A translation layer's losses live in fields, not in quality. The post names four casualties: tool-call block ordering, streamed tool arguments, prompt-cache markers and stop-reason fidelity, all of which live in fields with no equivalent on the other side [3]. A chat window reads none of those. An agent loop reads all four. That is why the first row asks which protocols an endpoint implements and which it translates, and treats those as two different products [2].
The max_tokens probe works because the Anthropic schema makes the field required. Send a `/v1/messages` call without it and an endpoint that validates the request itself has to reject it, which is the 400; an endpoint that returns 200 has supplied a value you did not send, on your behalf, before forwarding [4]. The 200 is the more informative outcome. A 400 tells you validation happens at the edge; it does not tell you whether something is rewritten between that edge and the model.
Streamed tool calls give you the same signal from the other direction. Incremental fragments carrying an `index` mean the endpoint is relaying deltas as they arrive; everything landing at once at the end means something buffered the upstream stream and re-emitted it [5]. Buffering is invisible in the final text and obvious to any consumer that acts on partial tool arguments.
Two more comparisons cost nothing. `GET /v1/models` should return the exact IDs your key can call rather than the vendor's full catalogue, and a completion response should echo a `model` field that matches what you asked for [7].
Then the availability number you generate instead of reading. A one-token completion against each model every five minutes is 12 calls an hour, 288 a day and 8,640 over a thirty-day month, appended to a JSONL file you own [18]. That buys a success rate and a latency distribution measured from your own network [12]. The author's framing of the alternative is the line worth keeping: a marketing-page availability figure is a claim, and a claim is not the same as a measurement [11].
One row is not about the vendor at all. Before a key is distributed, audit where it lands, because on a normal developer laptop that is an editor's settings store, a shell profile, two dotfiles, a CI secret and at least one Docker environment, five hiding places before the key has done any work [15].
The checklist is single-sourced and its author sells a gateway: he discloses working on daoxe, a multi-model gateway, and reports that two rows fail when he runs the list against it [16]. Treat that as a prompt for your own testing, not a settled finding. The tests cost curl calls, so run them against his endpoint, and against the one you are already paying for.
