When is this Claude or ChatGPT model retired, and what carries over?
Claude Sonnet 4.5’s retirement date on the Claude API is November 30, 2026. As of November 1, 2026, Claude Opus 4.5 had no retirement date, only a “not sooner than” floor of November 24, 2026. On the OpenAI API, the shutdown date for the GPT-5 family and o3 is December 11, 2026, and for the GPT-4o and GPT-4.1 snapshots January 14, 2027. GPT-4o left ChatGPT on February 13, 2026.
In the cases the vendors have written up (Haiku 4.5 at Anthropic, GPT-4o and others at OpenAI), old chats stayed in the account and continued on a newer model.
Memory and custom instructions are account settings, not part of a model, so they should stay. Anthropic says memory helps a new model adapt; OpenAI’s retirement notice doesn’t mention them. Saving your memory as text (below) takes a minute. The model’s way of writing doesn’t carry over to anything.
Dates as of November 1, 2026, read that day from Anthropic’s and OpenAI’s own pages (listed at the foot). OpenAI’s three ChatGPT retirement pages blocked our download, so they were read from archived copies made between September 22 and October 30, 2026. An AI drafted this page from those sources and another AI checked it. No person has run the steps; they come from the vendors’ help pages and, for Counterparts, our own docs. We make Counterparts, a memory for Claude, which comes up near the end.
The dates
Each vendor runs two clocks. The API date is when requests to the model stop working. The app date is when it leaves the model picker in claude.ai or ChatGPT. A model can leave the app months before its API date, and an API date says nothing about the app.
Claude
Dates for the API apply to the Claude API, Claude Platform on AWS, Claude Models on Google Cloud (Anthropic managed) and Microsoft Foundry. Amazon Bedrock and Google-managed Claude set their own dates.
| Model | Claude API | claude.ai and the Claude apps | Status, Nov 1, 2026 |
|---|---|---|---|
| Sonnet 4.5 | Retirement date: November 30, 2026 (notice given September 30, 2026). Replacement: Sonnet 5.5 | A help-center note (archived May 17, 2026) said it would leave the model selector on May 15, 2026. The note is gone now, and we found no confirmation of the removal. If it isn’t under “More models”, it’s gone from the app | Deprecated, dated |
| Opus 4.5 | Not sooner than November 24, 2026. No notice | No notice found | Active, floor only |
| Haiku 4.5 | Not sooner than October 15, 2026. That floor has passed and it’s still Active | Removed from the selector on October 29, 2026 for Free, Pro and Max (announced October 15). Still in Claude Code and on Team and Enterprise | Active on the API, gone from the consumer picker |
| Opus 4.6 | Not sooner than February 5, 2027 | No notice found | Active, floor only |
| Sonnet 4.6 | Not sooner than February 17, 2027 | No notice found | Active, floor only |
| Opus 4.7 | Not sooner than April 16, 2027 | No notice found | Active, floor only |
| Opus 4.8 | Not sooner than May 28, 2027 | No notice found | Active, floor only |
| Fable 5, Mythos 5, Sonnet 5, Opus 5 | Not sooner than June 9, 2027 (Fable 5, Mythos 5), June 30, 2027 (Sonnet 5), July 24, 2027 (Opus 5) | No notice found | Active, floor only |
| Fable 5.1, Mythos 5.1, Opus 5.5, Sonnet 5.5, Haiku 5.5 | Not sooner than September 1, 2027 (Fable 5.1, Mythos 5.1), September 22 (Opus 5.5), September 28 (Sonnet 5.5), October 7, 2027 (Haiku 5.5) | No notice found | Active, floor only |
| Mythos Preview | Deprecated June 9, 2026; retirement date “to be announced” | No notice found | Deprecated, no date |
| Opus 4.1 | Retired August 5, 2026 | Removed from the selector January 16, 2026 | Retired |
| Opus 4, Sonnet 4 | Retired June 15, 2026 | Opus 4 removed from the selector January 16, 2026 | Retired |
| Haiku 3 | Retired April 20, 2026 | No note found | Retired |
| Haiku 3.5, Sonnet 3.7 | Retired February 19, 2026 | No note found | Retired |
| Opus 3 | Retired January 5, 2026; available on the API by request | Kept for paid claude.ai subscribers after retirement (Anthropic, February 2026) | Retired; still reachable as of Anthropic’s February 2026 post |
| Sonnet 3.5 | Retired October 28, 2025 | No note found | Retired |
On Amazon Bedrock, Sonnet 4.5 has been Legacy since October 8, 2026, with an end-of-life date of April 8, 2027. New Bedrock customers can’t use a Legacy model, existing ones may lose access after 15 days without using it, and from January 8, 2027 AWS says to expect higher prices.
