Fetching from the wire…
Public story · 2026-07-25 · high
Anthropic says turning thinking off can also make Opus 5 write a tool call into plain text instead of running it.
Why now: Anthropic's Opus 5 docs, current as of July 25, 2026, flag this alongside other breaking changes from Opus 4.8 that builders porting configs forward need to check first.
Opus 5 now rejects any request that pairs thinking: {"type": "disabled"} with effort xhigh or max, returning a 400 error, per Anthropic's Opus 5 docs. That's a break from Opus 4.8, where thinking mode and effort level were independent settings, and it will trip up any integration that hardcoded both.
The bigger issue sits below that ceiling. With thinking off, Opus 5 sometimes writes a tool call into user-facing text instead of a real tool_use block, per the docs.
The call looks like it ran. It didn't. In an agentic loop, that phantom call becomes part of the context every later turn reads.
Anthropic's fix is blunt: stop disabling thinking. Thinking enabled at low effort beats thinking disabled at similar cost, per the docs.
If you still need it off, Anthropic ships one combined mitigation instruction to paste in. It says that beats the DIY fix teams reach for, naming <thinking> tags in the prompt.
The pattern matters more than either fix. Anthropic's Opus 5 docs also tell builders to drop their own fallback model lists for a new default mode instead. Two workarounds swapped out in one doc update, not tuned, is a pattern worth watching before you port an Opus 4.8 config forward.
Each link below shares sources, entities, or timing with this story.
Two changes shipped within a day of each other that compound in exactly the wrong direction, and nobody connected them. Anthropic's Opus 5 prompting guide documents that the model "delegates to subagents more readily than prior models," and is blunt about the consequence: it "...
Three separate Anthropic changes over about two weeks point the same direction, and none of them announced themselves as a strategy. Claude Code 2.1.238 added claude self-hosted-runner --defer-shutdown-max-min, which keeps serving attached sessions on SIGTERM, parks whatever's...
August 6: Sapiom raised $35M for a router that sends each model call to the cheapest capable model. August 7: Databricks published its internal cost playbook. August 8: Toolport hit Product Hunt with a free MCP gateway cutting tool-definition overhead 96%. Three parties, zero...
The August 5 platform release notes confirm no serving at all, with Anthropic recommending Opus 5 and pointing researchers to the External Researcher Access Program. Anyone with that pinned model ID in production config or an eval harness breaks today rather than degrading gra...
Platform release notes: Claude Opus 4.1 retires from the API on August 5. The experimental prompt tools APIs retire alongside the Workbench on August 17. The temporary 50% weekly usage boost for Claude Code subscribers runs through August 19. And Claude Sonnet 5's promotional...
Endpoints moved to agent-memory-2026-07-22, and combining it with managed-agents-2026-04-01 on the same request returns 400 rather than merging. The traps: listing now returns a stable server-defined order and silently ignores order_by/order, depth accepts only 0, 1, or omissi...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.