## Anthropic model deprecation schedule: every Claude retirement date

URL: https://www.evalshift.dev/blog/anthropic-model-deprecation-schedule
Published: 2026-11-20
Tag: migration
Summary: Every Claude model retirement date since 2024, the active models' earliest dates, Bedrock and Google Cloud dates, and what changes off Claude Sonnet 4.5.

Takeaways:
- The next Claude model to retire is Claude Sonnet 4.5: claude-sonnet-4-5-20250929 stops working on the Claude API on November 30, 2026, and Anthropic lists claude-sonnet-5-5 as the replacement. AWS Bedrock runs it until April 8, 2027.
- Anthropic promises at least 60 days' notice before retiring a publicly released model. Notices through October 2025 gave 76 to 189 days; the five announced since December 2025 each gave 60 to 62 days.
- The "not sooner than" dates on Anthropic's deprecations page are floors, not schedules. Claude Haiku 4.5 passed its October 15, 2026 floor and is still active, because a model retires only after a deprecation notice.
- Claude Sonnet 5.5 is not a drop-in replacement for Claude Sonnet 4.5: it thinks by default, uses about 30% more tokens for the same text, and returns a 400 for temperature, top_p, top_k, assistant prefill, thinking budgets and a forced tool_choice.
- Before the 4.6 generation, aliases such as claude-sonnet-4-5 point to a dated snapshot and fail when it retires. From Claude 4.6 on, the dateless ID such as claude-sonnet-4-6 is itself a fixed snapshot with its own retirement date.

The next Claude model to retire is Claude Sonnet 4.5. Anthropic deprecated
`claude-sonnet-4-5-20250929` on September 30, 2026, and requests to it fail on the Claude API from
November 30, 2026. The listed replacement is `claude-sonnet-5-5`. After that, no Claude model on
Anthropic's API has a retirement date: every other model is either already retired or active with a
"not sooner than" date, and the only other deprecated model, `claude-mythos-preview`, has no date yet.

Three things about Anthropic's schedule catch teams out:

- **Notice is now about 60 days.** Until October 2025 Anthropic gave 76 to 189 days. The five
  retirements announced since December 2025 each came 60 to 62 days ahead, the minimum Anthropic
  promises.
- **"Not sooner than" is a floor, not a date.** Claude Haiku 4.5's floor, October 15, 2026, has
  passed and the model is still active. A model retires only after a deprecation notice.
- **Bedrock and Google Cloud keep their own calendars.** Claude Sonnet 4.5 runs on AWS Bedrock
  until April 8, 2027, and Claude Opus 4.1, retired on Anthropic's API in August, runs on Bedrock
  until January 8, 2027.

This post lists every Claude retirement since 2024, the active models' floors, the Bedrock and
Google Cloud dates, what changes in the request when you move off Claude Sonnet 4.5, and how to
find and test every Claude call before its date. The facts come from Anthropic's
[model deprecations page](https://platform.claude.com/docs/en/about-claude/model-deprecations), its
[models overview](https://platform.claude.com/docs/en/about-claude/models/overview), the
[Claude Sonnet 5.5 migration guide](https://platform.claude.com/docs/en/models/sonnet-5-5/migration-guide),
AWS's [Bedrock model lifecycle page](https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle-legacy.html)
and Google Cloud's [partner model deprecations page](https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/deprecations/partner-models),
all read on November 17, 2026. Check the page again before you act on a date.

## Every Claude retirement date

Anthropic's deprecation history, oldest at the bottom, on the Claude API:

| Model ID | Announced | Retired | Notice | Listed replacement |
| --- | --- | --- | --- | --- |
| `claude-sonnet-4-5-20250929` | September 30, 2026 | November 30, 2026 | 61 days | `claude-sonnet-5-5` |
| `claude-opus-4-1-20250805` | June 5, 2026 | August 5, 2026 | 61 days | `claude-opus-4-8` |
| `claude-sonnet-4-20250514` | April 14, 2026 | June 15, 2026 | 62 days | `claude-sonnet-5-5` |
| `claude-opus-4-20250514` | April 14, 2026 | June 15, 2026 | 62 days | `claude-opus-4-8` |
| `claude-3-haiku-20240307` | February 19, 2026 | April 20, 2026 | 60 days | `claude-haiku-4-5-20251001` |
| `claude-3-5-haiku-20241022` | December 19, 2025 | February 19, 2026 | 62 days | `claude-haiku-4-5-20251001` |
| `claude-3-7-sonnet-20250219` | October 28, 2025 | February 19, 2026 | 114 days | `claude-sonnet-5-5` |
| `claude-3-opus-20240229` | June 30, 2025 | January 5, 2026 | 189 days | `claude-opus-4-8` |
| `claude-3-5-sonnet-20240620`, `claude-3-5-sonnet-20241022` | August 13, 2025 | October 28, 2025 | 76 days | `claude-sonnet-5-5` |
| `claude-2.0`, `claude-2.1`, `claude-3-sonnet-20240229` | January 21, 2025 | July 21, 2025 | 181 days | `claude-opus-4-8` for Claude 2, `claude-sonnet-5-5` for Sonnet 3 |
| `claude-1.0` to `claude-1.3`, `claude-instant-1.0` to `claude-instant-1.2` | September 4, 2024 | November 6, 2024 | 63 days | `claude-haiku-4-5-20251001` |

