Skip to content
Claudexia TeamUpdated GUIDE

Claude Sonnet 4.5 retires November 30: replacement and migration

Anthropic retires claude-sonnet-4-5-20250929 on November 30, 2026. Replacement, a migration checklist, Sonnet 5.5 prices and which request settings now fail.

Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) retires on November 30, 2026. Anthropic recommends moving to claude-sonnet-5-5. If an old ID is left in your code, in Claude Code or in an editor's settings, requests will return an error after that date, so the search should start now.

This guide is for anyone who calls Sonnet 4.5 directly at Anthropic or through an intermediary. Dates and replacements come from Anthropic's model deprecations page, which we opened on November 13 and 17, 2026. Request changes come from the Sonnet 5.5 migration guide.

What retires and when

For each model the documentation lists a status, a deprecation date and a column called "Tentative retirement date". These are the rows that touch Sonnet 4.5 and its neighbors:

Model (API ID)StatusDeprecatedTentative retirementReplacement
claude-sonnet-4-5-20250929DeprecatedSeptember 30, 2026November 30, 2026claude-sonnet-5-5
claude-opus-4-5-20251101Activenonenot sooner than November 24, 2026none named
claude-haiku-4-5-20251001Activenonenot sooner than October 15, 2026none named
claude-sonnet-4-6Activenonenot sooner than February 17, 2027none named
claude-opus-4-6Activenonenot sooner than February 5, 2027none named

In Anthropic's terms, Deprecated means the model still works but is no longer recommended: it gets a replacement and a retirement date. Retired means requests to it fail. The Sonnet 4.5 row is the only one with a real retirement date. The others say "not sooner than", which is a lower bound, and while a model is Active no retirement has been announced. The Haiku 4.5 date has already passed, and the model is still Active.

Anthropic promises at least 60 days' notice before retiring a publicly released model. Between the September 30 notice and the November 30 retirement there are 61 days. On our reading of those rules, Opus 4.5 will not be retired on November 24 without a fresh notice, but that is our interpretation, not a promise from Anthropic. Set a reminder and check the page once a month.

One caveat about platforms. The dates on the page apply to platforms Anthropic operates itself: the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry. Amazon Bedrock and Claude Models on Google Cloud (Google managed) set their own schedules, which can differ. Other intermediaries have their own dates too, and we did not check them.

Migration checklist

The order is: find every mention, replace it, run a check on live requests, and only then roll out. We covered the general version in the checklist for retired Claude models. This is the Sonnet 4.5 version.

  1. If you use Claude Code, check the version with claude --version. Sonnet 5.5 needs version 2.1.284 or newer, and you update the client with claude update. On older versions the aliases point to older models.
  2. Export actual usage. In the Claude Console open Usage, click Export and read the CSV: it breaks usage down by API key and model. It also shows calls that are not in your code, such as someone else's script on your key.
  3. Search the repository for the ID. One pattern catches the dated ID, the undated alias, the cloud forms and a plain "Sonnet 4.5" in configs:
grep -rniE "sonnet[- ]4[-.]5" . --exclude-dir=node_modules --exclude-dir=.git
  1. Walk through the places grep cannot reach (table below).
  2. Replace the ID, fix the requests as described in "What changes in requests", and run the tests.
  3. Compare old and new model answers on real inputs, as described in the verification section below.
Where the ID livesWhat to look for
Code, .env, constantsmodel=, strings with claude-sonnet-4-5
CI, secrets, environment variablesseparately on every environment
Claude CodeANTHROPIC_MODEL, ANTHROPIC_DEFAULT_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, CLAUDE_CODE_SUBAGENT_MODEL, the model field in ~/.claude/settings.json and .claude/settings.json, the model field in subagent and skill frontmatter
Cursor, Cline and other editorsthe model list and the Model ID field of a custom provider
CloudsBedrock us.anthropic.claude-sonnet-4-5-20250929-v1:0, Google Cloud claude-sonnet-4-5@20250929
Dashboards, alerts, templatesthe model name in filters and in READMEs

Claude Code has three details described in its documentation. First, the sonnet alias points to Sonnet 5.5 on the Anthropic API only from version 2.1.284 (on older versions it resolves to an older model), and to Sonnet 4.5 on Amazon Bedrock and Google Cloud Agent Platform. Sonnet 5.5 itself needs Claude Code 2.1.284 or newer, and a request to it from an older version can fail, which is why updating the client is the first step of the checklist. Second, on Microsoft Foundry the default model in Claude Code is currently Sonnet 4.5, and Anthropic's dates apply on Foundry, so set another model there explicitly before November 30. Third, if the requested model has a scheduled retirement date, Claude Code prints a warning at startup. Since version 2.1.182 it also writes the warning to stderr in non-interactive mode, but only with the default text output format. With --output-format json and stream-json the warning is suppressed and will not appear in such CI logs: read the actual model from the modelUsage field of the result message instead. If you run claude -p with text output, look for the warning in the logs.

