Inferock: Serverless inference, with credits back for failed calls.

Credits come with firm limits. They are non-cash, expire after 12 months, and are capped at $1,000 in the wallet and $1,000 applied per month. They cover failures Inferock can verify: downtime, timeouts, empty or truncated output, and invalid structured output or tool-call arguments when a schema exists. Material billing or cache anomalies also qualify if there is observable evidence. The product has two modes. Inferock Watch is free and runs on a team's own OpenAI, Anthropic or Gemini keys. It produces per-call receipts and a monthly loss report, but the provider still bills the customer directly. Inferock Serve is pay-per-token inference through Inferock, and it is the only mode where service credits are applied. A local proxy called inferock-bench captures token usage, failures and retries. The router runs on Google Cloud Run, which is why the launch sits in the Cloud Run hackathon category. For founders, the pitch is a cost-per-usable-response view. Inferock's own example shows that a 10% failure rate on billed calls raises the effective cost of each usable response by 11.1%. Few providers offer a remedy for bad output. Routers, observability tools and cost dashboards mostly report what happened without paying anything back. Teams with high call volume or strict schema requirements could use the receipts to audit spend. The credit cap and the Managed-only remedy limit how much can be recovered.