The Canyon ReviewAI model retirement notice, ranked by how much warning you get
Bedrock promises a year, OpenAI six months, Anthropic sixty days, Google a date that may move. Four ways to buy the same models, ranked by fuse length as of August 2026.
You want the model you build on to still answer next spring. Whether it does depends less on anyone's roadmap than on the notice policy written into the vendor's docs, and those policies differ by a factor of six. Bedrock wins on warning. Anthropic direct gives you the shortest fuse. Google gives you a date with an asterisk on it.
Criteria, stated before the table: the written minimum notice before a retirement (50%), the notice actually given on the most recent retirement we could find (30%), and whether you can keep calling a retired model at all (20%). Desk research, as of August 22, 2026. Every number links to the policy page it came from.
The table
| Rank | Where you buy | Written minimum | Last retirement, notice given | Verdict |
|---|---|---|---|---|
| 1 | Amazon Bedrock | 12 months on the platform, then 6 months in Legacy before end of life | Claude Opus 4.1: Legacy July 8, 2026, EOL January 8, 2027 | the calm option |
| 2 | OpenAI | 6 months for GA models, 3 for variants, about 2 weeks for previews | Assistants API: announced August 26, 2025, off August 26, 2026 | fine, read the model name |
| 3 | Anthropic direct | 60 days | Claude Opus 4.1: deprecated June 5, retired August 5, 2026 | newest models, shortest fuse |
| 4 | Google Gemini API | none stated; the table lists "earliest possible dates" | gemini-2.0-flash: June 1, 2026 | skip anything marked preview |
Notes per rung
Bedrock's edge is a single fact. The same Opus 4.1 that Anthropic retired on August 5 stays callable on Bedrock until January 8, 2027. Five extra months, for the same model, is the entire product. Read the catch in the same document: for models with an EOL after February 1, 2026, the final stretch of Legacy is "public extended access", where the doc says you "should expect higher pricing, which will be set by the model provider". You are buying time. The price of the time is not published, which costs Bedrock points but not the rank.
OpenAI's six months is real and the calendar backs it: the gpt-5 family was announced for retirement on June 11, 2026 and shuts down December 11. The tier that bites is the third one. Anything with preview in the name can go in two weeks, and the docs say so plainly. Twelve months for the Assistants API was generous by the same policy, and still ended with teams migrating in August.
Anthropic's 60 days is the shortest written floor on this page, and the observed number matches it: 61 days for Opus 4.1. Two things soften it. Active models carry a "not sooner than" date in the same table (Sonnet 4.5 reads not sooner than September 29, 2026), which is the most useful column any vendor publishes. And Bedrock exists, above, if you want the same model on a longer leash.
Google gets the bottom rung for the same reason a pricing page that says "contact sales" does: no written minimum means no number to hold anyone to. The deprecation table gives shutdown dates as the earliest possible and promises advance notice of the real one. Preview models there have lived a few months; treat them as demos.
The skip
Building anything customer-facing on a model id with preview in it, from any vendor on this list. Every policy above carves previews out of its own guarantee. Pin a dated, generally available id, read the not-sooner-than column where one exists, and budget one migration per model per year. The vendors have already told you the fuse length. They wrote it down. Most teams find the docs after the email.