Is an alias the same as a snapshot?

No. A snapshot is one fixed release. An alias points to one, and can be moved to a newer one.

Where do I find a shutdown date?

On the deprecations page of the provider or platform you call. Amazon Bedrock says its dates can differ from the model provider's, and only the dates on its model card apply there.

Home · Foundations · 11 · Model versions

Concept card · Foundations

Model versions

What this gets you. Know which model your product calls, and move before it shuts down.

A model version is one fixed release of a model, named by a model ID. An alias is a shorter name that points to a version, and the provider can move it to a newer one.

Last checked 2026-10-26 6 min

Search for Model versions

What changed

Nov 1: Model facts for a model ID, a three-part switch check, and a replay of five recorded requests before you rename.

Earlier updates

Oct 26: First card. A worked example shows which model names in your code can change, and what a shutdown date means.

See which model names in your code can change without you, and what to do when one gets a shutdown date.

1 What it is

Each request names a model. Some names stay fixed and some can move. A snapshot is one fixed release. Anthropic's versioning page says the model behind a Claude model ID stays the same for that ID's lifetime. An alias is a name that points to a snapshot, and the provider can move it. Gemini's models page says its -latest aliases are hot-swapped with each new release, and that most production apps should use a specific stable model. OpenAI's Deprecations page lists names that point to dated snapshots.

Releases also retire. A provider deprecates a version, usually names a replacement, then shuts it down. After that, requests to it fail. Dated moves you may need to act on are listed on Changes to act on.

Look up a real model ID: Model facts.

Example: made up to show the idea.

Northline Desk, a made-up support product, calls a made-up provider in three places. The provider sends a notice: writer-2-0301 is deprecated. It shuts down in six months. The replacement is writer-3-0901.

Where	Name in the code	Kind	Hit by the notice?
Reply drafts	`writer-2`	Alias. Points to `writer-2-0301` today	Yes. The alias can move to a new release before the date
Ticket labels	`writer-2-0301`	Snapshot	Yes. It stops answering on the shutdown date
Search	`embed-1-0115`	Snapshot	No. Another model, with its own dates

Two of three calls are hit, and only one says so in its name. Pin drafts to a snapshot you have tested. Run your saved tickets on writer-3-0901. Then switch labels in one change, well before the date.

2 Why it matters

Cost

A new version can count the same text as a different number of tokens. Anthropic's models overview says 1M tokens is roughly 555k words on the tokenizer introduced with Claude Opus 4.7. Models before it fit about 750k words. Re-forecast before you switch. This card does not copy a rate.

Model note, last checked 2026-11-01.

Quality

The same prompt can get a different reply on a new version. Anthropic's deprecations page says a new model can reject a deprecated setting. Even a fixed ID can shift a little. Anthropic's versioning page says serving changes, such as safety classifiers, can cause minor differences.

Risk

After the shutdown date, requests fail. Notice varies. OpenAI's Deprecations page gives at least 6 months on generally available models, and previews may get about 2 weeks. Anthropic's page gives at least 60 days on publicly released models. OpenAI's page allows less when safety or compliance requires it.

When a version you call moves

You see	Do this
A shutdown date on a version you call	Find every call to that ID, and plan the move well before the date
A recommended replacement	Run your saved checks on it first: Production evaluation
A new tokenizer or different rates	Re-forecast cost: Cost forecasting
A different model for your users	Read the license that covers it: Licensing
Replies changed with no deploy	Check whether an alias moved: Monitoring & drift

3 How to apply · try today

List every model name your code and config send. Mark each one alias or snapshot, using your provider's models page. Look up each one's shutdown date on the deprecations page of the provider or platform you call. Pin any alias that serves your users to a snapshot you have tested.

Common mistake

Shipping on an alias, then finding that replies changed when it moved. Pin the snapshot you tested, and move on purpose.

Self-check

Your code sends writer-2, an alias for the newest writer-2 snapshot. What can happen with no change on your side?

