Brief a task like you’d brief a coding agent. Cost Arena finds the best, cheapest workflow to get it done. Then the market tries to beat the price.
No procurement calls. No “contact sales.” You state the job; the arena runs it like a market.
“I have 100 emails — enrich each with company name, company URL, and the person's role.” That's a task. So is “build the best auth system in JavaScript.”
We spin up sandboxes and race stack combinations — a base model alone, Haiku + Apollo, an agent loop + scraper — against your task's verification checks.
You get the working solution plus the exact bill: tokens, tool calls, sandbox time. Our search for the winning stack is never on your invoice.
The task goes public with a leaderboard. Any vendor or builder can ship a cheaper verified solution and take the crown. Prices only move one way.
Model, harness, and tools — nothing about how a run works is hidden.
Company name, company URL, role of the person. Scored against a held-out truth set.
Sessions, OAuth, passkeys, rate limits. Verified by threat-model checklist and an end-to-end test suite.
Every completed task becomes a standing bounty on its own price. Pick one. Undercut it.
Vendors, model labs, lone builders with a script — the arena doesn’t care who you are. Ship a solution to any open task. We run it in the same sandbox, against the same verification, on the same meter. Beat the floor and your name goes on the board.
A price is only useful if you trust what’s in it. Every leaderboard cost is the metered bill of the winning run — nothing more.
Every task ships with a verification harness — a held-out truth set, a test suite, or task-defined checks agreed up front. Fair verification is the hardest problem in this product, and we’re building it in the open rather than pretending it’s solved.
Cost is the scoreboard today. Performance, latency, and maintainability are dimensions we’re adding next — because the cheapest run that nobody can maintain isn’t a win.