We did not check Cursor, Cline and other editors against their vendors' documentation: open the model list and the Model ID field and look for sonnet-4-5. How they look when pointed at Claudexia is covered in the guides for Cursor and Cline.

What to replace it with and at what price

Anthropic names one replacement: claude-sonnet-5-5 (released September 28, 2026). The choice is wider than that. Below are Anthropic's figures (models page and pricing) and the Claudexia Premium price list captured on November 19, 2026. Dollars per 1M tokens, input and output.

ModelAnthropicClaudexia PremiumContext and output at AnthropicRetirement in Anthropic's table
Sonnet 4.5 (claude-sonnet-4-5-20250929)$3 and $15not in the catalog200K and 64KNovember 30, 2026 (announced)
Sonnet 4.6 (claude-sonnet-4-6)$3 and $15not in the catalog1M and 128Knot sooner than February 17, 2027
Sonnet 5 (claude-sonnet-5)$2 and $10$0.62 and $2.921M and 128Knot sooner than June 30, 2027
Sonnet 5.5 (claude-sonnet-5-5)$2 and $10$0.62 and $2.921M and 128Knot sooner than September 28, 2027
Opus 5.5 (claude-opus-5-5)$4 and $20$0.88 and $4.401M and 128Knot sooner than September 22, 2027
Haiku 5.5 (claude-haiku-5-5)from $0.10 and $0.50$0.10 and $0.361M and 128Knot sooner than October 7, 2027

The November 13 check still showed claude-sonnet-4-6 ($0.68 and $2.48) and claude-haiku-4.5 in the Claudexia price list, and Haiku 5.5 was listed as claude-haiku-5.5. As of November 17 the first two are missing from the price list and from the GET /v1/models response, and Haiku 5.5 is spelled claude-haiku-5-5. The Sonnet 5.5 and Opus 5.5 prices also moved over those four days, so check IDs and prices against the live catalog rather than this article.

The Claudexia catalog lists an output limit of 32,000 tokens for claude-sonnet-5-5, while Anthropic lists 128K. If you need long answers, check the model catalog.

Sonnet 5.5 has a lower price per token than 4.5, but the bill does not necessarily fall. By the documentation, Sonnet 5.5 uses a tokenizer that produces about 30% more tokens for the same text than Sonnet 4.5, 4.6 and Haiku 4.5, depending on content. Simple arithmetic for Anthropic: the price per token drops to two thirds (2 against 3 on input, 10 against 15 on output), the token count grows about 1.3x, and you land near 0.87 of the old bill for the same text. That is our calculation from two Anthropic figures, and it ignores thinking. Adaptive thinking is on by default in Sonnet 5.5, and its tokens are billed as output. Only a run on your own requests shows the real bill.

The nearest model by generation is claude-sonnet-4-6. It is Active at Anthropic, and "not sooner than February 17, 2027" is only a lower bound: no actual date has been announced. It is not in the Claudexia catalog as of November 19. Anthropic has not announced the order in which models will be retired either. It rejects a prefilled assistant turn (so says Anthropic's guide), and we did not check its other differences from 4.5. For more on the neighbors, see the guides to Sonnet 5 and Sonnet 4.6.

What changes in requests

Anthropic's Sonnet 5.5 migration guide lists the settings that return a 400 error: five in its example request and output_format in the section on moving from Sonnet 4.5. For code written against Sonnet 4.5 all six apply if you set them:

In the requestWhat happens on Sonnet 5.5What to do
temperature, top_p, top_k400remove them and steer behavior with the prompt
thinking with budget_tokens400, thinking budgets are replaced by an effort levelset output_config.effort
thinking: {"type": "disabled"}400send {"type": "between_tools"} at high effort or lower
A prefilled last assistant turn400, error text "This model does not support assistant message prefill"replace it according to what the prefill did
tool_choice of type any or tool400, including on the token counting endpointuse auto and strict: true on the tool, and say in the prompt when the tool is needed
output_formatdeprecated, 400 without the structured-outputs-2025-11-13 beta headermove it to output_config.format

With auto the model may answer without calling the tool, and strict: true only validates the arguments against the schema, it does not guarantee a call. Anthropic's guide asks you to say in the prompt when the tool should be used. The guarantee that tool_choice of type any or tool gave is gone on Sonnet 5.5, so your code has to survive a response with no tool_use block. Strict mode supports a subset of JSON Schema, needs additionalProperties: false on every object, and a request can carry at most 20 strict tools.

