Your AI Stack Has an Expiration Date: Model Sunsets Are the New Patch Tuesday

On August 5, 2026, Anthropic retired Claude Opus 4.1 on its API, and requests to claude-opus-4-1-20250805 started to fail. On Amazon Bedrock, the same model stays callable until January 8, 2027. AI models now retire on published schedules that differ by provider and by platform, so every model ID in your code needs a date on a calendar, just like a Patch Tuesday.

Verified as of 25 September 2026 against the sources linked inline.

The pattern is older than AI. Identity infrastructure, which I have built since founding LoginRadius in 2013 and scaling it past a billion identities, depends on upstream API versions and their shutdown dates. A missed upstream deprecation does not look like a code bug. It looks like an outage.

What is happening: model retirement is now a recurring event

Every major AI provider now runs a model lifecycle. A model starts as active, moves to legacy or deprecated when a newer one ships, and is finally retired. After retirement, the model ID stops serving requests, or it quietly serves a different model. The providers publish these dates on their own deprecation pages, and the dates keep coming.

Anthropic: seven Claude models retired in 2026

Anthropic's model deprecations page lists seven Claude models retired so far in 2026. The run started with Claude Opus 3 on January 5 and ended, for now, with Claude Opus 4.1 on August 5. Anthropic notified Opus 4.1 users on June 5 and named claude-opus-4-8 as the replacement. The page states plainly that requests to retired models will fail.

Anthropic commits to at least 60 days of notice for publicly released models. The same page lists "not sooner than" dates for models that are still active. claude-sonnet-4-5-20250929 carries a floor of September 29, 2026, and claude-haiku-4-5-20251001 a floor of October 15, 2026. A floor is the earliest possible retirement, not an announced one.

OpenAI: a six-month promise, with exceptions

OpenAI's deprecations page promises at least six months of notice for generally available models. On June 11, 2026 it announced that gpt-5-2025-08-07 shuts down on December 11, 2026, with gpt-5.6-sol as the replacement. Five sibling snapshots, including o3-2025-04-16 and gpt-5-mini-2025-08-07, shut down the same day.

An earlier batch lands first. On April 22, 2026 OpenAI announced that legacy snapshots such as gpt-4-0613, gpt-4-turbo and o1-2024-12-17 shut down on October 23, 2026. The page also lists gpt-5.4-cyber, announced September 11 with an October 1 shutdown. That is 20 days, because the six-month promise covers generally available models only.

xAI and DeepSeek: the old name kept answering

xAI retired eight model slugs at noon Pacific on May 15, 2026, per its retirement guide. Requests to the retired reasoning slugs are now served by grok-4.3 with low reasoning effort. Non-reasoning slugs are served by grok-4.3 with no reasoning. Old slugs are billed at $1.25 per million input tokens and $2.50 per million output tokens.

DeepSeek's API changelog announced on April 24, 2026 that deepseek-chat and deepseek-reasoner would be discontinued on July 24, 2026. During those three months, the two names routed to the non-thinking and thinking modes of deepseek-v4-flash. Callers were talking to a new model before the old name ever stopped working.

Google: dates are the earliest possible

Google's Gemini API deprecations page says its listed shutdown dates are "the earliest possible dates" a model might be retired. gemini-2.0-flash carried a June 1, 2026 shutdown date with gemini-3.6-flash as the replacement. Google also now restricts Gemini 2.5 access to users who have used those models before, and steers new projects to newer models.

Why it matters to engineering and security teams

A retirement looks like a one-line change. Engineering swaps a model string, the tests pass, and the ticket closes. The real exposure sits in five places that a string change does not cover.

Hard failures you can see

When a model retires on the Claude API, requests to it fail. That is the safe outcome, because monitoring catches it. The risk is timing: a forgotten batch job or internal tool breaks on retirement day, and nobody owns it. Parameters can break too. Anthropic's page lists temperature, top_p and top_k as deprecated for Claude Opus 4.7 and later, and a non-default value returns a 400 error.

