15 AI models hit end-of-life in the next 90 days — the full table

If you build on hosted LLM APIs, model retirements are the quiet breakage: a model ID keeps working right up until the day it returns a 404, and the deprecation notice was a line in a changelog you didn't read.

I maintain a catalog of model lifecycles, so here's a concrete, dated snapshot: 15 models across 6 providers are scheduled to hit end-of-life within the next 90 days (by 2026-09-30). Dates and replacements are from each provider's own docs.

The table

Retires	Provider	Model (API id)	Official replacement
2026-07-23	OpenAI	o3-deep-research	gpt-5.5-pro
2026-07-23	OpenAI	computer-use-preview	gpt-5.4-mini
2026-07-23	OpenAI	gpt-5-codex	gpt-5.5
2026-07-24	DeepSeek	deepseek-chat (legacy alias)	deepseek-v4-flash
2026-07-24	DeepSeek	deepseek-reasoner (legacy alias)	deepseek-v4-flash
2026-07-31	Mistral	mistral-small-3-2	mistral-small-2603
2026-07-31	Mistral	mistral-nemo	ministral-3-8b-latest
2026-08-05	Anthropic	claude-opus-4-1	claude-opus-4-8
2026-08-31	Mistral	mistral-medium-3-1	mistral-medium-latest
2026-09-08	Alibaba	qwen3-max	qwen3.7-max
2026-09-08	Alibaba	qwen3-max-preview	qwen3.7-max
2026-09-08	Alibaba	qwen3-6-max-preview	qwen3.7-max
2026-09-14	Amazon	amazon-nova-premier	amazon.nova-2-lite-v1:0
2026-09-24	OpenAI	sora-2	— (no drop-in successor)
2026-09-24	OpenAI	sora-2-pro	— (no drop-in successor)

A few things worth flagging

• The DeepSeek ones are aliases, not models. deepseek-chat and deepseek-reasoner are pointers; after 2026-07-24 they resolve to deepseek-v4-flash. If you pinned the alias expecting stable behavior, the behavior changes under you — pin the concrete snapshot instead.
• claude-opus-4-1 → claude-opus-4-8 is not a free swap. The 4.7-generation tokenizer counts roughly 30–35% more tokens for the same text, so your per-request cost and context math shift even if the price-per-token looks similar.
• Sora 2 / Sora 2 Pro have no listed drop-in replacement. If a video pipeline depends on them, that's a migration to scope now, not in September.

How to not get surprised by this

The durable fix is to treat model lifecycle as a dependency you monitor, the same way you watch for CVEs or deprecated framework APIs:

1. Pin concrete model snapshots, never floating aliases, in anything you can't hotfix fast.
2. Keep a list of the model IDs you actually call in production, and check their retirement dates on a schedule.
3. Test the named replacement before the retirement date — pricing, tokenizer, and output format can all move.

I keep the full, dated deprecation list (all providers, not just the next 90 days) updated here: aimodelwatch.dev/deprecations. If you want a heads-up when a model you use gets a retirement date, there's a free email alert on the site.

Data pulled 2026-07-02 from provider docs. If you spot a date that's drifted, tell me — accuracy is the whole point.