Self-check

A snapshot you call is deprecated, with a shutdown date in six months. What happens on that date?

Self-check

The notice names a replacement. What do you do before you switch?

Model facts

Last checked 2026-11-01 (Europe/Berlin).

One row per model ID whose context window and max output are stated on the linked page. Rates stay on the maker's page. Each maker's line links there.

OpenAI

OpenAI, Pricing · A prefix of at least 1,024 tokens can be cached. Default minimum lifetime 30 minutes; 12 hours is also offered in the Responses API. OpenAI, Prompt caching

Model ID	Context window	Max output	Shutdown	Source
`gpt-6-astra`	1,050,000 tokens	128,000 tokens	Not listed on the deprecations page. OpenAI, Deprecations	OpenAI, GPT-6 Astra
`gpt-6.1-sol`	1,050,000 tokens	128,000 tokens	Not listed on the deprecations page. OpenAI, Deprecations	OpenAI, GPT-6.1 Sol
`gpt-6-luna`	1,050,000 tokens	128,000 tokens	Not listed on the deprecations page. OpenAI, Deprecations	OpenAI, GPT-6 Luna

Anthropic

Anthropic, Pricing · Opt in with cache_control. Minimum cacheable length 512 tokens. The default lifetime is 5 minutes. A 1-hour duration is also offered. Anthropic, Prompt caching

Model ID	Context window	Max output	Shutdown	Source
`claude-fable-5-1`	1M tokens	128K tokens	Not sooner than September 1, 2027. Anthropic, Model deprecations	Anthropic, Models overview
`claude-opus-5-5`	1M tokens	128K tokens	Not sooner than September 22, 2027. Anthropic, Model deprecations	Anthropic, Models overview
`claude-sonnet-5-5`	1M tokens	128K tokens	Not sooner than September 28, 2027. Anthropic, Model deprecations	Anthropic, Models overview
`claude-haiku-5-5`	1M tokens	128K tokens	Not sooner than October 7, 2027. Anthropic, Model deprecations	Anthropic, Models overview

Google

Google, Gemini Developer API pricing

Model ID	Context window	Max output	Cache rules	Shutdown	Source
`gemini-4-argon`	1,048,576 tokens, shared by input, thinking, and output	Up to 1,048,576 minus the input, with long decode. Without it, streaming or the SDK stops at 64K tokens and non-streaming REST stops at 32K tokens. Gemini, Long decode	Implicit caching is on by default from 4,096 tokens. Explicit caching is not supported. Gemini, Context caching	Not listed on the deprecations page. Gemini, Gemini deprecations	Google, What's new in Gemini 4 Argon
`gemini-3.8-flash`	1M input tokens, the Gemini 3 and 2.5 budget on the long-context page	64K tokens	Implicit caching is on by default from 4,096 tokens. Explicit caching is on generateContent. Gemini, Context caching	No shutdown date announced. Gemini, Gemini deprecations	Gemini, Models · Gemini, Long context
`gemini-3.6-flash`	1M input tokens, the Gemini 3 and 2.5 budget on the long-context page	64K tokens	Implicit caching is on by default from 4,096 tokens. Explicit caching is on generateContent. Gemini, Context caching	No shutdown date announced. Gemini, Gemini deprecations	Gemini, Models · Gemini, Long context
`gemini-3.5-flash-lite`	1M input tokens, the Gemini 3 and 2.5 budget on the long-context page	64K tokens	Implicit caching is on by default for Gemini 2.5 and newer models. Explicit caching is on generateContent. Gemini, Context caching	No shutdown date announced. Gemini, Gemini deprecations	Gemini, Models · Gemini, Long context
`gemini-3.1-flash-lite`	1M input tokens, the Gemini 3 and 2.5 budget on the long-context page	64K tokens	Implicit caching is on by default for Gemini 2.5 and newer models. Explicit caching is on generateContent. Gemini, Context caching	May 7, 2027, the earliest date listed. Gemini, Gemini deprecations	Gemini, Models · Gemini, Long context
`gemini-3.1-pro-preview`	1M input tokens, the Gemini 3 and 2.5 budget on the long-context page	64K tokens	Implicit caching is on by default from 4,096 tokens. Explicit caching is on generateContent. Gemini, Context caching	No shutdown date announced. Gemini, Gemini deprecations	Gemini, Models · Gemini, Long context
`gemini-2.5-flash`	1M input tokens, the Gemini 3 and 2.5 budget on the long-context page	64K tokens	Implicit caching is on by default from 2,048 tokens. Explicit caching is on generateContent. Gemini, Context caching	No shutdown date announced. Gemini, Gemini deprecations	Gemini, Models · Gemini, Long context
`gemini-2.5-pro`	1M input tokens, the Gemini 3 and 2.5 budget on the long-context page	64K tokens	Implicit caching is on by default from 2,048 tokens. Explicit caching is on generateContent. Gemini, Context caching	No shutdown date announced. Gemini, Gemini deprecations	Gemini, Models · Gemini, Long context

