The first hour after a provider announcement

Nothing in your codebase changes in the first hour. The hour is for reading, in a fixed order, so that the action you take afterwards is the smallest one that fits. Five questions, and most announcements stop at the first.

1. Is a model id you actually run named?

Open the entry and compare the model ids against your own inventory: every id in production, in staging, in batch jobs, in internal tools, and on every platform you call, including partner platforms that serve the same model under their own schedule.

If nothing matches, log it and stop. That is the outcome most of the time, and the discipline of checking rather than assuming is what makes the exceptions cheap. When something does match, carry on down the list.

One shape deserves attention here. On 18 September 2026 Google limited access to Gemini 2.5 models to projects already using them, and pointed developers to newer Flash ids instead. An entry like that changes nothing for a project already calling the id, and quietly closes the door for the next one. The Google change log carries Google's own wording.

2. Is it a deprecation with a date?

A deprecation is the only kind of entry with a deadline in it, so it is the only kind that goes in a calendar during the first hour.

Read the date, then check how much warning the provider that owns this model commits to for itself. The three providers publish three separate commitments, never a shared industry norm, and reading a provider changelog has the exact figure for each. The one that applies to you is the one for the platform your feature calls.

Check the vocabulary too: deprecated does not mean the same thing on every provider's page, and the same post has Anthropic's own four lifecycle states in full. On 14 April 2026 Anthropic notified developers using Claude Sonnet 4 and Claude Opus 4 of their retirement, and both retired on 15 June 2026. Two months between the notice and requests failing, recorded in the Anthropic change log.

3. Is it an alias repoint?

This is the entry that changes behaviour with no deploy on your side, which is why it needs its own question.

Google says a Gemini "latest" alias points to the most recent release of a model variation and is hot-swapped with every new release of that variation.[Google] That is the provider documenting the mechanism in its own words. It has fired twice in the period we track: on 21 January 2026 Google changed the latest aliases to point to Gemini 3 models, and on 24 June 2026 OpenAI updated the chat-latest snapshot to point to the latest Instant model in ChatGPT.

Neither announcement asks you to do anything. That is the point. If a feature was served through one of those aliases, the model behind it moved on the announcement date whether or not anyone read the entry.

4. Does your code pin a dated id, or ride a floating one?

The answer to question 3 only matters once you know this one, and a surprising number of teams have to go and look.

Whether it pins or floats decides whether question 3 already happened without anyone noticing. Reading a provider changelog covers what each provider says about stable versus preview ids, including Google's own hedge on the point. A re-pin (moving your code from a floating alias to a fixed, dated model id, then running your suite once against that id so later runs have something to compare with) removes one source of change. It does not make a model stable, and a pinned id is still retired on a schedule. The trade is covered in pin the id, or ride the alias.

5. Which of the four actions applies?

Every entry we publish carries exactly one, and the four are not interchangeable.

Rerun. A new model exists, or a capability shifted, and you want a comparison before anything moves. When Google released gemini-3.8-flash as generally available on 2 September 2026, and when OpenAI released GPT-6 Astra on 3 September 2026, the action was the same: add the new id as a second model in the next scheduled run of your golden dataset (the fixed set of saved inputs and expected results agreed at baseline, versioned so two runs can be compared), read the differences, change nothing yet.

Re-pin. An alias moved, or can move. Replace it with an explicit dated id and run the suite once against that id, so the next repoint is a decision rather than an event.

Re-prompt. The request contract changed, so your prompts or request builders have to change with it. Anthropic's 1 September 2026 launch of claude-fable-5-1 and claude-mythos-5-1 is the clearest recent example: tool_choice values of any or tool now fail. Google's 21 July 2026 entry deprecates the sampling parameters temperature, top_p and top_k on Gemini calls. Both mean code edits, then a rerun on the edited requests.

Migrate. An id is going away and there is a date. Inventory, replace, rerun on the replacement while the old id still serves, so you have two runs to read against each other rather than one number with no baseline.

What the hour does not include

Do not move a threshold (the agreed pass or fail line for a scorer, recorded in your plan and changed only by a dated decision) because an announcement arrived. Thresholds are agreed once and revised on their own record, not to make a rerun come out green. Our threshold protocol keeps each breach as a dated entry with what moved and what was decided, which is only useful if the line itself stays still.

Do not paraphrase the provider either. An entry without the provider's own source and wording is not published on this site, and a note you wrote from memory in the first hour is the one that will mislead you in week three.

Our rerun protocol runs this sequence on every tracked announcement, and one of our six written guarantees is that every tracked provider update triggers a rerun within 2 business days of the provider's announcement. What that covers per plan is on the plans page. If an update already changed what your feature returns, the week by week version is in the situation guide.