About AgentBurn
AgentBurn exists because agent costs are quoted in units that cannot be compared: tokens per million, credits per action, seats per month, dollars per solved ticket. We publish one dataset in one unit set — tokens per task, cost per solved task, runs per day — with a source behind every number.
Who runs this
AgentBurn is an independent project maintained by a small editorial team. It is not affiliated with any model provider or agent vendor. The site is self-funded; the only commercial surface is clearly labelled sponsorship, which never changes a computed number (see pricing).
Contact the editor at [email protected]. We answer data corrections and submission questions within two business days.
Editorial team and roles
AgentBurn is run by a small editorial desk. Roles are listed so readers know who owns each decision, even when individual names are kept private for a dataset project:
- Editorial desk — owns the project index, the metric definitions and the public changelog.
- Data desk — maintains sources, provenance tiers and the JSON exports.
- Corrections desk — triages reported errors and ships fixes with the next snapshot.
- Sponsorship desk — sells labelled placements only and has no access to computed metrics.
Every page is authored by the AgentBurn editorial team; the byline links back to this page, and the machine-readable author is the Person entity AgentBurn Editorial in the JSON-LD.
Where the data comes from
- Public benchmarks and papers for per-task cost and token counts (SWE-bench Verified, Terminal-Bench 3.0, Artificial Analysis Coding Agent Index, research papers).
- Vendor documentation and rate cards for consumption-priced agents and model list prices.
- Community reports where a practitioner published a real bill or usage breakdown.
- Telemetry from agents instrumented with our SDK or an OpenTelemetry bridge — measured, not modelled.
Each project carries one of three provenance tiers: Verified (a public source states the number), Community (a credible public report) or Estimated (modelled from adjacent public data, with a metrics note). The full rules are on the methodology page.
Editorial independence
Rankings are computed from the same dataset for every project. Sponsorship buys placement in a labelled slot — never a metric, a rank or a success rate. We reject submissions that cannot cite a source, and we mark every number we could not verify.
Update cadence
The curated dataset is revised when model pricing or a tracked project changes materially; the dataset snapshot date is shown on every leaderboard and in meta.json. Live telemetry aggregates roll over a 30-day window. Product changes are logged in the methodology changelog.
Citing AgentBurn
Cite the dataset snapshot and keep the provenance tier with any figure you reuse:
AgentBurn, "Agent project library", dataset snapshot 2026-10-21,
https://getagentburn.com/projects, accessed YYYY-MM-DD.Machine-readable context for language models is published at /llms.txt (summary) and /llms-full.txt (full context).
Contact
Editorial, corrections and sponsorship: contact page · [email protected]. Privacy questions: see the privacy policy.