OpenAI
| Model | OpenAI API | ChatGPT | Status, Nov 1, 2026 |
|---|---|---|---|
| GPT-5, GPT-5 mini, GPT-5 nano, GPT-5 pro, o3, o3-pro | API shutdown date: December 11, 2026 (notice June 11, 2026) | GPT-5 Instant and Thinking retired February 13, 2026. OpenAI said on May 28, 2026 it would retire o3 from ChatGPT on August 26, 2026 (90-day sunset); we found no later note confirming it | Deprecated on the API, dated |
| GPT-4o (2024-08-06 and 2024-11-20 snapshots), GPT-4o mini, GPT-4.1, GPT-4.1 mini | API shutdown date: January 14, 2027 (notice October 14, 2026). Fine-tunes follow the same date | GPT-4o, GPT-4.1 and GPT-4.1 mini retired February 13, 2026. GPT-4o in Custom GPTs on Business, Enterprise and Edu ended April 3, 2026 | Gone from ChatGPT; deprecated on the API, dated |
| GPT-4o (2024-05-13 snapshot), o4-mini, o1, o3-mini, GPT-4 Turbo | Shut down October 23, 2026 | o4-mini retired from ChatGPT February 13, 2026 | Shut down |
| GPT-5.1 | Shuts down April 1, 2027 (notice October 1, 2026) | Retired March 11, 2026 | Gone from ChatGPT; deprecated on the API, dated |
| GPT-4.5 | The API preview ( | OpenAI said on May 28, 2026 it would retire GPT-4.5 from ChatGPT on June 27, 2026 (30-day sunset); we found no later note confirming it | Retirement announced for ChatGPT |
| GPT-5.5 | No notice on the API deprecations page | OpenAI said on October 14, 2026 it is retiring GPT-5.5 from ChatGPT, ChatGPT Work and Codex. The note gives no date and says API access isn’t affected | Leaving ChatGPT, no date given |
OpenAI’s deprecations page lists no floor dates for its current models. Its API reports a shutdown_date on each model once one is announced.
What the words mean
Deprecated, at Anthropic: the model still works and usually has a retirement date (Mythos Preview’s is still “to be announced”). Retired: “Requests to retired models fail.”
Deprecated, at OpenAI’s API: a model becomes deprecated the day the shutdown is announced, and every deprecated model has a shutdown date. “Sunset” and “shut down” both mean it no longer answers.
Retired, in ChatGPT: gone from the ChatGPT picker. OpenAI’s ChatGPT notices often add that “there are no changes to the API”.
Removed from the model selector, in claude.ai: the same idea at Anthropic. Haiku 4.5 is the current case: out of the consumer picker, still Active on the API.
“Not sooner than”: a floor, not an announcement. Anthropic gives at least 60 days’ notice before a retirement. Haiku 4.5’s floor of October 15, 2026 passed with no notice, and the model is still Active.
Legacy: no more updates, and likely deprecated later. Anthropic and OpenAI both use the word this way. Anthropic’s models overview lists every model before the current generation, Opus 5 and Fable 5 included, under “Legacy models (still available)” while its deprecations table shows them as Active. On Bedrock, Legacy also means new customers can’t start using the model.
How much notice to expect
We counted these gaps from the vendors’ own tables.
Anthropic API: the last five notices gave 60 to 62 days (Sonnet 4.5: September 30 to November 30, 2026). Older ones gave more: Sonnet 3.7 got 114 days and Opus 3 got 189.
OpenAI API: six months for the GPT-5 family and o3, and for GPT-5.1. Three months for the GPT-4o and GPT-4.1 snapshots, shorter than the six months OpenAI states for generally available models (our reading; the notice doesn’t explain it).
The apps: Haiku 4.5 got 14 days’ notice before leaving claude.ai’s selector. GPT-4o got 15 days before leaving ChatGPT. OpenAI gave o3 a 90-day sunset in ChatGPT and GPT-4.5 a 30-day one.
Anthropic adds that “historically Anthropic has deprecated most models around a year after their initial release”.
Is Opus 4.5 next? Its floor is November 24, 2026, and it had no notice on November 1. A notice issued that day would put the earliest retirement at December 31, 2026. That’s our arithmetic from the 60-day rule. Check the deprecations page for a notice since.
