Claude Sonnet 5.5: Price, Specs and When to Use It

Claude Sonnet 5.5 is Anthropic's everyday model: $2 per million input tokens, $10 per million output tokens, a 1M-token context window, 128K max output, and text and image input. Anthropic calls it "the best combination of speed and intelligence." It is closed-weight, launched 28 September 2026, and costs half as much as Opus 5.5. Checked 31 October 2026.

Specs and prices

Rows come from Anthropic's Sonnet 5.5 model page and pricing page on 31 October 2026. The 1M window is included at standard pricing. The minimum cacheable prompt is 512 tokens. US-only inference through inference_geo adds a 1.1x multiplier.

Anthropic's launch post says Sonnet 5.5 is priced the same as Sonnet 5 but typically needs far fewer tokens, so it costs up to 30 percent less per task. It also says cache reads were halved from $0.20 to $0.10. The Haiku 5.5 post repeats that and says this makes Sonnet 5.5 about 20 percent cheaper on most agentic work. Both are vendor claims about typical workloads.

What Sonnet 5.5 is for

Anthropic describes Sonnet 5.5 as "strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets," with "a sharp eye for design." It positions it as a "faster, lower-cost complement" to Opus 5.5, which it calls "clearly stronger at complex, open-ended work requiring sustained judgment."

That gives a simple rule for coding. Use Sonnet for defined tasks: a bug with a failing test, a refactor of one module, a review, a first draft of a feature. Move to Opus when the task spans many files, needs many steps of planning, or has already failed twice on Sonnet. Claude Code's opusplan mode formalises this by using Opus to plan and Sonnet to execute.

Published benchmarks (vendor-reported)

The table below is from Anthropic's Sonnet 5.5 launch post. Every figure is vendor-reported by Anthropic. The GPT column in that post is GPT-6 Sol, and Anthropic says it substituted GPT-5.6 Sol on benchmarks where OpenAI had not published GPT-6 Sol.

A few points about reading it. Sonnet scoring above Opus on Terminal-Bench (70.6 against 66.4) looks odd next to Anthropic's own statement that Opus is stronger on complex work, and the two columns use different effort settings, with the Opus figure carried from its own post. The Sonnet post states no effort level for its Terminal-Bench cells, and its footnote only says the Opus 5.5 figure is at Xhigh effort, its highest score. It does not explain why Sonnet scores above Opus on this row. The post's chart text adds that at Medium effort, the default in the Claude apps, Sonnet 5.5 far exceeds Sonnet 5's best Terminal-Bench score for under a tenth of the cost per task. It also says Sonnet 5.5 at High effort matches GPT-6 Sol's best FrontierCode score at about a fifth of the cost per task.

The Haiku 5.5 post lists slightly different Sonnet 5.5 numbers (for example 1840 on GDPval-AA and 83.9% on an offline OSWorld subset), probably from different settings. We treat both tables as vendor-reported and would not mix them. We have not run any of these tests.

Customer quotes in the post are also vendor-selected. Examples include a game studio reporting 90 percent of tasks passed in a multi-step editor benchmark and a finance team reporting about 121K tokens per answer against 497K for Sonnet 5. Those are single-customer results, and we cannot weigh them.

Behaviour changes to plan for

The Sonnet 5.5 "what's new" page lists five breaking changes from Sonnet 5.

1. Thinking. thinking: {"type": "disabled"} returns a 400 error. To switch off up-front thinking, send thinking: {"type": "between_tools"}. It works at low, medium and high effort and returns an error at xhigh or max.
2. Forced tool use fails. tool_choice of any or a named tool returns a 400 error.
3. Thinking blocks are bound to the model and conversation. Sonnet 5.5 reads blocks from Sonnet 5, Opus 4.8, Haiku 4.5 and earlier, and from Haiku 5.5 on the Claude API, but not from Opus 5, Opus 5.5 or Fable.
4. Old computer use tool rejected on the Claude API and Google's managed platform. Use computer_toolset_20260801.
5. Advisor tool pairings. The advisor tool rejects four models as advisors: Opus 4.8, Opus 4.7, Sonnet 5 and Haiku 5.5.

One more change alters output without failing a request: text between tool calls now comes back in thinking blocks. An app that streams that text goes quiet between calls until it sets a display value or uses between_tools. Also, setting temperature, top_p or top_k returns a 400 error. The tokenizer is the same as Sonnet 5's, so token counts do not change against that model.

Where it is served

Anthropic lists the Claude API, Amazon Bedrock, Claude Models on Google Cloud (both managed variants), Microsoft Foundry and Claude Platform on AWS. Of the hosts in our directory, AWS Bedrock is the one that applies; see also AI hosting. The big gotcha is the alias. Claude Code's docs say sonnet means Sonnet 5.5 on the Anthropic API, Sonnet 4.6 on Claude Platform on AWS, and Sonnet 4.5 on Bedrock, Google Cloud's agent platform and Microsoft Foundry. If your team runs Claude Code through a cloud, name claude-sonnet-5-5 or set ANTHROPIC_DEFAULT_SONNET_MODEL.

