Run a benchmark on a model

The skill	`benchmark-launch`
Changes anything	Yes. A run spends your credits, exactly as a launch from the Evaluate page does.
It picks this up when you say	”what benchmarks can I run”, “run this on Claude”, “is my run done”, “show my last runs”
It hands over to	`benchmark-analyze` to look inside a finished run

What it can do

List the benchmarks you can run
List the models you can run them on, both your own and the shared platform models
Tell you which models are live right now and which have to be brought up
Join a benchmark and queue a run on your account
Report a run’s state, and its score once it finishes
List your recent runs with their scores and links
Narrow that list to one benchmark or one model

What benchmarks can I run, and which of my models are live right now?

Run benchmark <id> on <model name>.

Is run <id> done? What did it score?

Show my last 5 runs with their scores and links.

State	What it means
`queued`	Waiting for capacity
`preparing`	The environment and the model are being brought up
`running`	The model is answering
`scoring`	The answers are being scored
`finished`	Done, with a score
`failed`	Stopped with an error, which the skill reports
`cancelled`	Stopped before it finished

Launch without both a benchmark and a model. If either is unclear it shows the candidates and asks
Launch twice for one request, or retry a failed launch on its own
Wait for a run. A chat turn times out after about a minute
Add a finished run to the leaderboard. That is done from the run’s page in the web app

The agent cannot message you first. Ask “is it done?” rather than waiting to be told.

Why each answer was marked wrong, and whose fault it is.

The walkthrough, including what the model actually answered.

Launch runs from your own code.

Poll a run programmatically.