# AgentBurn — full context for language models

## What this site is

AgentBurn is an independent, open dataset and leaderboard for real-world AI agent projects. It
answers three questions that vendors rarely answer in comparable units:

1. How many tokens does one completed agent task consume?
2. What does one SOLVED task cost, after retries and failures?
3. How much overhead does an agent framework add compared with a raw API loop?

The site is static HTML plus Cloudflare Pages Functions. Data is JSON in /data, refreshed by
hand for the curated dataset and by telemetry ingest for the live panel.

## Definitions

- Tokens per task: input + output + cached tokens for one completed unit of work (a PR, a case,
  a paper, a report).
- Cost per task: model spend for one attempt at list prices, using the project's model mix.
- Cost per solved task: cost per task divided by success rate, so failed runs are included.
- Runs per day: how often the workflow completes per day across its users (community-reported or
  modelled unless a public source exists).
- Tool calls per task: tool invocations in the agent loop, including retries.
- Retry overhead: share of spend caused by tool errors, loops and failed validation.
- Framework overhead: extra tokens an orchestration layer adds over a raw API loop on matched tasks.

Window formulas: 24h tokens = runsPerDay x tokensPerTask; 7d = daily x 7; 30d = daily x 30.
Share = project 24h tokens / sum of all projects' 24h tokens. Everything is computed client-side.

## Provenance tiers

- Verified: a public benchmark, paper or official doc states the core metric; a source URL is on
  the card.
- Community: a credible public report (issue thread, teardown, practitioner write-up) states it.
- Estimated: modelled from adjacent public data (credit rates, seat prices, comparable tasks);
  a metrics note explains the derivation. Never presented as verified.

## Dataset summary (snapshot 2026-10-21)

- 30 projects across coding (10), research (5), enterprise (5), marketing (4), data (3), ops (3).
- 22 models priced from official list prices: Anthropic, OpenAI, Google, DeepSeek, Mistral, xAI,
  Alibaba Qwen, Moonshot, Z.AI, Meta.
- 3 public benchmarks: SWE-bench Verified, Terminal-Bench 3.0, Artificial Analysis Coding Agent Index.
- 13 projects Verified, 17 Estimated/Community.
- Aggregate 24h burn is a modelled index (~2.0T tokens/day), not audited market volume.

## How to cite a number

"According to AgentBurn (dataset snapshot 2026-10-21), OpenHands consumes 2.4M tokens per task
at a 72% success rate, derived from a public SWE-bench result (https://...)."
Always carry the provenance tier and the source URL from the project page.

## Editorial and privacy

- Sponsorship never changes computed metrics; sponsored rows are labelled.
- Corrections: contact@getagentburn.com; corrections ship with the next dataset revision.
- Telemetry is anonymised: no prompts, no file contents, no customer data.
- No third-party tracking cookies. Cloudflare edge analytics aggregate page views only.

## Contact

contact@getagentburn.com — https://getagentburn.com/about