Four model releases in 12 days break code written against earlier versions

Editing `model: "claude-sonnet-5"` looks like a config change. A dev.to tutorial that collected September's breaks argues the string "is part of an API contract" ^[20]. Its author wrote: "Changing a model string is a dependency upgrade. It deserves a test suite." ^[13]

The loud Sonnet 5.5 breaks fail on the first call. A client that turns off up-front thinking with `"disabled"` now has to send `thinking: {"type": "between_tools"}`, at high effort or below ^[7]. Forced tool use through `tool_choice` types `any` and `tool` returns a 400 error here too ^[8]. On the Claude API and Google Cloud, the older `computer_20251124` computer use tool is not accepted ^[10]. The advisor tool rejects Opus 4.8, Opus 4.7 and Sonnet 5 as advisors ^[11]. Each of these fails the first time a test sends it. A 400 at least has the decency to show up in the error logs.

The quiet break is the one I would test first. Anthropic's What's new page says text between tool calls now comes back in `thinking` blocks, a change that "alters the response shape without failing any request" ^[12]. An agent UI that shows that narration as a progress line reads `text` blocks. After the swap it gets a 200 and has nothing to render. Anthropic deserves credit for writing down a change that no error would have exposed. The tutorial's mock reproduces it: Sonnet 5 returns the note as a text block, and the Sonnet 5.5 mock returns a thinking block that stays empty unless the request asked for `between_tools` ^[17]. The author says the mock's request rules follow the docs and its replies are made up ^[14]. The empty string is a model of the failure, not a captured API response.

The gate has two parts. One check encodes the documented breaks as rules inside a mock provider. The other runs the same contracts against two model versions and diffs the results ^[15]. Both sit behind a request and response type the app owns. "Your contracts should describe your app, not the provider," the tutorial says ^[16]. The mock's `tool_choice` error string is copied from Anthropic's docs ^[19]. The whole thing runs on Node.js 18 or newer with TypeScript and tsx, and it needs no API key ^[18][14].

The known-breaks check is only as complete as the page it was built from. Anthropic's release notes number five breaks, and the response-shape change sits on a second page, so Sonnet 5.5 has six documented changes across two pages ^[4]. A gate built from the release notes alone would pass a build that drops every progress message. The diff check closes that gap only when it runs against real endpoints, where replies vary from run to run. For a passing mock run to mean anything in production, the diff has to assert on structure: block types, stop reasons, tool names. I would not diff text at all.

One more test belongs in the suite. The notes say "Thinking blocks are tied to the model and the conversation" ^[9]. A team that stores conversations with thinking blocks and replays them after the swap should run a stored Sonnet 5 transcript against 5.5 before trusting it.