Dashboard

AI Model Preview: What the Label Promises

An AI model preview label is not a lifecycle stage. Compare the documented notice periods and find out what you are actually promised before you ship.

Cecilia Iona
Cecilia Iona
Senior Editor, AI & Product
10 October 20261 min read

An AI model preview label tells you less than it appears to, because preview is usually not a stage in the vendor's formal model lifecycle at all. It is a stability signal bolted on beside the lifecycle, and what it promises you differs sharply between providers. One documents two weeks of notice before a preview model goes away. Another gives 60 days, but only for models it describes as publicly released.

If you are deciding whether to ship on a preview model, that gap is the whole question.

The formal lifecycle, and where ai model preview sits

Anthropic's model deprecations page defines four states, and preview is not among them:

Stage

Definition as published

What it means for you

Active

Fully supported and recommended for use

Build on it

Legacy

No longer receives updates and may be deprecated in future

Working, on notice

Deprecated

Still functional but no longer recommended, with a replacement and a retirement date

Migrate now

Retired

No longer available; requests fail

Already broken

Notice how cleanly those four describe a model's support status and how little they say about stability. A model can be Active and still be labelled preview, which is why reading the stage alone does not answer the question you care about.

The documented commitment is "at least 60 days' notice before model retirement for publicly released models", and that qualifier does real work. A preview is arguably not a public release, so the 60 days is not obviously yours.

What the label is worth in practice

A current example from that same table makes the point better than any argument. One preview model listed there is marked Deprecated, deprecated on 9 June 2026, with its retirement date given as "to be announced". Every generally available model in the table instead carries a concrete retirement date, usually a year or more out, phrased as "not sooner than".

So the practical difference is not that preview models are unstable in some vague way. It is that a preview model can enter deprecation without a date attached, while a generally available model arrives with a floor under it from the start.

Google's version patterns, which name a number

Google's Gemini documentation splits the same territory three ways and is unusually explicit about the trade. Its models page says preview models "may be used for production", typically have billing enabled, "might come with more restrictive rate limits", and "will be deprecated with at least 2 weeks notice". Experimental models are described as "typically not suitable for production use", not stable, with endpoint availability subject to change. Stable models "usually don't change", and the page recommends most production apps pin a specific stable model.

Two weeks is a number you can plan against, which already puts it ahead of a vaguer assurance. It is also short. Two weeks of notice means a migration you can execute if someone is watching the changelog, and a broken product if nobody is.

Preview

Stable or generally available

Production use

Permitted, with caveats

Recommended

Documented notice before removal

2 weeks (Google); not clearly covered by the 60-day commitment (Anthropic)

60 days for publicly released models (Anthropic)

Rate limits

May be more restrictive

Standard tier limits

Behaviour changes

Expected

Usually none

When shipping on a preview model is the right call

It often is, and the risk is manageable once you name it. Three conditions make it reasonable.

  • The capability does not exist in a stable model. If preview is the only way to do the thing, the comparison is not preview versus stable, it is preview versus not shipping.

  • You can switch models behind a configuration change. If the model identifier is a constant in one place and your prompts are not tuned to one model's quirks, a two-week migration is an afternoon.

  • Somebody reads the changelog. This is the one that fails silently. A two-week notice period is only a notice period if it reaches a human.

Against that, the strongest reason to wait is tuning. Prompts tightened against a preview model's particular behaviour are the expensive part of a migration, not the identifier. The more you optimise, the more a change costs you, which is a good argument for keeping prompts plainer while on preview.

Three habits that cost nothing

  1. Pin the exact identifier, never an alias. A dated or fully qualified name changes when you change it. An alias changes when the vendor decides.

  2. Keep a second model configured and tested, even if you never use it. Lock-in is cheap to avoid in advance and expensive to escape under a deadline, which is the argument in how to avoid lock-in to one AI provider.

  3. Check the platform, not just the model. Anthropic notes that partner-operated platforms such as Amazon Bedrock and Google Cloud set their own retirement schedules, so the same model can carry a different status and different dates depending on where you call it. Our note on the same model behaving differently across providers covers the behavioural half of the same problem.

When a retirement notice does land, the sequence to follow is in what to do when an AI model gets deprecated, and the broader mechanics of what these labels describe sit in how AI models work.

Frequently asked questions

Is a preview model cheaper?

Not reliably. Google's documentation notes preview models typically have billing enabled, so preview is not a free tier. Any price difference is a pricing decision rather than a property of the label.

Can a preview model change its answers overnight?

Yes, and that is the main practical difference from a stable model, which vendors describe as usually not changing. If identical inputs must give identical outputs, pin a stable model and keep your own regression set.

What is the difference between preview and experimental?

Where a vendor distinguishes them, experimental is the weaker promise: not suitable for production, not stable, and endpoint availability subject to change. Preview sits between experimental and stable.

Does preview mean the model is worse?

No. It is frequently the newest and most capable option available. The label describes the commitment you are being offered, not the quality of the output.

How did this land?

About the author

Cecilia Iona
Cecilia Iona

Senior Editor, AI & Product

Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.

Share

Get the next post in your inbox

One email a month. Product updates, engineering posts, and the best of Built with Swarmz.

I agree to receive emails about AI building tips and Swarmz product news. Unsubscribe any time.