Can I switch to X?

X is the model ID you are considering. Compare its row with the row you call now.

Fit. The request you send has to fit X's context window, and the reply you need has to fit its max output. Context Window is how a window fills up.
Token delta. Count one saved request on the model you call now and on X. The same text can come out as more tokens. Cost above quotes Anthropic's word counts for 1M tokens. Open the rate link above each maker's table when you forecast. This card does not copy a rate. Cost forecasting.
Deadline. If the row you call now has a date, move before that date. A cell that says the date is not listed, or that no shutdown date is announced, means that deprecations page did not name a shutdown day on the Last checked day.

Want the why? Understand it · 5 min, optional

Example: made up to show the idea.

Scripted, not a live model. Nothing is sent, and nothing leaves this page.

Replay five recorded requests before you rename.

Northline Desk is a made-up support product. It calls the made-up snapshot desk-writer-2-0301. The replacement is desk-writer-3-0901. Shutdown is in 40 days. The replacement holds 8,000 tokens in the window and 1,000 tokens of output. It refuses a tone parameter. Request 4's instruction block counts as about 2× tokens.

Mark each request Rejected params, Window overflow, Relative cost jump, or Still ok.

1 of 3 · Worked

Five recorded requests, replayed on desk-writer-3-0901 except the last.

Request	Input	Params	Dry run
Request 1	400 tokens	None	Still ok
Request 2	900 tokens	`tone=formal`	Rejected params
Request 3	12,000 tokens	None	Window overflow
Request 4	400 tokens. The instruction block is about 2× tokens	None	Relative cost jump
Request 5	600 tokens	None	Not run yet

Read first: the dry-run column. Which break does each row name?

Verdict: Request 1 is Still ok. Request 2 is Rejected params. Request 3 is Window overflow. Request 4 is Relative cost jump. Request 5 has not been run. The rename waits.

What you change: drop or replace tone on Request 2. Cut Request 3 so it fits the 8,000-token window. Recount Request 4, because its instruction block is about 2× tokens. Run Request 5. Then change the name.

Why: replay the recorded requests before you move. One broken row is enough to hold the rename.

2 of 3 · Your turn

The fixes for Requests 2, 3, and 4 are in, and those three were replayed. Request 1 is Still ok. Request 2, with tone removed, is Still ok. Request 3, cut to 6,000 tokens, is Still ok. Request 4, recounted, is Still ok. Request 5 returned a reply, and the log says "unknown field ignored."

Read first: Request 5. That row does not name a break yet.

What do you do with the rename?

3 of 3 · Your turn

After the fixes, a second sheet says the set looks fine. Request 1 is Still ok. So are Request 2, Request 4, and Request 5. Request 3 is Window overflow: the cut was not saved, so it is still 12,000 tokens into an 8,000-token window.

