AI model-gateway reAPI enforces spend caps at runtime, keeps only call metadata, not content

reAPI, a pay-as-you-go gateway that routes requests across multiple third-party AI models, enforces a daily spend ceiling on every API key at runtime rather than only after the fact, according to the company's site.

"Per-key spend caps Set a daily ceiling on every reAPI key. A runaway worker hits the wall, not your card," the company states. Alerts fire before the cap trips.

The service also draws a line on what it keeps. "Request bodies and model outputs travel through the AI API aggregator but are never persisted. We retain only billing and audit metadata: model name, token counts, latency, status code, and key identifier," reAPI says. That is a metadata-only audit trail: a customer can reconstruct what was called, when and at what cost, but not the content of a request or response, which the company does not retain at all.

reAPI presents a single API surface over multiple providers, saying customers can switch models by changing the model name in a request rather than rewriting integration code. Its own listing names OpenAI, Anthropic, Google, ByteDance, Black Forest Labs and Suno among the routed providers.

Pricing is metered per completed generation, not by subscription. "Pay-as-you-go credits, priced per completed generation — per token, image, or video second depending on the model. There is no subscription and no monthly minimum, credits never expire, and failed generations are refunded automatically," the company says — tying payment to whether a job actually produced output rather than to time or seats.

reAPI also claims 99.96% uptime "with automatic failover" and offers region pinning so a customer can keep traffic inside the EU, US or APAC: "reAPI respects the boundary you set in the dashboard, even when a faster cross-region option exists," per the site.

That combination touches two trust-gap dimensions this outlet tracks directly: runtime governance, which the per-key spend cap enforces before a bill runs away rather than after, and audit trail, which the metadata-only retention partially satisfies — enough to reconstruct what was called and at what cost, not enough to review the content of a request or response, since neither is kept at all.