Skip to content
Putting technology to work.
Insights to guide decisions and action.

Search articles

What to check before switching to Claude Sonnet 5.5

Table of contents · 8 items

Correction on October 16, 2026: We corrected four points in line with official announcements made after publication: the price of cache reads for Sonnet 5.5 (cut from $0.20 to $0.10 on October 7), fast mode for Opus 5.5 leaving research preview (October 15), Sonnet 4.5 being deprecated and retiring on the Claude API on November 30 (announced on September 30), and the release of Haiku 5.5 (October 7).

You have built the Claude API into internal AI apps and business systems, with claude-sonnet-5 as the default model. Then Sonnet 5.5 comes out at the same price, and you are told, "If just changing the ID makes it faster and cheaper, let's switch right away." Meanwhile, you also have not decided which of the workloads running on Opus 5.5 can be moved down to Sonnet 5.5.

Sonnet 5.5 keeps the same pricing, but some settings in requests written for Sonnet 5 will return 400 errors if sent as is. Some changes also alter only what appears on screen without causing errors. Within what we could confirm in the official documentation, we sort out the items to check before switching.

Same pricing as Sonnet 5, and the ID is claude-sonnet-5-5 with no date

Sonnet 5.5 was released on September 28, 2026, and is the second model in the Claude 5.5 family. Anthropic positions it as "a faster, lower-cost model that complements Opus 5.5." Among the current models listed on the models overview page, Sonnet 5.5 and the models above and below it compare as follows.

ItemSonnet 5.5Opus 5.5Haiku 4.5
Claude API IDclaude-sonnet-5-5claude-opus-5-5claude-haiku-4-5-20251001
Input / output (per million tokens)$2 / $10$4 / $20$1 / $5
Cache reads$0.10$0.20$0.10
Context window / max output1M / 128K1M / 128K200K / 64K
ThinkingAdaptive (on by default)Adaptive (always on)Extended thinking
Default effort in the APIhighmediumNot supported
Retirement no earlier thanSeptember 28, 2027 or laterSeptember 22, 2027 or laterOctober 15, 2026 or later

Sonnet 5.5's input and output prices, as well as its cache write and batch processing prices, are the same as Sonnet 5 ($2.50 for 5-minute cache writes, $4 for 1-hour cache writes, and 50% off both input and output with the Batch API). Cache reads were $0.20, the same as Sonnet 5, at publication, but dropped to $0.10 (0.05 times the input price) on October 7, 2026. The tokenizer is also the same as Sonnet 5, so the same text produces the same number of tokens.

Anthropic's announcement page says output is more than 30% faster than Sonnet 5, and that in its own tests the cost per task was up to 30% lower (because fewer tokens are used). Your own workloads will not necessarily see the same ratios.

It is available on the Claude API, Amazon Bedrock (ID anthropic.claude-sonnet-5-5), Claude Platform on AWS, Google Cloud and Microsoft Foundry. On every platform other than Amazon Bedrock, the ID is claude-sonnet-5-5.

Five places where Sonnet 5 code can return 400 errors

The official "What’s new in Claude Sonnet 5.5" lists five breaking changes that affect code running on Sonnet 5.

Setting sent to Sonnet 5Handling in Sonnet 5.5Replacement Destination
thinking: {"type": "disabled"}400 error{"type": "between_tools"} (effort at high or below)
any / tool in tool_choice400 error (also when counting tokens)auto + strict tool use, with the prompt specifying when to call the tool
Editing earlier history and resending thinking blocksMay return a 400 errorMake the conversation append-only
computer_20251124 (Claude API, Google Cloud)400 errorcomputer_toolset_20260801
Opus 4.8, Opus 4.7 or Sonnet 5 as the advisor400 errorAn advisor that Sonnet 5.5 accepts

The first row is the change also singled out in the "Getting started" section of the announcement page. In Sonnet 5.5, adaptive thinking runs even if nothing is specified. Workloads that turned thinking off should be rewritten to use between_tools, the lowest setting. It cannot be combined with xhigh or max effort. For requests that do not use tools, the response will be text only.

# Sonnet 5
client.messages.create(model="claude-sonnet-5", thinking={"type": "disabled"}, ...)
# Sonnet 5.5
client.messages.create(model="claude-sonnet-5-5", thinking={"type": "between_tools"},
                       output_config={"effort": "high"}, ...)

The second row applies to workloads that force a specific tool call for classification or field extraction. With auto, the model may answer without calling a tool, so specify in the prompt when it should call one. The migration guide at the time of publication noted that structured outputs (including strict tool use) were not available for Sonnet 5.5 on Amazon Bedrock, but the November 13 version no longer has that note. The strict tool use page also lists Amazon Bedrock as supported. If you use Amazon Bedrock, it is safer to confirm before switching that requests with strict go through.