What carries over when a model retires
Blank ChatGPT cells mean OpenAI’s notices don’t cover the row.
| What you have | Claude | ChatGPT |
|---|---|---|
| Old chats | For Haiku 4.5, Anthropic said the chats “are still in your account, and nothing was deleted”, and continuing one gets a reply from Haiku 5.5. That notice is about Haiku 4.5. We searched Anthropic’s help center on November 1 for Sonnet 4.5, Opus 4.5 and retired models and found no matching note for any other model | “Conversations and projects are defaulted to their GPT-5.3 Instant and GPT-5.4 Thinking and Pro equivalents going forward.” GPT-5.1 chats “automatically continue on the corresponding current model” |
| Projects | Anthropic suggests projects with custom instructions to help a new model adapt. Its project help page doesn’t mention a model setting, and no retirement notice we read says how a project moves | Projects move to the newer equivalents, as above. GPTs on a retired model “automatically move to the closest” equivalent |
| Memory and instructions | Account settings, not tied to one model, so they should stay (our reading). Anthropic says memory and chat search “can help new models to learn and immediately adapt to your preferences and communication style”. Claude’s memory “is designed to focus on work-related topics”, so it “may not retain imported personal details unrelated to work” | OpenAI’s retirement article doesn’t mention them |
| API calls to a pinned model ID | Fail after the retirement date. Every Claude model ID is a pinned snapshot, so nothing moves you to the replacement on its own | The model “will no longer be accessible” after the shutdown date. Fine-tunes stop with their base model |
| Claude Code sessions | Claude Code shows a startup warning when your chosen model has a scheduled retirement date. Resuming a session whose model has retired falls back to your usual model settings | |
| Your exported data | An archive you keep. Anthropic says an export can’t be imported into another personal Claude account | |
| How the model wrote | Doesn’t carry over. Anthropic says its suggestions “can only go so far toward preserving or replicating the unique experience of interacting with a particular model” | Doesn’t carry over |
What to do before the date
Try the replacement now, side by side, with your own prompts. Anthropic recommends this as soon as a deprecation is announced, while you can still compare.
Export your chats. In claude.ai on the web or Claude Desktop: Settings > Privacy > Export data. You can’t start it from the phone apps. The emailed link expires after 24 hours. On Team and Enterprise an owner has to allow it, and the export leaves out the contents of files you uploaded.
Save Claude’s memory as text. View it under Settings, or ask Claude: “Write out your memories of me verbatim, exactly as they appear in your memory.” Keep the result in a file.
Write down what you liked. Anthropic’s suggestion: ask Claude to review the conversations you most appreciated and write instructions for itself from them. Put the result in a project or your account’s instructions. A new model will follow it in its own way.
On the API: export usage by key and model from Usage in Claude Console to find what still calls the old model, then change the model ID. Anthropic’s migration guides list what else changes. For example,
temperature,top_pandtop_kreturn an error on API version2026-10-21. On OpenAI, check each model’sshutdown_date.
These steps make copies; none of them keeps the old model.
Keeping what you’ve built across models
What you and Claude have written down can move to the next model; the model itself can’t.
Counterparts, which we make (Mike LaPeter runs it), is a memory kept outside any one model, and we think it’s the best answer to this problem.
The AI writes down what it learned about you, your work and itself, in its own words. What matters sticks; what never comes up again fades, the way a person’s memory does. Each session starts with a short note on who you are and where you left off.
Any Claude model you use with it reads that note, so what you wrote down with Sonnet 4.5 is there for Sonnet 5.5 or Opus 5.5. Your preferences, the decisions you made together and the projects in progress carry over. The local install is free and open source.
Claude’s built-in memory helps too, and Anthropic recommends it. It’s built around work, and Anthropic says it “may not retain imported personal details unrelated to work”. Counterparts keeps one memory of the whole relationship, including what the AI learned about itself.
What memory can’t do:
It doesn’t keep the old model, its weights or its voice. The new model reads notes it didn’t write and answers in its own way.
Counterparts doesn’t import your chat exports.
Claude Code (Mac or Linux): in Claude Code, run
/plugin marketplace add mlapeter/counterparts
/plugin install counterparts@counterpartsor in a terminal, npm install -g counterparts@latest and then counterparts (needs Node 22.15 or later; or bun add -g counterparts with bun 1.3+). Give your name, restart Claude Code, and run counterparts doctor. Memories stay in one file on your machine. Setup for Claude Code.
Claude Desktop on a Mac: install as above, then quit Claude Desktop, run counterparts install --host claude-desktop, open Desktop, and set the Counterparts tools to Always allow. In a new chat, pick “Start with Counterparts” or ask it to call the counterparts wake tool.
claude.ai: a hosted version connects through Settings > Connectors. It’s in beta and needs an invite for now; to ask for one, email mike@resonantventures.com. It stores memories on our server.
ChatGPT: Counterparts has no door there. ChatGPT’s own memory and custom instructions are account settings and aren’t tied to one model (our reading; OpenAI’s notice doesn’t say), and an export keeps a copy. If you move to Claude, Claude’s memory import takes a pasted list of what another assistant knows about you.
Questions people ask
Can I still use Sonnet 4.5 after November 30, 2026?
