Put your result on a board
Run a benchmark with hbench on your own machine, upload the job, and publish it to its board. If the model needs a key, the key stays in your own shell: the site never sees it.
Uploading is free. A model you serve on your own machine has no model fee. A commercial model bills your key for every trial: a run of SWE-bench Verified on Claude Opus 5.5 costs about $80, by that board's estimate. Is it free?
Install hbench
hbench runs each task in Docker on your machine, on Linux or macOS. This command installs its latest release (more on installing):
Run a benchmark
For a board, the run builder issues the run and writes its commands, so the site knows of the run before it starts and checks your upload against its
spec.json. You can also run any dataset from a registry, with any agent and model, and 1 to 5 attempts per task. Your first run, start to finish uses a model on your machine, so it needs no key.Upload the job
Drop the job's folder on Your bench, with its run token if you issued the run. Only you see it there. Upload a job also shows how to upload from a script.
Publish it
Press Publish on the job's page. The board then lists your score in its harness's row, beside the other harnesses that ran the same tasks on the same model (what a board shows). Your bench keeps every trial: what the agent did, and why the tests passed or failed. Read your result explains each outcome.
If something goes wrong, see troubleshooting. Limits lists how many runs you can issue and how large an upload can be.