What we mean by 100% pass-through

There is a well-known trick in this market: accept a request for an expensive model, serve it with a cheaper one, and hope the customer cannot tell. On single-turn chat, most of the time, they cannot. On evals, distillation runs, and anything with a compliance story attached, it is a serious problem.

The contract

• The model you name is the model that runs. Always.
• Response bodies are forwarded unmodified, including the provider's own usage object.
• No prompt rewriting, no injected system messages, no truncation you did not ask for.
• If we cannot serve the model you asked for, you get a 503 — not a substitute.

How to check

Do not take our word for it. Provider models have measurable fingerprints: token-level logprob distributions, tokenizer boundaries, and characteristic refusal phrasing. Run the same prompt against your direct provider account and against the gateway, and compare the usage accounting and the output distribution.

# usage objects should match within rounding
direct = openai_client.chat.completions.create(**payload)
via_gw = bestagi_client.chat.completions.create(**payload)

assert direct.usage.prompt_tokens == via_gw.usage.prompt_tokens

If a gateway will not survive that test, the price it quotes you is not the number that matters.