The third row is the prefix check (a mechanism that checks whether the system prompt, tools and earlier messages that come before a thinking block were modified after the block was created). The What’s new page, the migration guide and the prefix check page, as checked on November 13, describe the result when Sonnet 5.5 fails this check as "Possible 400 error". This means it may return a 400 error, or it may not. According to the errors page, the 400 error message includes "The block is bound to a different conversation." When no error is returned, the block that failed the check is still sent to the model and billed, and the model answers based on thinking about the content before it was modified. Message Batches items do not return an error unless you specify a setting, and the block may be dropped.

At publication (September 30), the What’s new page described the cases where an error is returned by default as "the Claude API, Amazon Bedrock and Google Cloud, for accounts created on or after August 31, 2026 (UTC)". The November 13 version no longer has this account creation date condition. The September 28 release notes and announcement do not mention a creation date condition for Sonnet 5.5 either. As for where the check runs, the What’s new page says "the Claude API and some cloud platforms", and the linked support table lists Amazon Bedrock, Claude Platform on AWS, Google Cloud (Google managed) and Microsoft Foundry for Sonnet 5.5.

Either way, code that rewrites the history will either stop with an error or answer with thinking based on outdated assumptions. Regardless of when the account was created, only append to the conversation, and use a system message inserted partway through the conversation to change instructions. For requests that do not return an error, adding the beta header thinking-binding-controls-2026-08-01 puts a thinking_mismatch_allowed entry in the response's input_transformations, so you can notice them there.

There is one more point where the shape of responses changes without any error. Notes longer than one or two sentences written between tool calls are, by default, returned as thinking blocks with empty contents. Screens that stream intermediate progress to users go silent without any error. Using between_tools or specifying thinking.display restores the previous behavior.

Behavior that changes without errors

Even without changing your code, the following differences appear.

  • The effort scale has changed. The same high does not produce the same amount of thinking as on Sonnet 5, and the official docs ask you to re-measure. As a starting point, use high for general workloads, medium for agentic workloads with well-defined steps, and medium or low for latency-sensitive chat.
  • More kinds of requests are refused. Refusals are returned as stop_reason: "refusal", and stop_details contains one of five categories: cyber, bio, frontier_llm, reasoning_extraction or general_harms. The server-side fallback in the Claude API (beta) retries only cyber and frontier_llm refusals on Sonnet 5.
  • The minimum cache length is shorter. The minimum length of a cacheable prompt has dropped from 1,024 tokens on Sonnet 5 to 512 tokens.
  • Thinking blocks can be used only by the account that created them. If sent from a different account, the blocks are dropped and the request succeeds.

If you are moving from Sonnet 4.6 or earlier, there are even more changes. Remove temperature, top_p and top_k (the model page says specifying them returns a 400 error). Thinking budgets (budget_tokens) are not available either. The same text also uses about 30% more tokens than on Sonnet 4.6, Sonnet 4.5 and Haiku 4.5.

Workloads to move down from Opus 5.5 and up from Haiku 4.5

We covered Opus 5.5 pricing and how to review costs in A cheaper new model does not automatically lower your bill. Here we look at where Sonnet 5.5 fits.

The official "Choosing a model" page positions Opus 5.5 for complex agentic coding and enterprise work, and Sonnet 5.5 for everyday coding and business workloads where you want both speed and capability. The announcement page lists, for Sonnet 5.5, Sonnet 5 and Opus 5.5 respectively, 70.6%, 10.3% and 66.4% on Terminal-Bench 4.0 (the Opus 5.5 figure is at xhigh effort) and 55.5%, 34.1% and 57.8% on CursorBench 4.0. The same page also says Opus 5.5 is clearly stronger on complex work involving long chains of judgment, so the closeness of the numbers alone cannot decide which workloads to move down.

When routing workloads from Opus 5.5 to Sonnet 5.5, the following points change.

  • Sonnet 5.5 does not read thinking blocks from Opus 5, Opus 5.5, Fable or Mythos. If you switch from Opus 5.5 partway through a conversation, subsequent responses do not carry over that thinking (the request succeeds, and dropped blocks are not billed).
  • Fast mode is not available on Sonnet 5.5 (as of November 13, it covers Fable 5.2, Mythos 5.2 and Opus 5.5, plus Opus 5 in research preview and the deprecated Opus 4.8). Fast mode on Opus 5.5 left research preview on October 15, 2026, and no longer requires an access request or a beta header. Fast mode on Opus 4.8 was deprecated on October 30 and will stop working on November 30, 2026. Pricing and supported platforms are covered in Opus 5.5 fast mode no longer requires an application: what to check before using it.
  • Opus 5.5 cannot turn thinking off, but Sonnet 5.5 can turn off upfront thinking with between_tools.