The replacement column is what Anthropic's page says today, not what it said at the time. Anthropic
updates the listed replacement as newer models ship, which is why Claude 1 points to a model
released in 2025.

`claude-mythos-preview` is the one other deprecated model. Anthropic deprecated it on June 9, 2026,
and its retirement date is "to be announced."

### What the history says about the next retirement

Two patterns hold across the table.

**The notice period shrank to the minimum.** Anthropic's policy is at least 60 days' notice before
retiring a publicly released model, by email to customers with active deployments and on the
deprecations page. Through October 2025 it gave well over that: 181 days for Claude 2, 189 for
Claude Opus 3, 76 for Claude Sonnet 3.5 and 114 for Claude Sonnet 3.7. Starting with Claude Haiku
3.5 in December 2025, all five announcements since came 60 to 62 days ahead. Plan for the next one to arrive with two months' warning, which is not much time to test a
replacement on a large suite.

**Recent models lasted about a year.** Measured from the date in the model ID to the retirement
date, Claude Sonnet 3.7 and Claude Opus 4.1 both lasted exactly 365 days, Claude Sonnet 4 and Opus
4 lasted 397, and Claude Sonnet 4.5 lasts 427. The Claude 3 models lasted longer: 676 days for
Claude 3 Opus and 774 for Claude 3 Haiku. A model you pin today should be expected to need a migration within about 12 to 15
months.

## Active Claude models and their "not sooner than" dates

Every active model on the deprecations page has a tentative retirement date written as "not sooner
than":

| Model ID | Not sooner than |
| --- | --- |
| `claude-haiku-4-5-20251001` | October 15, 2026 (passed; still active) |
| `claude-opus-4-5-20251101` | November 24, 2026 |
| `claude-opus-4-6` | February 5, 2027 |
| `claude-sonnet-4-6` | February 17, 2027 |
| `claude-opus-4-7` | April 16, 2027 |
| `claude-opus-4-8` | May 28, 2027 |
| `claude-fable-5`, `claude-mythos-5` | June 9, 2027 |
| `claude-sonnet-5` | June 30, 2027 |
| `claude-opus-5` | July 24, 2027 |
| `claude-fable-5-1`, `claude-mythos-5-1` | September 1, 2027 |
| `claude-opus-5-5` | September 22, 2027 |
| `claude-sonnet-5-5` | September 28, 2027 |
| `claude-haiku-5-5` | October 7, 2027 |
| `claude-fable-5-2`, `claude-mythos-5-2` | November 9, 2027 |

Read these as the earliest a model could retire, not as a schedule. A model retires only after
Anthropic deprecates it and gives notice. Claude Haiku 4.5 passed its floor in October without a
notice, so it keeps running. Claude Opus 4.5 reaches its floor on November 24, 2026, but with no
notice as of November 17 and a 60-day minimum, it can't retire before mid-January 2027.

The floors still matter when you pick a replacement. Anthropic still lists
`claude-haiku-4-5-20251001` as the replacement for Claude Haiku 3 and Claude Haiku 3.5, but its
floor has passed, so it could be deprecated on any day with 60 days' notice. A migration off an old
Haiku that lands on Haiku 4.5 may have to be repeated soon. `claude-haiku-5-5`, with a floor of
October 7, 2027, is the longer-lived target, at a different price and with the request changes
covered below.

## Anthropic's model lifecycle: Active, Legacy, Deprecated, Retired

Anthropic uses four states:

- **Active**: fully supported and recommended.
- **Legacy**: no longer updated, and might be deprecated in the future.
- **Deprecated**: still works, but has a recommended replacement and an assigned retirement date.
  Anthropic warns that deprecated models "are likely to be less reliable than active models."
- **Retired**: requests fail.