Which coding agents can use it

Claude Code uses it as the sonnet alias, and the sonnet[1m] alias has no effect now because Sonnet 5.5 and 5 have a native 1M window. GitHub Copilot lists Sonnet 4, 4.6, 5 and 5.5. Cursor has a Sonnet 5.5 page and bills it at the list price from prepaid credit on Premium and above, with a $0.10 cache read. OpenCode lists Sonnet 5.5 at $2 and $10 among its provider prices. Cline and Aider can reach it with an Anthropic key. Kiro lists Sonnet 4.5 on its free tier and Sonnet 4.6 at a stated multiplier, and its models page now lists Sonnet 5.5 and Sonnet 5 at 1.3x credits relative to Auto.

Cost in practice

Our arithmetic from the list prices, not a measurement. A call with 50,000 input tokens and 2,000 output tokens costs $0.10 for input and $0.02 for output, so $0.12 before thinking tokens and caching. With 45,000 tokens read from cache at $0.10 per million, input falls to about $0.014, and the call costs about $0.034. Batch halves either figure.

A coding agent that makes 40 calls in a session, each with a 100,000-token cached prefix, 5,000 fresh input tokens and 1,500 output tokens, would cost roughly: cache reads 4 million tokens at $0.10 is $0.40, fresh input 200,000 tokens at $2 is $0.40, output 60,000 tokens at $10 is $0.60, for about $1.40 in total. These figures leave out cache writes and thinking tokens, which add to it. Run the same sum at Opus rates ($0.20, $4, $20) and you get about $2.80. Opus is twice the price on every line, so the session cost doubles. Whether that is worth it depends on how many retries it saves.

Sonnet price history

The pricing page lists several Sonnet generations side by side, which helps if you are still running an older one.

Sonnet 5's $2 and $10 was first announced as introductory pricing through 31 August 2026, with a rise to $3 and $15 planned for 1 September. The pricing page now says the introductory price is standard and the rise will not happen.

Effort and thinking in plain words

Sonnet 5.5 decides how long to think, and the effort parameter steers it. Anthropic's levels run low, medium, high, xhigh and max. The API default is high, while the Claude apps use medium. That split matters when you compare experiences: a developer who tries Sonnet in the app at medium effort and a team that calls the API at high effort are not running the same configuration.

Anthropic's launch post says that on several evaluations Sonnet 5.5 at max effort performs comparably to Opus 5.5, and that at low or medium effort it beats Sonnet 5's best score for about a tenth of the cost per task. The shape of the argument is that effort is a cheaper way to buy intelligence than moving up a model tier, at least up to a point. At the top of the curve, the post says, "Opus 5.5 remains clearly stronger" on open-ended work.

The between_tools setting is worth a note. It turns off up-front thinking but keeps the short progress updates the model writes between tool calls. For an interactive agent, where users watch a spinner, that cuts the wait before the first action. It is not available at xhigh or max, and with it effort cannot change mid-conversation.

Choosing by task

These are our mappings from Anthropic's descriptions. We have not measured them.

Local hardware needs

None. Sonnet 5.5 is not downloadable. For local coding models, read best open source LLM for coding and open source LLMs.

Who should pick it

Our take

Sonnet 5.5 is the model we would put in front of most teams first. It has the same context and output limits as Opus, the price is half, and Anthropic's own benchmark table has it close to Opus on knowledge work, computer use and chart reading. Cache reads at $0.10 make long agent sessions cheaper than the headline price implies.

The caveats are real. The effort default is high, so a call can think more than you expect, and the breaking changes around thinking and tool choice mean a straight swap from Sonnet 5 can fail until you fix the request. We also treat the Terminal-Bench number with caution until the footnotes explain it. We have not tested Sonnet 5.5, and this opinion comes from published documents. For a direct view of the pair, see Claude Opus vs Sonnet, and for the whole family, Claude models.

Checklist

• Remove thinking: disabled and forced tool_choice from existing code.
• Drop temperature, top_p and top_k.
• Choose an effort level on purpose; the default is high.
• Stream or display thinking text if users need to see progress between tool calls.
• Cache static prefixes, minimum 512 tokens.
• Pin claude-sonnet-5-5 on every cloud.
• Recheck pricing before committing; this page is dated 31 October 2026.

Questions we hear

Should I move from Sonnet 5? The price is the same, Anthropic says it uses fewer tokens, and cache reads are half. Test the breaking changes first.

How does it compare with GPT-6.1 Sol? Both list at $2 and $10. See GPT models for the OpenAI side.

Is Haiku enough? For classification and subagent steps, per Anthropic. See Claude Haiku.