When moving up from Haiku 4.5, the unit price rises and the same text also uses more tokens, so the migration guide asks you to re-measure costs. Claude Haiku 5.5, aimed at high-volume workloads, was released on October 7, 2026. Its price is $0.10 for input and $0.50 for output (per million tokens) for prompts of up to 100,000 tokens, and five times that for prompts over 100,000 tokens. If you run volume-driven workloads on Haiku 4.5, the editorial team's view is to compare with Haiku 5.5 before moving up to Sonnet 5.5. We also cover how to assign models to each workload in How to decide which model to use for development.

Sonnet 5 remains available until at least June 30, 2027

The "Model deprecations" page lists claude-sonnet-5 as Active and states that it will not be retired before June 30, 2027. Customers using a public model are notified at least 60 days before it is retired.

It is actually the older generations whose deadlines are closer. claude-sonnet-4-5-20250929 was deprecated on September 30, 2026, and will be retired on the Claude API on November 30, 2026. The recommended replacement is Sonnet 5.5. claude-haiku-4-5-20251001 has passed its target retirement date of October 15, 2026, but as of November 13 it remains Active and no deprecation has been announced. These target dates mean "not retired before this date"; the actual retirement date is set by announcement. These dates apply to the Claude API, Claude Platform on AWS, Google Cloud (Anthropic managed) and Microsoft Foundry; Amazon Bedrock and Google Cloud (Google managed) set their own deadlines. You can check which models are used by which keys in the CSV exported from the Usage page in the Claude Console. What to check when moving from Sonnet 4.5 is summarized in Claude Sonnet 4.5 retires on November 30: what to check for migration.

Steps for switching

Editorial diagram showing the five items to check before switching to Sonnet 5.5 (model ID and pricing, settings that can return 400 errors, requests before and after rewriting, re-measuring effort, and deadlines for older models) and the official pages for checking each

The diagram maps the items to check to the official pages for verifying them. From here on, these are the editorial team's recommendations based on the official documentation.

  1. Find the relevant settings in your code. Search for "disabled", tool_choice, temperature, top_p, top_k, computer_20251124, advisor settings, and code that rewrites history and resends it, then compare them against the table above. The migration guide also describes using /claude-api migrate in Claude Code to handle the rewrite and build a checklist.
  2. Compare two or three effort levels. Do not reuse the value you used on Sonnet 5; measure quality and cost per request with the same kinds of inputs as production.
  3. Check screens that show intermediate progress. Look at the actual screens to see whether nothing is shown between tool calls anymore.
  4. Check how refusals are handled. Decide whether receiving stop_reason: "refusal" should stop processing as an error or retry with another model.
  5. Make the model ID a configuration value and switch per workload. Do not switch all workloads at once; switch them one at a time and keep the ability to revert to Sonnet 5.

If model names are scattered across multiple places in your code, step 5 turns into a refactoring project. We covered the idea of consolidating access into a single entry point in Should you put a gateway in front of your internal AI?.

On September 30, 2026, we cross-checked by directly opening the Claude release notes (the September 28, 2026 entry), the Claude API release notes, the models overview in Claude Platform Docs, the Sonnet 5.5 model page, What’s new, migration guide and prompting guide, pricing, Model deprecations, Choosing a model, and Anthropic's Sonnet 5.5 announcement page. Because we had no API key, we did not send requests to Sonnet 5.5 and did not verify error contents, speed, token counts, cost per request, or behavior on Amazon Bedrock, Google Cloud or Microsoft Foundry. Benchmark figures are Anthropic's published values. On October 16, 2026, we reopened the Claude API release notes (the September 30, October 7 and October 15 entries), pricing, Fast mode, Model deprecations, the models overview and What’s new, and corrected the four points above. On November 13, 2026, we opened What’s new, the migration guide, the model page, the release notes, the announcement, pricing, Fast mode, Model deprecations and the models overview, as well as the prefix check, Preserved thinking, errors, strict tool use and Claude in Amazon Bedrock pages, and updated the article (desk research).

For switching models in internal systems built on the Claude API, or reworking them into an architecture where models are easy to swap, contact GleamHub.

Revision history

November 13, 2026: We updated the conditions under which rewriting the history causes a 400 error to match the official documentation as of November 13 ("may or may not return a 400 error"), and noted that the account creation date condition in the documentation at publication is not in the current documentation. We also updated the descriptions of strict tool use on Amazon Bedrock, the models covered by fast mode, the sampling parameters, and the platforms to which the retirement dates apply.

Sources

Share this articleXFacebook
Kakeru Suzuki

Fascinated by the possibilities of technology, has had a deep interest in programming and digital art since student days

Turn this article's theme into your company's next step

Thinking together, starting from the work you entrust to AI.

We organize your current operations and data to define the scope entrusted to AI, what humans should review, and how to run trials.

  • Target operations
  • Data to use
  • How to verify effectiveness
Consult on AI adoption for your business

You can consult with us from the initial conceptual stage. Details from this article will be carried over to the inquiry form.

Receive the latest articles by email