Read four things in a provider changelog entry and you have covered most of what can break your feature: changes to the request contract, model ids with a retirement date attached, changes to defaults, and whether the id you call is a preview. The capability paragraph at the top is usually the least useful part of the page.
Read the request contract first
The fastest way to break a working feature is a change in what the API accepts, with no model swap on your side.
When Anthropic launched claude-sonnet-4-5-20250929, the release notes stated that a new stop reason, model_context_window_exceeded, had been added, and that the memory tool and context editing launched in beta. A new stop reason is a branch your error handling does not have yet. The change log entry carries the provider's own wording and the action we attach to it: add handling for the new stop reason so long inputs fail visibly rather than silently.
This is the category to scan for verbs like removed, renamed, returns a 400, moved to. They cost a deploy. A capability paragraph costs a conversation.
Read the ids with a date on them
A deprecation entry is the only kind that has a deadline in it, so it is the only kind that belongs in a calendar.
OpenAI announced the deprecation of the chatgpt-4o-latest snapshot on 18 November 2025 with a shutdown date of 17 February 2026, and named gpt-5.1-chat-latest as the replacement. On 26 September 2025 OpenAI announced the deprecation of gpt-3.5-turbo-instruct, babbage-002, davinci-002 and gpt-3.5-turbo-1106 with a shutdown date of 28 September 2026. Anthropic notified developers using Claude Sonnet 3.7 on 28 October 2025 that the model would retire, and it retired on 19 February 2026 with claude-sonnet-4-6 listed as the replacement. All three are in the change log with the provider page as the source.
Notice the shape of the second one. Four ids, a year of warning, and almost certainly some batch job or internal tool that nobody has looked at since it was written. That is the entry that bites, not the flagship launch.
How much warning you get is a commitment each provider publishes for itself, and the three commitments are different. Anthropic says customers with active deployments get at least 60 days[src] before a publicly released model is retired, and its page states that partner platforms set their own schedules.[Anthropic] OpenAI says at least 6 months[src] for a generally available model, at least 3 months for specialized variants, and that preview models may be retired with much shorter notice, such as 2 weeks, with no minimum committed.[OpenAI] Google says Gemini preview models may be used in production but will be deprecated with at least 2 weeks[src] notice.[Google] Three companies, three promises, no industry norm. Read the one that applies to the platform your feature runs on.
It also helps to know the vocabulary. Anthropic describes its model lifecycle in four states, active, legacy, deprecated and retired, and says requests to retired models will fail.[Anthropic] Those are Anthropic's words for Anthropic's models. Other providers use different ones.
When one of these entries names something you actually run, the first hour after a provider announcement is the ordered checklist for what to do about it before you touch code.
Read the defaults
A default is a value you never wrote down, which is why a change to one is easy to miss.
Effort levels, thinking modes, caching minimums and sampling parameters all sit in this category. They do not appear in your code, so a diff will not show them, and a change to one lands in production the moment the provider ships it. Our rerun protocol treats a default change as a tracked provider update, which means a run against your golden dataset (the fixed set of labelled inputs and expected outputs agreed at baseline, versioned so two runs can be compared) rather than an assumption either way.
Read whether you are on a preview
Google's Gemini documentation says stable models usually don't change and that most production apps should use a specific stable model.[Google] Keep Google's hedge intact: usually, not never.
Preview ids move faster and stop sooner. When Google launched gemini-3-flash-preview on 17 December 2025, the sensible reading was to add it to a comparison run and keep the preview id out of production pins until a stable id shipped. The same logic applies to any feature that is currently serving from a preview: it needs a shorter rerun cycle than one on a general release, because the notice period behind it is shorter. The tradeoff between the two is the subject of pin the id, or ride the alias.
What to skip
Skip the ranking language. When OpenAI released GPT-5.2 on 11 December 2025 it stated improvements in reasoning, vision, code generation and spreadsheet understanding, and when it released GPT-5.1 on 13 November 2025 it described a flagship optimized for steerability, code generation and agentic workflows. Those sentences are useful for one decision only: whether the model is worth a comparison run against yours. They are measured on the provider's own suite. They are not a statement about your prompt, your corpus or your users, and we never restate them as one.
Which action, in short
Every entry we publish gets exactly one of four actions: rerun, re-pin (move your code from a floating model alias to a fixed, dated model id), re-prompt, or migrate. The vocabulary is deliberately small, because an action list that needs interpretation gets skipped. Which one an entry gets depends on which of the fields above it touched, not on how the capability paragraph reads.
The first hour after a provider announcement walks through the checks in the order to run them and maps each answer to its action. If you want the schedule rather than the reading list, one of our written guarantees is that every tracked provider update triggers a rerun within 2 business days of the provider's announcement, and the plans page says which features that covers.