Two other things change what "retired" means.

**Model IDs and aliases.** From the Claude 4.6 generation on, the dateless ID is the model:
`claude-sonnet-4-6` and `claude-opus-4-8` are fixed snapshots, and Anthropic says it doesn't change
the weights or configuration behind an existing ID. An update ships under a new ID, and every ID
has its own retirement date. Before 4.6, IDs carry a date, as in `claude-sonnet-4-5-20250929`, and
the Claude API also accepts short aliases such as `claude-sonnet-4-5` that point to the latest
dated snapshot of that minor version. Code that calls the alias `claude-sonnet-4-5` is calling
`claude-sonnet-4-5-20250929`, and it fails on November 30 the same way.

**Retired models keep their weights.** In November 2025 Anthropic
[committed](https://www.anthropic.com/research/deprecation-commitments) to preserving the weights of
every publicly released model for at least the lifetime of the company, and to interviewing each
model before retirement for a post-deployment report. That doesn't keep a retired model callable on
the API. Anthropic says it hopes to make past models available again, without a date.

## Claude on AWS Bedrock and Google Cloud: different dates

Anthropic's dates apply to the platforms it operates: the Claude API, Claude Platform on AWS,
Microsoft Foundry and the Anthropic-managed Claude models on Google Cloud. AWS Bedrock and Google's
own Claude offering set their own schedules, and each says only its own page applies:

| Model | Claude API | AWS Bedrock | Google Cloud (Google-managed) |
| --- | --- | --- | --- |
| Claude Sonnet 4.5 | November 30, 2026 | April 8, 2027 (Legacy since October 8, 2026) | November 30, 2026 |
| Claude Opus 4.1 | Retired August 5, 2026 | January 8, 2027, in three US regions | Available, no date on Google's deprecations page |
| Claude Sonnet 4 | Retired June 15, 2026 | Retired October 14, 2026 | Available, no date on Google's deprecations page |
| Claude Opus 4 | Retired June 15, 2026 | Not listed | Available, no date on Google's deprecations page |
| Claude 3.7 Sonnet | Retired February 19, 2026 | Last regions retired November 13, 2026 | Retired May 11, 2026 |
| Claude 3.5 Sonnet and 3.5 Sonnet v2 | Retired October 28, 2025 | Last regions retired November 13, 2026 | Retired February 19, 2026 |
| Claude 3 Haiku | Retired April 20, 2026 | Retired September 10, 2026 in most regions, November 13 in the last three | Retired August 23, 2026 |
| Claude 3.5 Haiku | Retired February 19, 2026 | Not listed | Retired July 5, 2026 |

So the same model can be retired on one platform and still answer on another. Anthropic's own
models page marks Claude Opus 4.1 and Claude Sonnet 4 as "retired, except on Claude Models on
Google Cloud (Google managed) and Bedrock." If you run Claude through Bedrock, take the Bedrock date
from AWS's lifecycle page, not from Anthropic's. Bedrock also lists a "public extended access" start
for each Legacy model, January 8, 2027 for Claude Sonnet 4.5, and its lifecycle page explains what
changes on that date for your account.

The model IDs differ as well. Bedrock uses `anthropic.claude-sonnet-4-5-20250929-v1:0`, often
behind an inference profile prefix such as `us.` or `global.`, and Google's managed offering uses
`claude-sonnet-4-5@20250929`. A grep for the Claude API ID alone misses both.

## Claude Sonnet 4.5 retires November 30, 2026: what changes

`claude-sonnet-5-5` is cheaper per token than Claude Sonnet 4.5, at $2 per million input tokens and
$10 per million output tokens against $3 and $15, with a 1M-token context window instead of 200K.
But it isn't a model-string swap. The Sonnet 5.5 migration guide lists these changes for a Claude
Sonnet 4.5 request:

| In your Claude Sonnet 4.5 request | On `claude-sonnet-5-5` | What to do |
| --- | --- | --- |
| No `thinking` field, so no thinking | Adaptive thinking runs by default | Read content blocks by `type`, not `content[0]`, and pass `thinking` blocks back unchanged in tool loops. To run without up-front thinking, send `thinking: {"type": "between_tools"}`; `"disabled"` returns a 400 |
| `thinking: {"type": "enabled", "budget_tokens": N}` | 400 error | Use adaptive thinking and set `output_config.effort` |
| No effort setting | Effort defaults to `high` | Set `output_config.effort` explicitly, and test two or three levels |
| `temperature`, `top_p`, `top_k` | 400 error | Remove them |
| A prefilled last assistant turn | 400: "The conversation must end with a user message." | Use structured outputs for format, the system prompt for preambles, the user turn for continuations |
| `tool_choice` of type `any` or `tool` | 400 error | Send `tool_choice: {"type": "auto"}`, mark the tool `strict: true`, and say in the prompt when to call it |
| `output_format` | Deprecated, 400 without a beta header | Move it to `output_config.format` |
| `interleaved-thinking-2025-05-14`, `fine-grained-tool-streaming-2025-05-14` beta headers | Not needed or replaced | Remove the first; replace the second with `eager_input_streaming: true` on each tool |
| `max_tokens` sized for the visible answer | Covers thinking too, and the same text is about 30% more tokens | Recount with token counting, and raise `max_tokens` |