Behaviour changes you cannot see

The xAI and DeepSeek cases are harder. Your application keeps getting successful responses, but a different model now writes them. Refusal behaviour, output format, tool-call reliability and prompt-injection resistance are properties of the model, not your prompt. Your last security evaluation described a model you no longer run.

Data terms that change with the model

A replacement model does not always inherit the data terms of the old one. Anthropic's API and data retention page designates Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5 and Claude Mythos 5 as Covered Models. They require 30-day data retention, and zero data retention (ZDR, where the provider stores no prompts or outputs) is not available for them unless Anthropic expressly authorizes it.

On the Claude API, a request to Claude Fable 5 from an organization whose retention setup does not meet that rule returns a 400 invalid_request_error. An organization on ZDR can enable 30-day retention for one workspace to use these models. That is a data-flow change your privacy team has to approve before engineering ships it. My post on what zero data retention actually covers explains the contract side.

The same model on different clocks

Anthropic's dates apply to Anthropic-operated platforms, and its page says Amazon Bedrock and Google Cloud set their own schedules. Bedrock's legacy lifecycle table shows Claude Sonnet 4 reaching end of life on October 14, 2026, four months after it retired on the Claude API. Claude Opus 4.1 reaches Bedrock end of life on January 8, 2027.

Bedrock adds two traps. From October 8, 2026, Opus 4.1 enters "public extended access," where Bedrock says pricing may rise as set by the provider. And once a model is Legacy, existing customers "may lose access after 15 days of inactivity." A disaster-recovery account that rarely calls the model can lose it before production does.

Compliance records that go stale

Records of processing, vendor risk assessments and AI system inventories name specific models and specific data terms. After a migration they describe a system you no longer run. For teams pursuing ISO/IEC 42001, that mismatch is an audit finding, a point I covered in what ISO 42001 actually certifies. A model migration is a production change and needs the same evidence as any other.

What the AI Models reference shows

AI Models is a facts-only reference for the major language models. It publishes no rankings. Each model card records the exact API model IDs, the lifecycle status, the announced retirement date and replacement, and data-handling facts such as training default, retention and ZDR. It currently tracks 66 models.

How it is built

The methodology page sets three rules. Every fact links to a page the provider itself publishes, never to news coverage or aggregators. Every fact carries the date it was last checked. And stale facts fail the build: lifecycle status must be re-verified within 60 days and governance facts within 120, or the site does not deploy.

That design has a limit you should respect. The reference is a dated snapshot, and a provider can change its page the day after a check. Every model card shows its verification date and links to the provider's source. Treat the reference as your index and the provider's page as the final word before you schedule work.

Three findings from the data

• More than a quarter of tracked models are already gone. The lifecycle page lists 19 of the 66 models as retired, and 15 of those retirements happened in 2026. Claude Opus 4.1, Grok Code Fast 1 and DeepSeek Chat all sit in that group.
• Retired names do not all fail the same way. The Grok Code Fast 1 card records a replacement of Grok Build 0.1, a different model built for agentic coding. The DeepSeek Chat card records deepseek-v4-flash. In both cases the old name kept answering for a while.
• Data terms differ by model, not only by provider. On the data-handling comparison, the Claude Fable 5 card shows ZDR as unavailable, while other current Claude models show it on request. A same-provider upgrade can still cross a data boundary.

How to use it: build a model-retirement calendar

A model-retirement calendar is a shared list of every model ID you run, its next retirement date, its replacement and an owner. Building one takes an afternoon. Keeping it current takes a monthly check.

