Claude Sonnet 5.5 Guide: Effort Levels, Benchmarks and Migration

The headline on Claude Sonnet 5.5 is that the price did not change. Anthropic released it on 28 September 2026 at the same $2 per million input tokens and $10 per million output as Sonnet 5. The claimed saving comes from somewhere else: fewer tokens per task and output that Anthropic says is more than 30% faster, adding up to as much as 30% lower cost per task for most work.

That is a claim about behaviour, not a price list, which means it is the kind of claim you can and should check on your own workload. This guide covers what the model is, how to set it up so the claim has a chance of holding, and the API changes that will stop your old code at the door.

Where it sits in the lineup

Anthropic positions Sonnet 5.5 for well-scoped everyday work: bug fixes, clear feature tasks, and polished documents, slides and spreadsheets. Opus 5.5, at $4 input and $20 output, is the model for complex work that needs careful judgement. Haiku 5.5 has since joined the family for high-volume, latency-sensitive jobs.

The specification from Anthropic's model docs: 1M-token context as standard, 128K maximum output on the Messages API (300K on the Batches API with a beta header), adaptive thinking, a June 2026 knowledge cutoff, and text and image input. The API ID is claude-sonnet-5-5, and Anthropic commits to keeping it available until at least 28 September 2027.

Cache reads cost $0.10 per million tokens, half the Sonnet 5 rate, and the minimum cacheable prompt drops from 1,024 to 512 tokens. For agents that resend a large system prompt and tool list on every turn, that is a real reduction. Work out what it is worth by multiplying your cached prefix length by your turn count before assuming a saving.

Effort is now the main dial

Sonnet 5.5 has five effort levels: low, medium, high, xhigh and max. The Claude API defaults to high; Claude Code and the apps default to medium. Anthropic says the levels were recalibrated, so high on Sonnet 5.5 does not produce the same amount of thinking as high on Sonnet 5.

Anthropic's own guidance is specific:

• Agentic coding and multistep tool use — start at medium for well-specified tasks, move to high for harder or longer ones.
• Chat and latency-sensitive work — medium or low.
• Everything else — start at high.

The most useful data point in the launch post is a warning. On FrontierCode, max scored lower than xhigh. Anthropic attributes this to extra code-review subagents at max causing timeouts and out-of-scope edits. More effort is not monotonically better, and the top setting can cost more and do worse. The token cost of reasoning models explains why that hidden output adds up so quickly.

Anthropic also reports that on Terminal-Bench, medium effort beats Sonnet 5's best score at under a tenth of the cost per task. If that holds for you, most agent traffic should never run above medium.

The benchmark table, read carefully

Anthropic's launch table, Sonnet 5.5 against Sonnet 5 and Opus 5.5:

• Terminal-Bench 4.0 — 70.6%, against 10.3% for Sonnet 5 and 66.4% for Opus 5.5 at xhigh.
• CursorBench 4.0 — 55.5%, against 34.1% and 57.8%.
• GDPval-AA v2.1 — 1,844, against 1,449 and 1,846.
• OSWorld 2.1 (partial) — 80.1%, against 57.0% and 81.8%.
• Humanity's Last Exam with tools — 64.5%, against 54.9% and 67.7%.

The pattern is that Sonnet 5.5 lands within a few points of Opus 5.5 almost everywhere, at half the per-token price, and beats it on the terminal benchmark. The jump over Sonnet 5 is large across the board. The caveat is the usual one: these are vendor-run numbers on vendor harnesses, and a 70.6% on someone else's terminal tasks says little about yours until you test it.

Anthropic also notes Sonnet 5.5 is the first Sonnet to finish Pokémon Red from screenshots alone. It is not a business metric, but it is a decent proxy for staying coherent across a very long sequence of visual decisions.

What breaks when you migrate

From Sonnet 5, the list is short but sharp:

1. Forced tool use is rejected. A tool_choice of type any or tool returns a 400 error. Use auto with strict tool schemas.
2. Turning thinking off has a new name. Send thinking: {type: "between_tools"}, which allows short updates between tool calls but no up-front reasoning. It works at low to high effort and errors at xhigh or max.
3. Keep conversations append-only. Thinking blocks are signed over the conversation before them. On accounts created from 31 August 2026, replaying a block after editing earlier history is refused. Use mid-conversation system messages to change instructions instead.
4. Text between tool calls moves. It now arrives in thinking blocks, so a UI that only renders text blocks will go quiet during tool loops.
5. More refusal categories. Declines return stop_reason: "refusal" with a category such as cyber or reasoning_extraction. Handle them explicitly.

From Sonnet 4.6 or earlier, add the bigger changes: thinking runs by default, budgets are replaced by effort, sampling parameters are rejected, and the tokenizer produces about 30% more tokens for the same text. The Sonnet 5 guide covers the intermediate model and the cases where staying on it still makes sense.

Sonnet 5.5 or Opus 5.5

With the two models this close on published benchmarks, the decision rests on what a failure costs. Pick Sonnet 5.5 for bounded tasks where you can check the result cheaply and retry, which is most coding and document work. Pick Opus 5.5 for long unattended runs where one wrong judgement call wastes an afternoon, and confirm on your own tasks that the gap is real before paying double. The Opus 5.5 guide covers that side of the trade.

For bulk work that can wait, the :batch variant listed on OpenRouter is the same model on the asynchronous Batch API at half price, $1 input and $5 output per million tokens at the time of writing.