Open-source coding benchmark

Benchmarking models
on real engineering work.

Merged open-source pull requests, graded by deterministic hidden tests.

vulcanbench — suite: v1-micro · sandbox: docker · network: off

The harness, live — an actual v1-micro run on Claude Sonnet 5 (low): 26 sandboxed tasks, graded by hidden tests, 25 pass at $0.82 total.

04
Published reports
52
Decontaminated tasks
1,050+
Graded runs

A benchmark that measures the job, not the trivia.

Most coding benchmarks saturate: when every frontier model scores 98 to 100 percent, the benchmark is measuring its own ceiling, not the models. VulcanBench is built to stay discriminating by sourcing tasks from real engineering work that current models cannot all solve, and by treating economics as a first-class result.

Every task is a real pull request merged into a production open-source project (flask, sqlglot, aiohttp, urllib3, poetry, redis, and more), each merged after the models' training cutoffs so the fix is genuinely novel and cannot be recalled from memory. A model receives the repository sliced at the pull request's base commit plus a terse issue describing the symptom, exactly as an engineering team would hand off a ticket. It is graded by the project's own hidden tests: no LLM judge, no rubric, just pass or fail, run in a network-isolated sandbox.

Three rules, no exceptions.

01

Deterministic grading

Each task's pass or fail is decided by hidden tests validated to fail on the base commit and pass on the merged fix. No model judges another model.

02

Decontaminated tasks

Every pull request was merged after the evaluated models' training cutoffs, so no model has seen the solution. Memorization cannot help.

03

Economics as a result

Cost, wall-clock time, agent steps, and tokens are recorded alongside accuracy. When correctness converges, efficiency is the signal.

Same score. Four times the cost.

Total cost across fifteen evals, at identical accuracy
Every configuration scores 14 of 15 — each missing a different task.
Claude Opus 4.8 effort: high
$5.00
14/15
Claude Fable 5 effort: low
$5.99
14/15
Claude Fable 5 effort: medium
$11.10
14/15
GPT-5.5 effort: high
$20.43
14/15

Fig. 1 — When correctness converges, economics is the signal. Costs include each provider's prompt-cache discount. Read the full report →