AIToday
Large Language ModelsAI Business & IndustryZenn AI/MLPublished: Nov 21, 2026, 16:00 JST

Claude Opus 5.5 rejects forced tool calls; EveryInfra bills $0.001389 for bad values

Send on LINE
Claude Opus 5.5 rejects forced tool calls; EveryInfra bills $0.001389 for bad values

3 Key Points

  1. What happened

    On 2026年11月1日, EveryInfra tested Claude Opus 5.5, Sonnet 5.5, and Fable 5.1 and found forced tool calls return a 400 error; its trustpilot/reviews test showed invalid sort and date values accepted and billed at $0.001389 before a same-night fix.

  2. Why it matters

    the rejected forced-call pattern breaks existing Claude apps that relied on it to return fixed-shape JSON, and invalid values slipping through unnoticed can silently change results while still charging the usual fee.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

Anthropic's release notes explain the change as a consequence of always-on thinking in the new models: a forced tool call skips that thinking, so the model writes its reasoning into the tool arguments and the argument quality drops. The recommended migration is to let the model decide whether to call a tool while a strict grammar fixes the shape of the call. EveryInfra's own tests on 2026年11月1日 show why both layers of defense matter. Before the night fix, its trustpilot/reviews parameters had no allowed-value lists, so invalid values were accepted and billed at $0.001389, returning a different set of reviews with no error signal. That silent acceptance is worse for a model than an outright failure, because the model sees success and never retries or warns the user. During the same day, strict schemas on the client side would have blocked values like newest and last7days before any call was sent, since they are defined as enums there. The same-night server change added allowed values so that wrong input now returns a 422 before billing. The episode points to a division of labor: strict schemas keep the request shape and enum values correct, and the API's 422 responses handle the rules a schema cannot express.

FAQ
What replaces forced tool calls in the new Claude models?
Anthropic recommends keeping tool_choice: auto and adding strict: true to tools, or moving the schema to structured outputs.
What happened when EveryInfra sent invalid values to its review endpoint?
On the afternoon of 2026年11月1日, invalid sort and date values returned 200 with normal billing of $0.001389 and quietly returned a different set of reviews; after the night fix, the same invalid values returned a 422 before billing.
How is strict tool use different from structured outputs?
Strict tool use controls the tool arguments inside a tool_use block, while structured outputs controls the final response text; data-API workflows need both.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articlenference, Cedars-Sinai to pool 5M patients' records for AI research