LLM gateway

An LLM gateway that checks the answer.

The usual gateway job is done and largely solved. One key, many models, failover, a base-URL swap, and several good products give it away free. That is the floor now, not a product.

What no gateway does is tell you whether the answer that came back was any good. This one does, and that is what makes routing to a cheaper model a decision rather than a gamble.

Start free Migrate an existing app

The gateway parts, table stakes

OpenAI and Anthropic wire formats. Hundreds of models behind one key, via Vercel AI Gateway or OpenRouter. Automatic fallback when a provider fails. Nothing surprising, and nothing you should pay a premium for.

The part that is not

Every answer is checked before it leaves: schema always, groundedness against context you supplied, cross-model and resampling when a request earns them. A failed check escalates to your strong model automatically.

Priced so we cannot game it

Tokens at provider list price, zero markup. We make money on verification, which means we have no incentive to route you to a pricier model, and the savings figure on your receipt is not marketing.

One line, then nothing else changes

Same request shape, same response shape. The receipt arrives as an extra body block and a set of response headers, so anything already parsing your completions keeps working untouched.

Rolling back is the same line in reverse. Nothing to export first, because unless you explicitly turned retention on, there was never anything of yours stored to export.

the whole migration

- base_url="https://api.openai.com/v1"
+ base_url="https://www.llm11.com/v1"

Requests

last 30 days

Bars show how much of the baseline model’s cost each request avoided. Illustrative shapes, not a customer’s figures.

Gateways log. This one acts.

A gateway’s dashboard tells you what you spent and which model you spent it on. That is useful and it is not the same as knowing whether the cheap model coped.

The number to watch here is the escalation rate. It is the honest answer to whether a downgrade worked, because it counts the times a cheaper model produced something that did not survive a check.