Anthropic made Opus 5.5 cheaper. Then it broke four things your agent depends on.

Anthropic released Claude Opus 5.5 on Tuesday with lower prices: input tokens drop from $5 to $4 per million, and output tokens from $25 to $20 per million. The context window stays at 1 million tokens, and the maximum output stays at 128,000. So yes, the sticker price looks friendlier. The migration guide is doing a lot of the heavy lifting behind that calm face.

The first trap is thinking. Opus 5.5 rejects requests that try to disable thinking or use enabled with budget_tokens, and it now uses effort instead. It also starts every response with thinking blocks, which can trip code that expects plain text first. The default effort level has shifted from high to medium, so code that leaves the parameter alone may quietly get a different result than before.

Tool calls changed too. Forced tool use through tool_choice no longer works, including on token-counting requests, so cost estimates built that way can fail alongside the request they were meant to price. Anthropic’s workaround is auto-combined with strict tool use or structured outputs, with the prompt spelling out when the tool should apply. The broader pattern is familiar: the model gets cheaper, but the surrounding agent logic gets fussier.

The routing and history rules are stricter as well. Thinking blocks are tied to the model and conversation that produced them, and on the Claude API only Fable 5.1 and Mythos 5.1 can read Opus 5.5 thinking blocks besides Opus itself. Conversations also have to stay append-only for those blocks to remain valid; trimming messages, changing tool definitions, summarizing client-side, or rewriting the system prompt can invalidate them. Anthropic says accounts created on or after August 31, 2026 at midnight UTC will get a 400 error by default if they replay an invalid block after those edits.

The last big break is computer use. On the Claude API and Google Cloud, Opus 5.5 rejects the computer_20251124 tool and only accepts computer use through computer_toolset_20260801. The request shape is simpler in some ways, but the agent loop has to change: each action now comes as its own tool_use block, a single turn can contain several of them, and every result has to echo toolset_name. There are also quieter shifts that won’t throw errors, including progress narration turning into empty thinking blocks unless developers opt into updates or summarized display.