What changed
Anthropic's building guide lists the per-million-token rates below. Every Sonnet 5.5 price, including batch processing and prompt caching, matches Sonnet 5, so an ID swap leaves the per-token bill unchanged. US-only inference, with inference_geo set to "us", costs 1.1 times the standard price. Anthropic says the task bill still moves because Sonnet 5.5 typically uses fewer tokens, and that most work costs up to 30% less. A 2000 by 1500 image on the high-resolution tier, up to 2576 pixels on the long edge, costs about 2.5 times the tokens of Sonnet 4.6, Sonnet 4.5, or Haiku 4.5. Downscale when that detail is unnecessary.
On the Claude API, which the Claude Code changelog calls the Anthropic API, the ID is claude-sonnet-5-5. Claude Platform on AWS, Google Cloud, and Microsoft Foundry are separate hosts in the same guide, and they take that same ID. Amazon Bedrock uses anthropic.claude-sonnet-5-5. The guide says Sonnet 5.5 is available today on those platforms, with Foundry limited to Global Standard deployments. Context is 1 million tokens, native, with no beta header. Max output is 128k tokens, or up to 300k on the Message Batches API with the output-300k-2026-03-24 beta header. The knowledge cutoff is June 2026. The tokenizer matches Sonnet 5. The shortest cacheable prompt is 512 tokens, versus 1,024 on Sonnet 5, and a cache read costs a tenth of the input price.
The Claude Code changelog says claude-sonnet-5-5 is now the default Sonnet on the Anthropic API, with 1M context, $2/$10 per million tokens, and cache reads at $0.20 per million tokens. The building guide says that from Claude Code v2.1.284 and Agent SDK for TypeScript v0.3.284 or later, the sonnet alias is Sonnet 5.5 at medium effort, thinking stays on, and there is no fast mode. The default model stays Opus 5.5. For well-scoped tasks, switch with /model sonnet.
| Rate | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Cache write, 5 minutes | $2.50 | $5 |
| Cache write, 1 hour | $4 | $8 |
| Cache read | $0.20 | $0.20 |
Who it is for
Anthropic's workload table starts well-scoped coding, high-volume everyday development, polished documents, slides, and spreadsheets, and repeated well-defined agent tasks on Sonnet 5.5. Careful judgment, long-horizon agentic work, and the hardest problems start on Opus 5.5. The guide says Sonnet 5.5 fits a clear spec with a way to check the result.
Anthropic quotes Daniel Vogel, COO of Epic Games, on early tests that held a higher-tier quality bar for a system design audit and a data-flow review, across tens of thousands of lines of gameplay code and multi-hour tasks, with less prescriptive prompting. That is Epic's early testing as Anthropic published it. Haiku 5.5 is planned for the coming weeks, for high-volume, low-latency work. The same day, @claudeai pointed to a foliage simulator by @_re_pete and bouncing-ball tests by @forwardeditor, each set up as Sonnet 5 beside Sonnet 5.5. Those posts include no scores or prompts. Lance Martin's caption on the building guide shows each model writing code that repaints one photograph, ordered as the photograph, Sonnet 5, Sonnet 5.5, then Opus 5.5, and names no winner.
What to change in a Sonnet 5 integration
Send model claude-sonnet-5-5 and read each content block by type. Adaptive thinking is on by default, so a response can begin with a thinking block, and code that reads content[0].text breaks. The guide's sample sets output_config effort to medium as an example. Default effort is not one setting across products. The model-details table gives high on the Claude API, which the guide says is so you start from strong results, and medium in Claude Code. The pricing section says surfaces can differ, for example high on Claude Platform and medium in Claude Code.
Effort levels are low, medium, high, xhigh, and max, and they are recalibrated, so a Sonnet 5 setting does not carry over. Start at high unless the workload is agentic or latency-sensitive. For agentic coding and multistep tool use, start at medium on well-specified tasks and move to high when the work is harder or longer. For chat and other latency-sensitive work, start at medium or low. Use xhigh or max only where your evals show a quality gain. Thinking counts toward max_tokens. For agentic coding, set max_tokens to 128,000 and stream the response. Asking in the system prompt to think less does not reliably reduce thinking. In Claude Code, /claude-api migrate this project to claude-sonnet-5-5 applies the model ID swap and the breaking parameter changes.