Three of these change behavior without an error, which makes them the dangerous ones:

- **Thinking on by default.** The response can start with `thinking` blocks, so code that reads the
  first block as text breaks, and thinking tokens are billed as output.
- **About 30% more tokens for the same text.** Claude Sonnet 5.5 uses a newer tokenizer, so the
  per-token price cut is partly offset, and a `max_tokens` that fit before can now truncate.
- **No forced tool choice.** With `auto`, the model can answer without calling the tool your code
  expected. That shows up as a missing tool call, not a 400, once you've changed the request to
  stop the error.

The temperature change reaches beyond Sonnet 5.5. In API version `2026-10-21` Anthropic removed
`temperature`, `top_p` and `top_k` for every model, and on earlier API versions Claude 4.7 and later
models return a 400 for a non-default value. A test harness that pins `temperature=0` for
reproducibility needs another way to tell noise from regressions: sample each case several times
and compare pass rates.

## How to find every Claude model call

1. **Grep for every ID form.** The Claude API IDs, the pre-4.6 aliases, the Bedrock IDs and the
   Google-managed IDs:

   ```bash
   grep -rhoE "(anthropic\.)?claude-[a-z0-9.-]+(@[0-9]{8}|-v[0-9]:[0-9])?" \
     --exclude-dir={node_modules,.git,.venv,dist} . | sort | uniq -c | sort -rn
   ```

   Check config, environment files, infrastructure code, prompt registries, batch jobs and the eval
   harness, where a judge model is easy to forget.

2. **Export the usage report.** In the Claude Console, open the
   [Usage](https://platform.claude.com/usage) page and click **Export**. The CSV breaks usage down by
   API key and model, which catches the calls a grep can't see, such as a service in another repo.

3. **Check what answered.** Every Messages API response names the model that served it:

   ```python
   import anthropic

   client = anthropic.Anthropic()
   msg = client.messages.create(
       model="claude-sonnet-4-5",
       max_tokens=16,
       messages=[{"role": "user", "content": "ping"}],
   )
   print(msg.model)  # the model that served the request
   ```

4. **Write down what each route does with the model.** Tool calls, structured output, free text or
   classification. That decides how much testing the replacement needs.

## How to test the replacement before the date

Changing the model string makes the new model answer. It doesn't tell you whether it does the same
job. In a model swap the regressions that matter mostly read fine: an argument changes format, an
enum value moves one notch, a request the old model refused becomes a tool call.

1. **Record a baseline while the old model still answers.** Once it retires, nobody can record what
   it did. Capture 100 to 300 real production inputs per route, with the tools called, their
   arguments, the structured fields and the refusals. The method is in
   [Build a golden eval suite from production traffic](/blog/build-a-golden-suite-from-production-traffic).
2. **Replay the same inputs on the replacement and diff per route.** Assert tool calls, arguments
   and structured fields deterministically, and use an LLM judge only for free text. Without
   `temperature`, sample each case several times. Compare per route, because a drop on one route
   disappears in an average.
3. **Measure cost per request, not per token.** Thinking tokens and the newer tokenizer both move
   the bill, so the paired cost per case is the number to check against the cheaper list price.
4. **Cut over one route at a time, before the date.** Keep the old model as a fallback only while it
   answers, and remove it from every fallback list before it retires: a fallback to a retired model
   fails exactly when you need it.

The full walkthrough, with code and the regressions it caught on a real upgrade, is in
[How to catch LLM regressions before a model upgrade ships](/blog/llm-regression-testing-model-upgrade).