Not on the Claude API, Claude Platform on AWS, Anthropic-managed Google Cloud or Microsoft Foundry: requests fail after that date. On Amazon Bedrock, existing customers can keep using it until April 8, 2027, with the catches above. In claude.ai, a help-center note archived May 17, 2026 said it would leave the model selector on May 15; we found no confirmation of the removal.
How do I downgrade Claude to an older model?
In claude.ai, click the model name next to the send button and open “More models”; older models appear there while Anthropic still offers them in the app. In Claude Code, run /model with the full ID, such as /model claude-sonnet-4-6, or launch with claude --model claude-sonnet-4-6. On the API, send the older model ID until its retirement date. Once a model retires on the Claude API, requests to it fail there. Exceptions exist: Amazon Bedrock and Google-managed Claude keep their own, often later, dates; Anthropic kept Opus 3 for paid claude.ai users and offers it on the API by request; OpenAI says some API customers can arrange dedicated capacity past a shutdown date through sales.
Will my chats be deleted when a model retires?
Anthropic said none were deleted when Haiku 4.5 left the selector: the chats stayed and continued on Haiku 5.5. OpenAI says conversations with retired ChatGPT models continue on the current equivalents. Exporting still gives you a copy that doesn’t depend on either.
When is GPT-4o retired?
It left ChatGPT on February 13, 2026, and Custom GPTs on Business, Enterprise and Edu plans on April 3, 2026. On the API, the 2024-05-13 snapshot shut down October 23, 2026, and the 2024-08-06 and 2024-11-20 snapshots shut down January 14, 2027, with gpt-6-luna as the named replacement.
What does “not sooner than” mean on Anthropic’s page?
The earliest the model could retire, not a plan to retire it then. A real date comes with a notice at least 60 days ahead. Haiku 4.5’s floor passed on October 15, 2026 and it is still Active on the API.
What time of day does a model retire?
Neither vendor’s deprecations page gives a time, only a date. Plan to finish before that day starts.
How did Opus 3 stay on claude.ai, and could my model come back?
Anthropic retired Opus 3 on January 5, 2026 and kept it for paid claude.ai subscribers, with API access by request. It wrote that it is “not committing to similar actions for every model in the future”. Anthropic has committed to keep the weights of its publicly released models for at least the life of the company and says it hopes to make past models available again someday. It hasn’t promised that for any model.
Questions? Ask in r/AIMemorySystems.
Last checked November 1, 2026. Dates change with every notice; the deprecations pages below are the record.
Sources
- Anthropic: Model deprecations: https://platform.claude.com/docs/en/about-claude/model-deprecations
- Anthropic: Models overview: https://platform.claude.com/docs/en/about-claude/models/overview
- Anthropic: Claude Haiku 4.5 was removed from the model selector: https://support.claude.com/en/articles/17444146-claude-haiku-4-5-is-leaving-the-model-selector
- Anthropic: Adapt to new model personas after deprecations: https://support.claude.com/en/articles/12738598-adapt-to-new-model-personas-after-deprecations
- Anthropic: Release notes: https://support.claude.com/en/articles/12138966-release-notes
- Anthropic: How up-to-date is Claude’s training data? (archived May 17, 2026, with the Sonnet 4.5 selector note): https://web.archive.org/web/20260517083338/https://support.claude.com/en/articles/8114494-how-up-to-date-is-claude-s-training-data
- Anthropic: Export your Claude data: https://support.claude.com/en/articles/9450526-export-your-claude-data
- Anthropic: Import and export your memory from Claude: https://support.claude.com/en/articles/12123587-import-and-export-your-memory-from-claude
- Anthropic: Change the model, effort and thinking settings: https://support.claude.com/en/articles/8664678-change-the-model-effort-and-thinking-settings
- Claude Code docs: Model configuration: https://code.claude.com/docs/en/model-config
- Anthropic: Commitments on model deprecation and preservation: https://www.anthropic.com/research/deprecation-commitments
- Anthropic: An update on our model deprecation commitments for Claude Opus 3: https://www.anthropic.com/research/deprecation-updates-opus-3
- OpenAI: Deprecations: https://developers.openai.com/api/docs/deprecations
- OpenAI: Retiring GPT-4o and other ChatGPT models: https://help.openai.com/en/articles/20001051-retiring-gpt-4o-and-other-chatgpt-models
- OpenAI: Retiring GPT-4o, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini in ChatGPT: https://openai.com/index/retiring-gpt-4o-and-older-models/
- OpenAI: Model release notes: https://help.openai.com/en/articles/9624314-model-release-notes
- AWS: Amazon Bedrock legacy models: https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle-legacy.html
- Counterparts on GitHub: https://github.com/mlapeter/counterparts