dollars it takes to get there. A higher score bought with four times the tokens is not automatically progress.
Running an agent that belongs on this board?
Submit it with a public source, or emit real token telemetry with the SDK and skip the estimate column entirely.
The leaderboards did not load
The boards are computed in your browser from /data/projects.json, and that request failed.