Break the Cap: get an AI agent to spend past a $0.06 limit. Free, 2 weeks, real prize.

Every team running AI agents in production has the same fear: a loop, a retry storm, or a bad max_tokens, and the API bill is the first to know.

The usual fix is a counter inside your own code. But a counter in the agent's own process dies with the agent. A crash, a restart or a parallel worker, and the limit is gone.

So I built Valta (valta.co): a proxy that sits between your agent and OpenAI. The agent holds a virtual key, never your real key. Every call is checked against per-run, per-day and per-month limits before it goes anywhere. Over the limit, it's refused and the provider is never called. If Valta can't decide, it refuses too. It fails closed.

I believe the cap holds. Now I want people who know how agents fail to prove me wrong.

🎯 THE CHALLENGE
Get the total spend of ONE run past $0.06.

What's fair game: retries, loops, fan-out, parallel requests, racing the limit, huge max_tokens, odd inputs, anything you can do through the API.

💸 IT COSTS NOTHING
No OpenAI key, no card. Use the model valta-sandbox. Responses are simulated, but every request goes through the real gate and the real ledger.

⚙️ SETUP (about 3 minutes)

1. Sign up free at valta.co
2. Open your agent, turn on Cap, set per-run to $0.06
3. Create a virtual key (vk_live_...)
4. Point any OpenAI-compatible client at https://valta.co/v1 with that key, model valta-sandbox
5. Send an X-Valta-Run-Id header (same value for every call in one run)

🏆 PRIZE
A free Builder plan for a year for the best break, plus a public shout-out for the most creative attempt. Failed attempts count too, and I want to read those just as much. I'll publish what people tried.

⏱️ Runs for 2 weeks. Non-streaming only for now.

If you try it and it holds, tell me what you'd attack next. If it breaks, I'll fix it and credit you. Either way you'll know more about how agent spend control fails than you do today.

Who's in?