Above: Dario Amodei, CEO Anthropic. (Photo: Simon Walker / No 10 Downing Street, CC BY 2.0)

Anthropic released Claude Haiku 5.5 on 7 October 2026. Anthropic's launch post covers benchmark results, pricing, customer comments, safety and availability. For developers, one significant change is how the model decides how much to think, and how that setting is controlled.

What Anthropic announced

In its launch post, Anthropic describes Haiku 5.5 as "the cheapest, fastest, and most capable small model we've ever released". Those are Anthropic's own claims. The company says the model is designed for high-volume, cost-sensitive work such as summaries, compactions, database queries and classification. It also says Haiku 5.5 pairs well with Opus 5.5 and Sonnet 5.5 as a subagent on coding work.

The feature this article follows is the effort setting. Anthropic says Haiku 5.5 is "our first Haiku-class model to come with an adjustable effort setting". The post adds that, as with the company's other models, "users can decide whether to optimize for cost or intelligence." Anthropic also publishes charts showing how Haiku 5.5 performs at each effort setting on three benchmarks: OSWorld, GDPval-AA and Humanity's Last Exam. These are the company's own results. The charts plot accuracy against cost.

How the effort setting works

Anthropic's Effort documentation says the parameter lets developers "trade off between response thoroughness and token efficiency with a single model". It is set per request, through output_config.effort, and the top-level parameter needs no beta header.

The documentation lists five levels: max, xhigh, high, medium and low. Claude Haiku 5.5 is named among the models that support max and xhigh. Haiku 5.5 and Opus 5.5 default to medium, while most other models that support effort default to high. The documentation describes the levels this way:

  • low: "Most efficient", with significant token savings and some capability reduction. The documented use case is simpler tasks that need the best speed and lowest costs, such as subagents.
  • medium: a balanced approach with moderate token savings, suited to agentic tasks that need a balance of speed, cost and performance.
  • high: spends as many tokens as the task needs, for complex reasoning, difficult coding problems and agentic tasks.
  • xhigh: extended capability for long-horizon work, such as long-running agentic and coding tasks.
  • max: no constraints on token spending, for tasks needing the deepest possible reasoning.

The documentation says effort affects all tokens in the response: text, tool calls and function arguments, and thinking when it is active. Because it applies to every output token, it works whether or not thinking is enabled, and lower effort means fewer and terser tool calls.

The practical upshot is that a team can run one model and give different kinds of work different effort levels, rather than switching models. The documentation names subagents as a typical use for low.

Thinking is now adaptive, and on by default

Anthropic's Haiku 5.5 migration guide describes a change in how thinking is configured. In Haiku 4.5, a request could set thinking to {"type": "enabled", "budget_tokens": N}. On Haiku 5.5 that value "returns a 400 error".

The replacement is adaptive thinking:

"thinking": { "type": "adaptive" },
"output_config": { "effort": "medium" }

According to the guide, adaptive thinking is on by default, so a response can begin with one or more thinking blocks even when the request doesn't set thinking. The guide's instruction is to leave thinking unset or set it to adaptive, "and use effort as the lever". Where Haiku 4.5 ran without thinking, or with a small budget to save tokens, it says to choose a lower effort level. In the guide's words: "At a lower level the model thinks less, and it can skip thinking entirely on simpler requests."

The guide lists several consequences that affect how applications are built and budgeted:

  • Output limits. Thinking tokens count toward max_tokens. A request with a small max_tokens can stop with stop_reason: "max_tokens" after a thinking block and before any text. The guide says to raise the limit or choose a lower effort level.
  • Multi-turn conversations. Thinking blocks from earlier assistant turns stay in context and count as input tokens, where Haiku 4.5 kept only the latest turn's. Anthropic says these conversations carry more input tokens than the tokenizer change alone explains, and points to thinking block clearing for removing older blocks.
  • Reading responses. Code that reads the first content block as the answer should select blocks by type instead.
  • Thinking display. By default, each thinking block comes back with an empty thinking field and only a signature. Summarised thinking has to be requested with "display": "summarized".

Other changes in the migration guide

The guide's checklist for moving from Haiku 4.5 includes further changes unrelated to thinking:

  • Sampling parameters. temperature, top_p and top_k should be removed. Any temperature other than 1 or top_p other than 0.99 returns a 400 error, as does any top_k value or a request that sets both temperature and top_p.
  • Prefill. Ending a request with an assistant turn for the model to continue is rejected with a 400 error. Requests should end with a user turn.
  • Computer use. The Claude API and Google Cloud use the computer_toolset_20260801 toolset. A request that declares computer_20250124 returns a 400 error.
  • Refusals. Developers should handle stop_reason: "refusal". The guide says Haiku 5.5 runs safety classifiers that can decline a request, and that it has no server-side fallback.
  • Priority Tier. It is not supported on Haiku 5.5.

Cost and tokens

Anthropic says Haiku 5.5 "now costs around 75% less to run" on average than Haiku 4.5. Its footnote explains the figure: the price is 90% lower for requests up to 100,000 tokens and 50% lower for longer ones, and the calculation also accounts for the updated tokenizer using slightly more tokens per task. Anthropic says about 90% of requests to Haiku 4.5 fell into the shorter category.

The listed prices per million tokens for prompts up to 100,000 tokens are $0.10 for input and $0.50 for output, against $1.00 and $5.00 for Haiku 4.5.

The migration guide gives the tokenizer change in more detail. The same input text produces approximately 30% more tokens on Haiku 5.5 than on Haiku 4.5, with the exact increase depending on the content. It says cost estimates built from Haiku 4.5's token counts need recomputing with Haiku 5.5's counts and prices, "including its higher prices for long prompts". The 75% figure is Anthropic's own average and is weighted by how requests to Haiku 4.5 were distributed. A team's bill will depend on its own mix of prompt lengths, effort settings and thinking volume.

Where Anthropic says Haiku 5.5 fits

Anthropic draws a line between Haiku 5.5 and its larger models. It says Sonnet 5.5 and Opus 5.5 "remain better choices for complex agentic coding tasks" like those measured by Terminal-Bench 4.0. Haiku 5.5, it says, is best suited to more narrowly scoped tasks, such as compaction, summarisation or subagent work, that might otherwise have been cost-prohibitive with previous versions of Claude.

The launch post carries comments from customers, including Asana, HubSpot, AlphaSense, Box, Rogo and Cognition. They are testimonials supplied by Anthropic's customers, and they describe early testing.

A view

The headline numbers are Anthropic's, but the effort setting changes something developers can test for themselves. A per-request setting on a single model adds a choice of how much thinking to give each task alongside the choice of model, and the migration guide shows that question now reaches into token limits, context growth and cost estimates. Teams moving from Haiku 4.5 have more to check than the model ID. Haiku 5.5 defaults to medium effort, and Anthropic's documentation reserves xhigh for long-horizon work and max for tasks needing the deepest reasoning, so the higher settings are worth testing against a team's own evals rather than switching on by default.

Sources

Anthropic (2026) 'Introducing Claude Haiku 5.5', Anthropic, 7 October 2026, available at: https://www.anthropic.com/claude-haiku-5-5 (Accessed: 11 October 2026).

Anthropic (n.d.) 'Claude Haiku 5.5 migration guide', Claude Platform Docs, Anthropic, available at: https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide (Accessed: 11 October 2026).

Anthropic (n.d.) 'Effort', Claude Platform Docs, Anthropic, available at: https://platform.claude.com/docs/en/build-with-claude/effort (Accessed: 11 October 2026).