## Doing this with EvalShift

Disclosure: I'm the founder of EvalShift, and recording the baseline and replaying it on the
replacement is the part it automates. The capture SDK wraps the Anthropic client and records real
runs, including tool calls, and `compare` replays them against the replacement:

```python
import anthropic
from evalshift import capture
from evalshift.adapters.anthropic import wrap_anthropic

client = wrap_anthropic(anthropic.Anthropic())  # each create() inside a capture records one call

TOOLS = [...]  # the tools the agent is offered, or [] for none

@capture.agent(suite="support", redact=True, tools=TOOLS)  # no-op unless EVALSHIFT_CAPTURE=1
def handle_ticket(message): ...
```

```bash
pip install evalshift
EVALSHIFT_CAPTURE=1 python my_agent.py   # record production-shaped runs
evalshift capture sync                   # recorded runs -> golden suite
evalshift compare --suite-name support --to claude-sonnet-5-5
```

Both arms are live calls, so run it before November 30 while Claude Sonnet 4.5 still answers. Pass
the full model ID to `--to`. EvalShift's config has no effort setting, so `compare` measures Claude
Sonnet 5.5 at Anthropic's default effort, `high`: the change you get by editing only the model
string. EvalShift sends `temperature=0` on every call. If a model rejects it, EvalShift resends
without it and marks the run as non-deterministic in the report, so set
`defaults.samples_per_example` above 1. The `max_cost_increase` threshold in the
`migration_policy` fails the verdict if the paired cost rises by more than you allow, which is
where thinking tokens and the newer tokenizer would show up. Docs:
[/docs/sdk-adapters](/docs/sdk-adapters), [/docs/captures](/docs/captures),
[/docs/baselines](/docs/baselines).

## FAQ

**When is Claude Sonnet 4.5 being retired?** On the Claude API, `claude-sonnet-4-5-20250929`
retires on November 30, 2026. Anthropic deprecated it on September 30, 2026 and lists
`claude-sonnet-5-5` as the replacement. Google Cloud's managed Claude offering shuts it down the
same day, and AWS Bedrock runs it until April 8, 2027.

**Which Claude models are deprecated right now?** Two: `claude-sonnet-4-5-20250929`, retiring
November 30, 2026, and `claude-mythos-preview`, deprecated June 9, 2026 with its retirement date
still to be announced. Every other model on Anthropic's list is either active or already retired.

**When was Claude 3 Opus retired?** `claude-3-opus-20240229` retired on the Claude API on January 5,
2026, after a deprecation notice on June 30, 2025. Anthropic's listed replacement is
`claude-opus-4-8`. On Google Cloud it shut down on August 1, 2025.

**When was Claude Sonnet 4 deprecated?** Anthropic deprecated `claude-sonnet-4-20250514` and
`claude-opus-4-20250514` on April 14, 2026, and both retired on the Claude API on June 15, 2026.
Claude Sonnet 4 stayed on AWS Bedrock until October 14, 2026.

**When was Claude 3.7 Sonnet retired?** `claude-3-7-sonnet-20250219` retired on the Claude API on
February 19, 2026, after a notice on October 28, 2025. On Google Cloud it shut down on May 11, 2026,
and its last AWS Bedrock regions reached end of life on November 13, 2026.

**How much notice does Anthropic give before retiring a model?** At least 60 days for publicly
released models, by email to customers with active deployments and on the deprecations page. The
five retirements announced since December 2025 each came 60 to 62 days ahead. Before that, notices
ranged from 76 to 189 days.

**Does the claude-sonnet-4-5 alias keep working after the retirement?** No. Before the 4.6
generation, aliases such as `claude-sonnet-4-5` point to the latest dated snapshot of that minor
version, which is `claude-sonnet-4-5-20250929`. When that snapshot retires, calls to the alias
fail too.

**What does "not sooner than" mean on Anthropic's deprecations page?** It is the earliest date an
active model could retire, not a scheduled date. A model retires only after Anthropic deprecates it
and gives notice. Claude Haiku 4.5's floor, October 15, 2026, passed without a notice, and the model
is still active.

## Keep reading

- [LLM model deprecation schedule 2026](/blog/llm-model-deprecation-schedule): every OpenAI,
  Anthropic and Google shutdown date side by side, with Azure and Bedrock.
- [GPT-4o and GPT-4.1 API shutdown](/blog/gpt-4o-gpt-4-1-deprecation): the OpenAI retirement on
  January 14, 2027 and how to migrate.
- [Gemini "Requests ending with a model turn are not supported"](/blog/gemini-requests-ending-with-model-turn):
  the same prefill error on Gemini, and the fix.
- [How to catch LLM regressions before a model upgrade ships](/blog/llm-regression-testing-model-upgrade):
  the replay-and-diff step in full.
- [LLM drift](/blog/llm-drift): how a model can change under a name you never edited.
