Claude 5.5: effort rules, and the history is append-only

Three new models and one constraint that did not exist before: the conversation is no longer a piece of text you can rewrite. The prompt still decides what you ask for. What decides how much thinking happens, and how much cache survives, sits outside it.

What has shipped

Opus 5.5, Sonnet 5.5 and Haiku 5.5, with the IDs claude-opus-5-5, claude-sonnet-5-5 and claude-haiku-5-5. Sonnet 5.5 landed on 28 September 2026 and Haiku 5.5 on 7 October, the first Haiku with effort levels. Opus 5.5 opened the family.

Thinking is on by default on all three. The control is effort, with five levels —low, medium, high, xhigh, max— and a different starting point on each model.

The starting level, which is not the same number on all three:
effort = medium    # Opus 5.5 and Haiku 5.5
effort = high      # Sonnet 5.5

What changes for reasoning

effort is the main lever on depth, quality, latency and cost. On Opus 5.5 thinking cannot be switched off: thinking: {"type": "disabled"} and thinking: {"type": "enabled", ...} both return 400. The field gets omitted, and depth stays where it belongs, in the level.

The level names do not measure the same amount of thinking on every model, and that is the work nobody sees: Opus 5.5 at medium matches or beats Opus 5 at high. A prompt swept at high on the previous generation is not swept on this one, because the ladder gets walked again per model. Opus 5.5 also thinks more per turn than Opus 5 at the same level.

That sweep has a cache cost: changing top-level effort between requests invalidates the prompt cache. The per-message change in the mid-conversation-output-config-2026-07-01 beta keeps it.

Each model has its own floor. On Sonnet 5.5 the minimum is between_tools, not disabled, and it is only accepted at high or below: at xhigh and max it returns 400. At those two levels thinking and replies get much longer, so they are reserved for work where the gain has been measured.

And the history stopped being yours: every assistant turn goes back byte for byte with its thinking blocks inside. Nothing is summarised in place. Nothing is rebuilt between requests.

What changes for writing prompts

The above has one direct consequence. Editing an earlier turn, changing the system block or touching the tools list invalidates every later thinking block. On Claude 5.1 and later a prefix that does not line up can return bound to a different conversation, and thinking.block_binding.prefix_mismatch_behavior decides whether the failure arrives as an error or as a dropped block.

The reminder you used to bolt onto system now goes in as a mid-conversation system message. It is the documented way to add a rule without losing the reasoning already emitted.

System message inserted mid-conversation:

Reminder from here to the end of the conversation: [the new rule].

temperature, top_p and top_k were removed in API version 2026-10-21: setting any of them returns 400. Anthropic's answer is to steer from the prompt, so what used to be a number becomes a sentence. If you had temperature: 0 to keep output steady, you now say what steady means: same format, same sections, same order.

Verification moves in opposite directions depending on the model in front of you. On Opus 5.5 it comes out: it verifies its own work without being told, and the inherited lines —"include a final verification step"— cause over-verification. On Sonnet 5.5 and Haiku 5.5 it goes in: both report a change as done without running the check that exercises it, Sonnet 5.5 at low and Haiku 5.5 at low and medium.

What you stop doing

• Rewriting the history to fix the prompt mid-conversation. This is the hardest habit to drop and the one that breaks most: the fix is a mid-conversation system message, not an edit backwards.
• Carrying the previous generation's verification instructions into Opus 5.5. They get deleted, not reworded.
• Sending temperature, top_p, top_k or budget_tokens: 400 on the last one from Claude 4.7 onward.
• Reusing the previous generation's effort level without re-sweeping it on the new model.
• Asking it to write its reasoning in <thinking>. It may decline, and cost comes down through the level, not through a sentence.

One last note on Opus 5.5, which this post does not date: Anthropic's API notes and its news listing disagree, so there is no launch date that can be asserted.

Sources

• Models overview
• Release notes
• Model deprecations
• Claude prompting best practices
• Prompting Claude Opus 5.5
• Prompting Claude Sonnet 5.5
• Prompting Claude Haiku 5.5