Key figure: In the migration case documented on this page, the replacement models were not substitutes: gpt-realtime-1.5 scores about 75%, and gpt-realtime-2 / 2.1 drop to 0–20% success (figures reported in the case write-up below).
This AI agents 2026 hall of shame documents real deprecation disasters: teams broken by AI model retirements in 2026, from the October 23 wave to silent sunsets. Learn the migration lessons cheaply.
Five documented cases: fine-print deprecations that broke production, widgets orphaned overnight, workbench amputations, voice agents that lost their voices, and sunsets that outlasted their own warnings. Each 2026 case includes what the vendor announced, what actually broke, and the migration lesson.
AI vendors retire models the way phone companies retire networks: on a schedule they publish and a timeline you discover too late. Every entry below is a verified, sourced incident — no invented companies, no anonymous horror stories. Read them, then take the 7-step never-again checklist so your product is never the next entry.
The October 23, 2026 wave is the largest OpenAI model shutdown of the year: seventeen model identifiers stop working in the API, announced April 22, 2026. The casualties include gpt-4-0613, gpt-4-turbo, o1, o3-mini, o4-mini, and gpt-image-1. But the trap is in the fine-tuned rows: six ft-* models carry the same 2026-10-23 shutdown date — while the same deprecations page contains a sentence saying fine-tuned models "are not affected." That sentence sits under 2023 entries and governs nothing in 2026; the reliable signal is the table. And the door closes twice: the ability to create new fine-tuning jobs ends entirely on January 6, 2027, so teams on ft-* models have roughly ten weeks to retrain onto a supported base.
Sources: our model obituaries (core retirement list); Digital Matters, "OpenAI Model Deprecations: The October 23 List" — digitalmatters.me/artificial-intelligence-ai/openai-october-2026-model-shutoff/ (the 17-identifier count, the ft-* rows, the contradictory 2023 sentence, the January 6, 2027 fine-tuning wind-down).
A solo founder built an AI chat widget that qualifies leads on business websites on a stack of Cloudflare Pages, Netlify Functions, Supabase, and Groq — with Llama 3.3 70B as the LLM. Then Groq deprecated the model his whole product was built around. What actually broke: the model name was hardcoded in several places, the system prompts were unknowingly fitted to that model's quirks (instruction following, JSON formatting, long system prompts), and the qualification flow depended on clean structured output. Swapping in a "similar" model was not a one-line change. His fix is now the template: model list in an environment variable with a fallback, a small regression set of real conversations to judge swaps, and paid inference with fallback protection for anything customers depend on.
Source: Ansar Ali, "What I Learned When Groq Deprecated the Model My Product Depended On" (DEV Community, Oct 7, 2026) — dev.to/ansarali/what-i-learned-when-groq-deprecated-the-model-my-product-depended-on-9g4. Product named in the article: Zappiq AI.
On August 17, 2026, Anthropic retired the legacy Claude API Workbench — and with it, every saved prompt, named variable, and eval configuration stored there. The notice window was 31 days, shorter than Anthropic's own documented 60-day minimum for model deprecations. Teams that didn't export before the cutoff have no documented recovery path — Anthropic states the data is simply no longer accessible. Three experimental API endpoints (/v1/experimental/generate_prompt, /v1/experimental/improve_prompt, /v1/experimental/templatize_prompt) now return errors, breaking any automation built on them. The replacement Playground is deliberately stateless: nothing is stored on Anthropic's servers.
Source: TechTimes, "Anthropic Kills Claude Workbench Today: Saved Prompts Gone, API Pipelines Broken" (Aug 17, 2026) — TechTimes article, citing Anthropic's July 17, 2026 release notes and the Claude Help Center.
WebPlanning-AI, a B2B voice-driven planning SaaS, built its entire architecture around gpt-realtime (snapshot gpt-realtime-2025-08-28): a non-reasoning model with strict, structured tool-calling instructions and very low variance — 98–100% success on their production test corpus. Then the model was flagged "Deprecated," with a shutdown date of January 20, 2027. The replacements aren't substitutes: gpt-realtime-1.5 scores ~75%, and gpt-realtime-2 / 2.1 drop to 0–20% success with up to 14s response times, reasoning-driven parameter interpretation, and commentary leaking into user-facing audio. A second team migrating a structured outbound voice agent from realtime-mini hit the same wall. "Upgrading" changes the behavioral contract in a way that breaks production systems built on deterministic instruction following.
Source: reported by the teams themselves on the OpenAI Developer Community forum — community.openai.com thread (July 2026). The January 20, 2027 date was confirmed in-thread by OpenAI staff citing the published deprecations page.
Twelve months after the notice went out, the Assistants API shutdown took effect in August 2026 — and the twelve months of warning ran out on "thousands of production apps," per Digital Matters' coverage of the deprecation wave. It's the clearest proof that a generous notice window doesn't save anyone: even a full year isn't enough when the migration competes with every other roadmap item. If your team filed the Assistants notice under "we'll deal with it later," August was later.
Source: Digital Matters, "OpenAI Model Deprecations: The October 23 List" — digitalmatters.me/artificial-intelligence-ai/openai-october-2026-model-shutoff/. The October wave article notes OpenAI has since published migration material for the Assistants API — for the October 23 model wave, there is no equivalent migration guide, script, or tool.
gpt-4o-2024-05-13 dies while bare gpt-4o survives. Grep for literal identifiers, not model names.model field from 30 days of your own request logs and count distinct values.model_not_found, which is indistinguishable from "does not exist" or an entitlement problem. Alert on a drop in successful completions instead.Checklist synthesized from the incidents above and the linked sources — defensive guidance, not vendor advice. Where a step rests on a specific source's claim, that source is cited in the entry above.
The vendor sets a shutdown date; after it, API calls to that model identifier stop working and must be moved to a replacement. Deprecations can also change behavior and pricing — price the migration before the deadline, don't scramble after it.
Seventeen model identifiers, including gpt-4-0613, gpt-4-turbo, gpt-4o-2024-05-13, gpt-image-1, o1, o3-mini, o4-mini, and six fine-tuned models (ft-gpt-3.5-turbo, ft-gpt-4, ft-gpt-4.1-nano, ft-o4-mini, ft-babbage-002, ft-davinci-002).
OpenAI's page states nothing about exporting weights or moving a tuned model onto a newer base — no porting path offered, none ruled out. Meanwhile creating new fine-tuning jobs ends entirely on January 6, 2027. Plan fine-tune exits early.
It varies: OpenAI announced the Oct 23, 2026 wave about six months out; Anthropic gave 31 days for the Aug 2026 Workbench retirement; some teams got no usable warning at all. Never assume a comfortable window.
Pin exact snapshots, keep model names in config with a fallback, maintain a regression set of real inputs, audit where model IDs hide (databases, CMS config, env dashboards), and run a deprecation radar like ModelMinder over your fleet.