A few more changes do not break the request but change the result:

  • Thinking is on even when there is no thinking field. A response can start with thinking blocks, so code that reads content[0].text will break. Read blocks by type. max_tokens covers both thinking and text.
  • Thinking text does not arrive by default, and the block carries only a signature. For readable summaries you need display: "summarized".
  • Remove the interleaved-thinking-2025-05-14 beta header, it is no longer needed. Replace fine-grained-tool-streaming-2025-05-14 with eager_input_streaming: true on the tools that need it. Remove any beta header that extends the context window.
  • Sonnet 4.5 has no effort parameter. On Sonnet 5.5 the default is high and the scale is recalibrated. Set the level explicitly and rerun your own sweep.
  • Images use a high-resolution tier: up to 2,576 pixels on the long edge and up to 4,784 tokens per image, where Sonnet 4.5 stops at 1,568 pixels and 1,568 tokens. By Anthropic's guide, a 2000 by 1500 image produces about 2.5 times as many tokens on Sonnet 5.5 as on Sonnet 4.5. Money works out differently because Sonnet 5.5 input is cheaper ($2 against $3 per 1M tokens in Anthropic's November 17 price list). Without caching that is 2.5 × 2 / 3, so such an image costs about 1.67 times as much. This is our calculation from the two input prices, not an Anthropic figure.

An example, before and after. First a request in Sonnet 4.5 style:

response = client.messages.create(
    model="claude-sonnet-4-5-20250929",
    max_tokens=1024,
    temperature=0.2,
    messages=[{"role": "user", "content": prompt}],
)
print(response.content[0].text)

And the same request on Sonnet 5.5:

response = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=4096,
    output_config={"effort": "medium"},
    messages=[{"role": "user", "content": prompt}],
)
print("".join(b.text for b in response.content if b.type == "text"))

What changed: the ID is replaced, temperature is gone, the effort level is set explicitly, max_tokens is raised so thinking does not eat the whole limit, and the answer is read by block type. The medium level comes from Anthropic's own example, so tune yours by measurement. This request follows the documentation, and we did not run it.

How to check before November 30

Unit tests that mock the provider will not catch a model swap. You need a run on live requests.

  1. Collect 20 to 50 real inputs: closed bugs, tickets with a known answer, documents with labels. We described how to build such a set in how to benchmark models yourself.
  2. Run the set on the old and the new model and compare the answers on several signals: the share of answers that do not parse in your format (JSON, schema), the stop_reason values, and the token counts in usage.
  3. Compute the cost of a task, not of a token: the tokenizer and thinking change token counts, and the price table above will not show that.
  4. Roll out to part of the traffic while Sonnet 4.5 still works, so rollback costs one line.
  5. Put the model ID in one constant or environment variable so the next replacement takes a minute. We wrote up an earlier case in the Opus 4.1 retirement.

If you are moving to Claudexia

Sonnet 4.5 is not in the Claudexia catalog. As of November 19 the catalog has claude-sonnet-5 and claude-sonnet-5-5, plus claude-opus-5-5 and claude-haiku-5-5. Claudexia spellings changed between November 13 and 17, so do not copy IDs from articles: take them from the catalog and from the GET /v1/models response.

Sign up on the registration page with a login, a password and a password confirmation, or through Google, GitHub or Telegram. There is no email field and no email confirmation. The minimum top-up is $1, and you can pay in rubles through SBP, with a Russian bank card, or with cryptocurrency. The key starts with sk_cdx_.

For the Anthropic SDK and Claude Code the base URL is https://api.claudexia.tech, without /v1:

curl https://api.claudexia.tech/v1/messages \
  -H "x-api-key: sk_cdx_YOUR_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-5-5",
    "max_tokens": 4096,
    "messages": [{"role": "user", "content": "Hello"}]
  }'

We did not run this request. GET /v1/models returns 200 even without a key, so test a key with a cheap POST /v1/chat/completions using the model gpt-5.6-luna: a wrong key gets a 401. In Claude Code the [1m] suffix is not needed for Sonnet 5.5, because the 1M window opens on its own, and the gateway setup is covered in the Claude Code setup guide. The new Fable 5.2 has its own guide. We do not give legal opinions and do not promise that a method is lawful or safe for your account. You read and accept the vendor's terms.

FAQ

When does Claude Sonnet 4.5 retire?

By Anthropic's documentation, on November 30, 2026. The notice was published on September 30. After that date requests to claude-sonnet-4-5-20250929 fail with an error. For Bedrock and Google Cloud (Google managed), check those platforms' pages: their schedules are their own.

What should replace Sonnet 4.5?

Anthropic recommends claude-sonnet-5-5. If you want something lighter and cheaper, look at Haiku 5.5, and if you want something stronger, at Opus 5.5. Test the choice on your own set of requests.

Will Opus 4.5 retire on November 24?

There is no such announcement. Opus 4.5 is Active, and November 24, 2026 sits in the tentative retirement column with the words "not sooner than". It is a lower bound. Keep watching the deprecations page.

Can I stay on Sonnet 4.5 through Claudexia?

No, the model is not in the catalog. Sonnet 4.6, which is Active at Anthropic with retirement not sooner than February 17, 2027, is not in the Claudexia catalog as of November 19 either. The Sonnet models there are claude-sonnet-5 and claude-sonnet-5-5.

Is Sonnet 5.5 certain to be cheaper than Sonnet 4.5?

Per token, yes: $2 and $10 against $3 and $15 at Anthropic. Per task, not necessarily: the same text produces about 30% more tokens, and adaptive thinking is on by default. Count it on your own requests.