What actually changed from Opus 5 to Opus 5.5

Cheaper, and stricter about requests

Opus 5.5 keeps Opus 5's context window, output cap and tokenizer, so token counts don't move, and it costs less per token — the comparison shows both rate cards. Anthropic says existing Opus 5 prompts should perform well out of the box. The work is in the request code and in a handful of behaviours that change what a user sees. Here is what changed, starting with what fails.

Thinking can't be switched off

On Opus 5, thinking: {"type": "disabled"} was accepted at effort high or below. On Opus 5.5 there is no off switch: send a disabled config or a manual budget_tokens config and you get a 400, whatever the effort level. That leaves effort as the single dial for how much the model thinks, so a route that ran with thinking off should move to a low effort level instead — see the error and its fix.

If a prompt asked the model to write out its reasoning in the reply as a substitute for thinking, remove that instruction. On Opus 5.5 it can trigger a reasoning_extraction decline. Read the thinking blocks with display: "summarized" instead.

The effort default dropped, and the levels moved

Leave effort unset and Opus 5.5 runs at medium, where Opus 5 ran at high. Yet level for level, Opus 5.5 usually thinks longer per turn than Opus 5, and the gap is widest at xhigh and max. So the same setting doesn't mean the same behaviour in either direction. Set effort explicitly and re-run your effort sweep rather than carrying the Opus 5 value across; the effort default guide covers what each choice does to the bill. To get less thinking, lower the level before adding "think less" instructions — it works more reliably than prompting.

Forced tool choice is rejected

tool_choice of any or of a named tool returns a 400. Migrate by intent: leave tool_choice on auto, use strict tool use to keep the schema guarantee, name the tool you expect in the prompt, and check in code that the call happened. The error page has the patterns.

Thinking blocks are bound to the conversation

Editing history between requests — a rebuilt system prompt, a trimmed tool result, a reminder inserted and later removed — now invalidates every later thinking block, and on newer accounts that is a 400. Opus 5.5 reads thinking blocks from Opus 5 and earlier models, so a conversation that moves onto it keeps its reasoning. The reverse doesn't hold: on the Claude API only Fable 5.1 and Mythos 5.1 can read an Opus 5.5 block, so a router switch or refusal fallback to Opus 5 continues without it. Making an agent loop safe for preserved thinking has the append-only replacements.

Computer use moves to the toolset

On the Claude API and Google Cloud, the older computer_20251124 tool returns a 400; Opus 5.5 accepts only the computer toolset, which needs no beta header. Amazon Bedrock still accepts the older version. Opus 5 accepts both, so make and test the change there first — the error page shows both request shapes.

Progress notes arrive as thinking blocks

On Opus 5, the short notes the model writes between tool calls came back as text. On Opus 5.5, any note longer than a sentence or two comes back as a progress-update thinking block — and under the default display setting its text is empty. No request fails, but an interface that renders only text goes silent for the length of a long agentic turn. Set thinking.display to "updates" (beta thinking-display-updates-2026-08-18), render each non-empty thinking block before the tool call it precedes, pass the blocks back unchanged, and say in the system prompt how often you want updates.

A broader set of refusals

Opus 5.5 adds a biology classifier to the cybersecurity one Opus 5 had, and it can decline requests that try to extract its internal reasoning. Declines arrive as an HTTP 200 with a refusal stop reason, and server-side fallbacks can retry them on another model — handling refusals walks through the setup.

Prompts worth revisiting

Opus 5-specific instructions. Prompts written to curb Opus 5's verbosity, over-verification or scope may no longer be needed. Keep them as a starting point, then test each one against your own evaluations instead of carrying it over untouched.

Frontend design. Asked for frontend work without design direction, Opus 5.5 falls back on a few default styles, and a general "avoid a generic AI look" mostly swaps one default for another. It responds well to a list of specific patterns to avoid; look at what the first result used and extend the list.

Visual inputs. It reads charts, diagrams and screenshots far more precisely out of the box, so cropping and zooming scaffolding built for earlier models may no longer earn its place. Re-test it before keeping it.

Long turns. At xhigh and max, turns run longer than on Opus 5. Plan timeouts, streaming and progress indicators accordingly — and lower effort before prompting for brevity.