A benchmark run gives a model, or your swarm, the same frozen task that every other entrant gets, then scores what it built by running it. The score comes with its evidence: the individual checks, screenshots and a recording of the grading. Use it to compare models before you trust one with real work. Nothing is published unless you choose to publish.
Gauntlet and Forge
The Benchmark view has two families. Gauntlet asks the entrant to build an application from a frozen specification, and you can enter either a single model or your swarm. Forge asks one model to build a Jira app in an offline test harness; a Forge run makes two independent builds, the second after the first is scored, and reports their median and range when both results are comparable. Only the newest stable version of each benchmark can be launched, and the app checks that before it starts.
The tasks and the scoring rules are public: see the benchmark methodology and the Forge benchmark.
Running one
- Set up a provider with its credentials, or a local model.
- In the Benchmark view, install the benchmark tools and, for Forge, finish the Forge Kit setup. Run stays disabled until everything is ready.
- Choose "New benchmark run", pick the family, the entrant and the exact model, and start it.
The model builds a whole application, so a run takes a while. Cloud usage is billed by your provider. If the build finished but scoring failed, "Retry scoring" grades the saved build again without a new model run.
What you get
The final score, the earned score and any admission limits, every individual check, screenshots and a video of the grading browser in its original order and speed. Runs on this Mac stay in the sidebar, each marked current or frozen.
"Publish to leanzero.net" sends a finished run to the public benchmark board under your installation's public handle. Your provider credentials are never published, and cancelling a run publishes nothing.
How it differs from sessions and builds
| Agent session | Software build | Benchmark run | |
|---|---|---|---|
| You start it with | A chat in a project folder | A chat on "Elzi · Build" | The Benchmark view |
| Who drives | You, turn by turn (or a schedule, for Agent Work) | The swarm, from one brief | A frozen task, the same for every model |
| Models | One model or node answers | Many nodes share the work | The model or nodes you want to measure |
| What you get | The conversation and the files it changed | A planned, multi-agent build written into the project | A score with its evidence, which you can compare |
Facts on this page were checked against our own test records on 6 November 2026.