How to read an AI tool changelog without assuming you have the feature

An AI release note rarely means every user received one identical change at the same moment. Before you rewrite a workflow around a new feature, identify six things: product, model, account plan, platform, region and rollout state.

Separate the product from the model

“ChatGPT,” “the API” and a named model are not interchangeable. A product update may change the interface, connectors or memory without changing the underlying model. A model update may reach an API endpoint before, after or separately from a consumer app.

OpenAI maintains separate ChatGPT release notes and model release notes. That separation is a useful reminder for any vendor: first ask which surface the announcement describes.

Record the entitlement

Look for plan language such as Free, Plus, Pro, Team, Enterprise or Education. Then look for administrator controls. An enterprise feature can be generally available while still disabled in a particular workspace. Likewise, a usage limit can differ by plan even when the button looks the same.

If the note does not name a plan, do not assume universal access. Check the linked help page, pricing table or administrator documentation. Save the date you checked; providers revise these pages.

Check platform and region

Web, iOS, Android, desktop applications and API clients can ship on different schedules. Mobile availability may also depend on an app-store release. A feature involving data residency, voice, payments or third-party services may be limited by country.

Use the exact wording. “Starting on web” is not “available everywhere.” “In the US” is not a global release. If a provider omits region information, report that the note does not specify it.

Treat rollout verbs as status information

The verbs matter:

• rolling out means some eligible accounts may not see it yet;
• preview, beta or experimental signals a feature may change or be withdrawn;
• generally available usually describes release stage, not necessarily every plan;
• deprecated starts a transition, while retired identifies an end state.

Record both the announcement date and any stated completion window. A screenshot taken during a staged rollout cannot prove what another account should see.

Test the smallest real task

When the feature appears, test it with a non-sensitive example that resembles your workflow. Note the selected model, settings and timestamp. Check output quality, citations or tool calls, export behavior and whether the result persists across sessions.

Do not use a marketing demo as the acceptance test. A demo shows an intended path; your test shows whether the current product, account and data produce the behavior you need.

Keep a compact change record

For each update, store the provider link, publication date, product, model, eligible plan, platform, region, rollout wording and the result of your test. If any field is unknown, say unknown. This turns a fast-moving announcement into a checkable operational note—and prevents “the feature launched” from becoming a claim broader than the evidence.