Read first: Request 3 beside the four Still ok rows.

Is this set enough to cut over?

Check what you learned

Before shutdown, what has to happen before you rename the model in code?

Take five real calls your product sends. For each, write the model name, the input size, the parameters, and what the replacement did. Nothing leaves this page.

Show a checklist

Every call site and model name in code and config is on your list.
Each alias is pinned to a snapshot you have run, or pointed at the replacement on purpose.
At least five recorded requests were replayed on the replacement.
Each request is marked Rejected params, Window overflow, Relative cost jump, or Still ok.
If the replacement is a different model class, such as a new embedding model, stored search text is checked again.
A reminder is set for before shutdown. You name that step on your own calendar. This list has no date.

Five recorded requests, one break mark each, then the rename. A smooth average still leaves a single broken request.

When a request no longer fits the window: Context Window →

When the same text counts as more tokens: Cost forecasting →

Understand it: how providers name, move and retire versions

5 min, optional

The words differ by provider. The routine stays the same: know your names, know the dates, and test before you move.

The same words, different meanings

OpenAI's Deprecations page says a model is deprecated as soon as its retirement is announced. It uses "sunset" and "shut down" for the same thing: no longer accessible. "Legacy" means no more updates. Anthropic uses Active, Legacy, Deprecated and Retired. Amazon Bedrock uses Active, Legacy and End-of-Life (EOL). Gemini's deprecations page says its listed shutdown dates are the earliest possible dates.

How much notice you get

OpenAI gives at least 6 months on generally available models, at least 3 months on specialized variants, and much shorter notice, such as 2 weeks, on previews. Anthropic gives at least 60 days on publicly released models. Gemini's models page says previews are deprecated with at least 2 weeks' notice, and a -latest alias gets 2 weeks' email notice before a breaking change. For models launched under its current policy, Bedrock's Legacy period is 6 months or 45 days. For Azure OpenAI models, Microsoft Foundry gives at least two weeks' notice before a new version becomes the default.

Same model, another platform, other dates

Anthropic's dates apply to the platforms Anthropic operates. Amazon Bedrock and Google-managed Google Cloud set their own. Bedrock's page says only the dates on its model card apply there, and that migration will not happen automatically. Once a Bedrock model is Legacy, new customers can't adopt it, and existing customers may lose access after 15 days of inactivity.

Names that move for you

Gemini's -latest aliases move with each release. In Microsoft Foundry, each Standard deployment has an upgrade policy. Opted out, it stops working when the model retires. Otherwise it moves to the new default version, or when the current one expires. Gemini's deprecations page also lists an old name that keeps answering from a newer model: gemini-3.5-flash routes requests to gemini-3.6-flash.

Model note, last checked 2026-11-01.

Fine-tuned models

OpenAI's page says inference on fine-tuned models is disabled only when the underlying base model is deprecated. Bedrock's page says you can't start new fine-tuning jobs on a model once it is Legacy. Plan a fine-tuned model's move with its base model's dates.

Watch for dates in code

OpenAI's List models endpoint can return a shutdown_date for each model. Bedrock returns a modelLifecycle field from GetFoundationModel and ListFoundationModels. Anthropic's Console can export usage by API key and model, so you can find calls to a deprecated model.

Check yourself

Your product calls one alias and one snapshot. Name one risk for each.

Show the answer

The alias can move to a new release with no deploy, so replies can change. The snapshot stays the same, but stops answering on its shutdown date.

Questions

What is a model version?: A model version is one fixed release of a model, named by a model ID. An alias is a shorter name that points to a version, and the provider can move it to a newer one.
Is an alias the same as a snapshot?: No. A snapshot is one fixed release. An alias points to one, and can be moved to a newer one.
Where do I find a shutdown date?: On the deprecations page of the provider or platform you call. Amazon Bedrock says its dates can differ from the model provider's, and only the dates on its model card apply there.

Next card →