Anthropic says existing Sonnet 5 prompts should perform well. Remove refusal steering, tool-call retry shims, and "do not be lazy" lines, then re-run evals before other tuning. At low effort, Sonnet 5.5 sometimes skips a check that exercises a code change. A decline is HTTP 200 with stop_reason set to refusal. stop_details is one of cyber, bio, frontier_llm, reasoning_extraction, or general_harms. Asking the model to write its reasoning in the response invites a reasoning_extraction decline. The guide says to read that reasoning from summarized thinking instead, with adaptive thinking and display set to summarized. Server-side fallback is beta on the Claude API, with fallbacks set to "default". It retries only cyber and frontier_llm declines, and it retries those on Sonnet 5. Anthropic says its automated behavioral audit shows Sonnet 5.5 matches or improves on Sonnet 5 for most alignment and honesty measures. The guide gives no item scores for that audit. Anthropic also says this is the first Sonnet with cybersecurity safeguards similar to its most capable models, and that most routine software development is unaffected. The guide adds per-message effort (beta), mid-conversation system messages, and mid-conversation tool changes (beta), without request shapes on this page. Changing top-level effort between requests invalidates the cache. Per-message effort keeps it.
- A request with no thinking field runs adaptive thinking. thinking type disabled returns 400. between_tools turns off upfront thinking, so thinking happens between tool calls, and only at low, medium, or high. At xhigh or max it returns 400, so those levels use adaptive thinking. The guide's own sample drops an xhigh request to high when it switches to between_tools. between_tools rejects display, budget_tokens, and block_binding, and effort cannot change mid-conversation. It works on every platform that offers Sonnet 5.5, with no beta header. Update the SDK if between_tools is missing.
- tool_choice of type any or tool returns 400, including on the token-counting endpoint. Send auto, strict true, additionalProperties false on every object, and a prompt that names when to call the tool.
- Sonnet 5.5 reads Sonnet 5 thinking blocks, so a conversation switched forward keeps that reasoning. The guide says no other model reads Sonnet 5.5 thinking blocks.
- On the Claude API and Google Cloud, computer use is only the toolset computer_toolset_20260801. Declaring the older computer_20251124 tool returns 400. Drop the computer-use-2025-11-24 header. Drop the fine-grained-tool-streaming-2025-05-14 header when it sits beside a toolset, because that pairing returns 400, and set eager_input_streaming to true on each tool that needs it. The agent loop then handles member tool_use blocks, batch actions, and toolset_name on results. Amazon Bedrock still accepts computer_20251124.
- With the advisor tool, a Sonnet 5.5 executor rejects Opus 4.8, Opus 4.7, and Sonnet 5 as advisors. Accepted advisors include Opus 5.5, Opus 5, and Sonnet 5.5. Advice comes back encrypted as an advisor_redacted_result block, and application code cannot read the text inside it.
- Notes between tool calls longer than a sentence or two return as progress thinking blocks, empty at the default display. With adaptive thinking, set display to summarized, or to updates with the thinking-display-updates-2026-08-18 beta header. With between_tools, that text comes back without display.
What this page leaves open
Rate limits are separate from Sonnet 5, with the same default tier values. Priority Tier is available on the Claude API. Zero data retention is for eligible customers. This page prints no tier numbers, no Priority Tier price, no eligibility rules, and no separate batch percentage. Haiku 5.5 is described as joining in the coming weeks. The Cyber Verification Program will soon include Sonnet 5.5 for legitimate security work. Neither line has a date. Computer-use rules here cover the Claude API, Google Cloud, and a Bedrock exception. Claude Platform on AWS and Microsoft Foundry are not given a computer-use rule on this page. Anthropic's automated behavioral audit is summarized without item scores. "Available today" is the guide's wording next to the 28 September 2026 announcement.
Source
Also from
Sources and editorial approach
This account relies on ClaudeDevs' announcement of 28 September 2026, Anthropic's building guide for Sonnet 5.5 on claude.dev, the Sonnet 5.5 note at the top of the Claude Code changelog, and two @claudeai posts from that afternoon about a foliage simulator and bouncing-ball tests. The date checked is 28 September 2026. Later changelog entries are outside this account of the model. Anthropic points to a separate migration guide and a prompting guide, and those pages are not covered here. No API call, Claude Code version check, cloud-console check, or token-bill comparison was run.