OpenHands: how many tokens per task, and what does it cost?
OpenHands publishes no per-task cost or token figure in the AgentBurn index. AgentBurn snapshot 2026-10-27. Every figure carries a provenance label: source-backed, derived, or left empty when no public basis exists.
What it is. Open-source CodeAct agent that resolves SWE-bench issues in a sandboxed container
OpenHands (formerly OpenDevin) gives an LLM a CodeAct loop over a Docker sandbox: it writes code as actions, edits the repository, runs tests and iterates until the issue is fixed. On SWE-bench Verified it reaches around 72% resolved with inference-time scaling and a critic model, at a few dollars of tokens per instance.
| Tokens per task | not published | Not published |
|---|---|---|
| Cost per task | not published | Not published |
| Success rate | not published | Not published |
| Cost per solved task | not published | — |
| Tool calls per task | not published | — |
| Retry overhead | not published | — |
| Minutes per task | not published | — |
| Estimated runs per day | null (model estimate) | Not published |
Where the tokens go
This is the per-task split AgentBurn records for OpenHands, step by step.
| Step | What happens | Share of tokens | Tool calls |
|---|---|---|---|
| Reproduce | Reads the issue, reproduces the bug inside the container | 12% | 8 |
| Localise | Greps the repo and opens candidate files | 20% | 22 |
| Patch | Writes the fix as code actions in the sandbox | 26% | 24 |
| Test loop | Runs the failing tests, reads tracebacks, retries | 28% | 26 |
| Critique | A critic model checks the patch against the issue | 14% | 10 |
What each number is based on
AgentBurn labels every field. Source-backed means the cited page states the figure; derived means a formula on this page produced it; not published means no public basis exists and the field is left empty rather than estimated.
| Field | Basis | Note |
|---|---|---|
| tokensPerTask | Not published | No usable public source for per-task tokens. |
| costPerTaskUsd | Not published | No usable public source for per-task cost. |
| successRatePct | Not published | No public task success rate. |
| runsPerDay | Not published | No public run volume. |
Frequently asked
How many tokens does OpenHands use per task?
No public tokens-per-task figure exists for OpenHands; AgentBurn leaves the field empty rather than estimating it.
What does OpenHands cost per task?
No public per-task cost exists for OpenHands, so it is recorded as not published.
Is this a vendor list price?
No. It is a per-task figure from a cited source, not a rate card; the two are not comparable.
Sources
- SWE-bench deep-dive: agent performance, costs and failure taxonomy (2026)
- Vendor site: https://openhands.dev
- Repository: https://github.com/All-Hands-AI/OpenHands
- Paper: https://arxiv.org/abs/2407.16741
- Open dataset: /data/projects.json
- Methodology: how these numbers are built
Cite this page
AgentBurn (2026-10-27). "OpenHands". https://getagentburn.com/projects/openhands/
Machine-readable: projects.json · Markdown version