AI companies rarely flip a switch and ship a product to everyone at once. Most launches move through a series of stages, and the label attached to a product tells you a lot about how much you can rely on it.

Research preview and waitlists

A research preview is an early, deliberately limited release. Its main purpose is to learn how people use the system and where it fails. Features can change without notice, usage limits are tight, and there is usually no service-level agreement. A waitlist controls how fast new users arrive, which protects scarce compute capacity during the first weeks.

Alpha, beta and “preview”

Cloud providers formalize these stages. Google Cloud, for example, distinguishes Preview from General Availability, and Microsoft’s Azure publishes separate supplemental terms for preview services. The common thread: preview features may have limited support, may be priced differently, and are not recommended for production workloads.

General availability (GA)

GA means the vendor stands behind the product: documented APIs, published pricing, support commitments and, for enterprise offerings, uptime guarantees. For companies, GA is usually the point at which legal and procurement teams will approve a tool.

Staged rollouts

Even after GA, many AI features roll out gradually: first to a percentage of users, then by region or plan. That is why two people on the same subscription can see different features on the same day. Regional rollouts are often shaped by regulation, for example data-protection or AI rules in the European Union.

Deprecation

Models also have an end of life. Providers announce deprecation dates for older models and give developers a window to migrate. Teams that build on AI APIs should track these notices as carefully as launch announcements.

What to check when a launch is announced

  • Stage: preview or GA?
  • Availability: which countries, plans and platforms?
  • Pricing: published, or “free during preview”?
  • Data use: are your inputs used for training, and can you opt out?
  • Limits: rate limits, context length, file sizes.

Reading the fine print of a launch takes five minutes and can save weeks of rework.

Why staged launches are common in AI

AI systems behave unpredictably in ways that are hard to test fully before release: users find unexpected uses, new failure modes and ways to misuse them. Staged rollouts let companies watch real-world behavior, fix problems and scale infrastructure gradually. Capacity is another reason — popular launches can overwhelm computing resources, leading to waitlists and usage limits.

Model versions and naming

Providers often release several versions of a model — larger and smaller, faster and slower — and update models over time. API users can often choose a specific dated version for stability or an alias that always points to the latest. Changes to the model behind an alias can alter results, so production applications usually pin versions and test before upgrading. Deprecation notices give developers time to migrate before old versions are retired.

Regional availability

Features may launch in some countries before others because of regulation, language support or infrastructure. In the EU, for example, data protection and AI rules have sometimes delayed or modified launches. Check availability and terms for your region before planning around a new feature.

See the numbers in AI announcements explained and our cheat sheet for choosing the right AI model.

Tip

Subscribe to providers’ changelogs to learn about version changes before they affect you.

What to watch next

  • Longer support windows for model versions.
  • Regional differences in AI feature availability.
  • Transparency about changes in model behavior after updates.
  • Launch practices for agents with more autonomy.

Key takeaways

  • AI products often move from research previews and betas to general availability.
  • Staged rollouts manage unpredictable behavior and capacity limits.
  • Pin model versions in production and test before upgrading.
  • Deprecations retire older versions on announced timelines.
  • Availability can differ by region because of regulation and infrastructure.

Frequently asked questions

What does “general availability” mean?

The product is officially released for broad use, usually with support commitments and stable terms.

Why is a feature not available in my country?

Regulation, language support, legal review or infrastructure can delay launches in some regions.

What happens when a model is deprecated?

It is scheduled for retirement; users must migrate to a newer version before the end date.