Anthropic made Opus 5.5 cheaper — then broke four things agent code depends on

Anthropic released Claude Opus 5.5 on Tuesday at a lower price than its predecessor: $4 per million input tokens and $20 per million output, down from $5 and $25. The context window and maximum output are unchanged.

The company's migration guide flags four breaking changes that can make requests built for Opus 5 return 400 errors.

Four changes that return 400 errors

Thinking controls come first. Opus 5.5 returns a 400 error when a request sets thinking to disabled or uses enabled with budget_tokens; effort is now how reasoning volume is controlled, so agents that switched thinking off for simple steps must give them a lower effort level instead. Because thinking is always on, responses begin with thinking blocks, and code that assumes the first content block is text must change too.

Forced tool use no longer works either. Setting tool_choice to any or to a named tool returns a 400 error, including on the token counting endpoint, so cost estimates built on those settings fail too. The replacement is auto combined with strict tool use or structured outputs.

Thinking blocks are now tied to the model and conversation that produced them. On the Claude API only Fable 5.1 and Mythos 5.1 can read Opus 5.5 thinking blocks, so a router or fallback that hands a conversation to another model runs those turns without the earlier reasoning instead of erroring. Conversations must also stay append-only: trimming old messages, changing tool definitions or rewriting the system prompt mid-conversation invalidates existing thinking blocks, and accounts created from August 31, 2026 onward get a 400 error by default when one is replayed.

The fourth change affects computer-use agents on the Claude API and Google Cloud, where Opus 5.5 rejects the computer_20251124 tool and accepts computer use only through the computer_toolset_20260801 toolset. Each action arrives as its own tool_use block identified by name rather than input.action, and every result must echo toolset_name.

Changes that will not throw an error

One change produces no error at all. Text Claude writes between tool calls now arrives as progress-update thinking blocks, empty at the default thinking.display setting of omitted, so interfaces that stream that narration go silent until display is set to updates or summarized. Opus 5.5 also ships with broader safety classifiers: a refusal stop_reason can carry bio and reasoning_extraction alongside cyber, and the server-side fallback will not retry requests declined under reasoning_extraction.

Upgrading from older models

Teams coming from Opus 4.8 must work through the Opus 5 migration first; those on Opus 4.7 or earlier also face rejected sampling parameters, no manual extended thinking, no prefill and a newer tokenizer. Claude Managed Agents users only change the model name, and Claude Code developers can run /claude-api migrate to apply the ID swap, parameter changes and prefill replacement. Anthropic recommends testing in development before switching production traffic.