1. List the model IDs you run. Search repositories, infrastructure config, environment variables and automation tools for model strings. Anthropic's Console also exports usage by API key and model, which catches callers your code search misses.
2. Match each ID to its model card. Cards list every ID string a model answers to. The GPT-5 card, for example, lists both gpt-5 and gpt-5-2025-08-07. The glossary explains the difference between an alias and a dated snapshot.
3. Read the lifecycle page by status. The lifecycle page groups models as retired, deprecated, legacy and active, with the retirement date, replacement and verification date on each row. Anything you run that is not active gets a ticket.
4. Confirm on the provider's page. Follow the source link on the card and check the date and status today. Record whether the date is firm or a "not sooner than" floor, since the two need different reminders.
5. Add platform dates as separate rows. If you call a model through Bedrock or Google Cloud, look up that platform's date as well. The Claude Sonnet 4 gap shows why one row per model is not enough.
6. Check the replacement's data terms. Before you pick a replacement, compare its row on the data-handling page with the model it replaces. Flag any change in training default, retention, ZDR or residency for privacy review.
7. Set two reminders and one owner. Put a 90-day and a 30-day reminder on each date, and assign an owner by role so the task survives turnover.

Here is what the next few months look like on such a calendar, using only dates from the providers' own pages.

Date	Model ID	Platform	Type of date	Named replacement
September 29, 2026	claude-sonnet-4-5-20250929	Claude API	Floor (active today)	None named yet
October 1, 2026	gpt-5.4-cyber	OpenAI API	Shutdown	gpt-5.6-cyber
October 14, 2026	anthropic.claude-sonnet-4-20250514-v1:0	Amazon Bedrock	End of life	Not listed
October 15, 2026	claude-haiku-4-5-20251001	Claude API	Floor (active today)	None named yet
October 23, 2026	gpt-4-0613, o1-2024-12-17 and others	OpenAI API	Shutdown	gpt-5.6-sol for most
November 24, 2026	claude-opus-4-5-20251101	Claude API	Floor (active today)	None named yet
December 11, 2026	gpt-5-2025-08-07 and five siblings	OpenAI API	Shutdown	gpt-5.6-sol, -terra, -luna
January 8, 2027	anthropic.claude-opus-4-1-20250805-v1:0	Amazon Bedrock	End of life	Not listed

If you already track tools and vendors in a register, add these rows there. The tool inventory template on The CISO Desk has room for model dependencies next to your other vendors.

What to do next

• Run a model-ID inventory this week, and compare it with each provider's usage export.
• Pin dated snapshot IDs in production, and keep aliases for development only.
• Log the model ID each response reports, and alert when it differs from the one you pinned.
• Route every provider's deprecation email to a monitored group mailbox, not one engineer's inbox.
• Re-run your full evaluation suite, including safety and prompt-injection cases, on every replacement.
• Get privacy sign-off on any change in retention, ZDR, training default or data location.
• Migrate well before the date, so the old model is still there if you need to roll back.
• Update your records of processing, AI system inventory and DPA schedule, then close the change ticket.

Frequently Asked Questions

What happens when an AI model is retired?

It depends on the provider. Anthropic says requests to a retired model fail, and Amazon Bedrock says the same after a model's end-of-life date. xAI instead routed its eight retired slugs to grok-4.3 and billed them at its rates. DeepSeek routed its two old names to deepseek-v4-flash for three months before discontinuing them.

How much notice do AI providers give before retiring a model?

Anthropic commits to at least 60 days for publicly released models, and OpenAI to at least six months for generally available models. Exceptions exist: OpenAI gave 20 days for gpt-5.4-cyber, which was not covered by that promise. Google lists earliest possible dates and confirms the real date later. Plan for the shortest notice you have seen.

Does a model retire on the same date everywhere?

No. Anthropic's dates cover its own platforms, and Amazon Bedrock and Google Cloud set their own. Claude Sonnet 4 retired on the Claude API on June 15, 2026, but reaches Bedrock end of life on October 14, 2026. Track each platform as a separate row on your calendar.

Should I use a model alias or a pinned version?

Pin a dated snapshot in production and use aliases only in development. An alias can point to a different model next month without any change to your code. The retired xAI and DeepSeek names show that even a retired ID can quietly answer with a new model.

Does zero data retention carry over when I switch models?

Not automatically. ZDR is scoped by model, endpoint and feature. Anthropic's Covered Models, including Claude Fable 5, require 30-day retention and are not available under ZDR unless Anthropic expressly authorizes it. Check the replacement's data terms before you migrate, not after.