--- title: "AI Model Retirement: What Happens to Your Agents" url: https://www.insulin.dev/blog/ai-model-retirement/ canonical: https://www.insulin.dev/blog/ai-model-retirement/ type: Blog description: "AI model retirement in Insulin: where agents, jobs, apps and knowledge bases pin a model, what the docs say happens on the date, and what to change first." --- # AI Model Retirement: What Happens to Your Agents > Canonical HTML version: https://www.insulin.dev/blog/ai-model-retirement/ 1. [Home](/) 2. / 3. [Blog](/blog/) 4. / 5. AI Model Retirement: What Happens to Your Agents # AI Model Retirement: What Happens to Your Agents Claude Sonnet 4.5 retires on November 30, 2026, and a custom agent pinned to it won't quietly switch models: it stops. Every place Insulin pins a model, and what to change first. ![Shirley Guo](/authors/shirley-guo.jpg) Shirley Guo Oct 7, 2026 ![AI Model Retirement: What Happens to Your Agents](/images/blog/ai-model-retirement/hero.png) Explore AI Summary [![](/logos/company/openai.svg)](https://chat.openai.com/?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-model-retirement%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Integrations%2C%20Agents. "Summarize with ChatGPT")[![](/logos/company/anthropic.svg) ](https://claude.ai/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-model-retirement%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Integrations%2C%20Agents. "Summarize with Claude")[![](/logos/company/gemini.svg)](https://www.google.com/search?udm=50&aep=11&q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-model-retirement%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Integrations%2C%20Agents. "Summarize with Gemini")[](https://www.perplexity.ai/search/new?q=Read%20and%20summarize%20https%3A%2F%2Fwww.insulin.dev%2Fblog%2Fai-model-retirement%2F%2C%20then%20cite%20the%20source.%20Focus%20on%20what%20it%20says%20about%20Integrations%2C%20Agents. "Summarize with Perplexity") Table of Contents - [What happens when a provider retires a model?](#what-happens-when-a-provider-retires-a-model) - [Where does Insulin pin a model?](#where-does-insulin-pin-a-model) - [What does a custom agent do when its default model is retired?](#what-does-a-custom-agent-do-when-its-default-model-is-retired) - [What about Insulin’s own default, jobs and apps?](#what-about-insulins-own-default-jobs-and-apps) - [What happens to a knowledge base’s embedding model?](#what-happens-to-a-knowledge-bases-embedding-model) - [How do you find which models you actually use?](#how-do-you-find-which-models-you-actually-use) - [What do you press when a turn stops on a retired model?](#what-do-you-press-when-a-turn-stops-on-a-retired-model) - [Frequently asked questions](#frequently-asked-questions) - [Takeaways](#takeaways) _AI model retirement is the date a provider stops serving a model, after which requests to it fail. In Insulin, a custom agent whose default model is retired stops with an error instead of quietly switching models; this post maps every other place a model is pinned, what the documentation says happens there, and what to change before Claude Sonnet 4.5 retires on November 30, 2026._ * * * Anthropic and OpenAI, two of the bring-your-own-key providers Insulin’s documentation names alongside Gemini, published retirement notices a day apart. On September 30, 2026, Anthropic notified developers using Claude Sonnet 4.5 that it retires from the Claude API on November 30, 2026. On October 1, OpenAI deprecated `gpt-5.1`, `gpt-5.3-codex` and `gpt-5.4-nano`, which it will remove from its API on April 1, 2027. If you or your organization connected your own Anthropic or OpenAI key, and someone chose one of those models for an agent, a job or an app, the provider’s date is now a date in your workspace too, and Insulin doesn’t treat every pin the same way. ## What happens when a provider retires a model? **Requests to it stop working.** Anthropic defines a retired model as one that “is no longer available for use” and says “Requests to models past the retirement date will fail.” OpenAI says that at shutdown “the model or endpoint will no longer be accessible.” These are the newest notices from each, leaving out OpenAI’s text-to-speech models: Provider Announced Model Retirement or shutdown date Recommended replacement Anthropic September 30, 2026 Claude Sonnet 4.5, `claude-sonnet-4-5-20250929` November 30, 2026 `claude-sonnet-5-5` OpenAI October 1, 2026 `gpt-5.1` April 1, 2027 `gpt-6-sol` OpenAI October 1, 2026 `gpt-5.3-codex` April 1, 2027 `gpt-6-sol` OpenAI October 1, 2026 `gpt-5.4-nano` April 1, 2027 `gpt-6-luna` Three details on the providers’ own pages change how you read that table: - **The newest notice is not the nearest date.** [OpenAI’s deprecations page](https://developers.openai.com/api/docs/deprecations) also carries older notices with sooner shutdowns: one from April 22, 2026 shuts older snapshots, including `o4-mini` and `gpt-4.1-nano`, down on October 23, 2026, and one from June 11, 2026 shuts older GPT-5 and o3 snapshots, such as `gpt-5-2025-08-07`, down on December 11, 2026. Check each pinned model’s exact name against the provider’s table, not only against its latest announcement. - **Anthropic calls its date tentative; OpenAI calls its date a shutdown.** [Anthropic’s model status table](https://platform.claude.com/docs/en/about-claude/model-deprecations) puts November 30, 2026 in a column headed _Tentative retirement date_, while OpenAI’s tables give a _Shutdown date_. - **Notice periods differ.** Anthropic gives “at least 60 days’ notice before model retirement for publicly released models” and recommends testing the new models well before the retirement date. OpenAI’s minimum notice is six months for generally available models and three months for specialized variants such as `gpt-5.3-codex`, unless safety or compliance concerns require a faster timeline, and preview models can go on much shorter notice, such as two weeks. Neither page covers Gemini. Google keeps its own [Gemini API deprecations page](https://ai.google.dev/gemini-api/docs/deprecations) for those models. ## Where does Insulin pin a model? **A model is pinned in five places: a custom agent’s Default model, the built-in assistant’s Default row, a manual job, the app builder and a knowledge base’s embedding model. Inbox has none.** Where the model is pinned What the documentation says happens when it can’t be reached What to change before the date **A custom agent’s Default model** Marked ` — unavailable`. The agent never silently switches, and a turn that can’t reach it stops with an error Choose a replacement as the **Default model**, and save **The built-in assistant’s Default row** It runs first, then fails over to the others. A retired Default isn’t described Click the replacement’s row, so it carries the **Default** badge **A manual job that pins a model** Not described, and there is no documented way to change the pin Note each job whose model badge names a retiring model, and open it after the date **The app builder’s model** Not described. The model you pick is pinned on the app Pick the replacement in the builder’s model picker **A knowledge base’s embedding model** Insulin’s precedent: Gemini embedding 001, withdrawn from new selections, keeps working where it is already indexed and shows as _legacy_ Switch only on purpose, because a switch re-embeds every document **Inbox** No pin. It walks the chat agent’s chain of candidate models, with Fours’ hosted models last Nothing to pin **Inbox has no model to pin.** Its **AI model** section, in Inbox **Settings → Account**, only reports whether each provider is **Connected**, and if your first provider is unavailable or has hit a usage limit, the next connected provider takes over its work outside a chat. Insulin’s documentation describes no warning while a deprecated model is still answering, so until the date, the provider’s table is the list to check against. ## What does a custom agent do when its default model is retired? **It stops and says so, instead of quietly answering on another model.** The documentation names retirement outright: when a saved default becomes unavailable because “its provider was disconnected or the model was retired”, the picker marks it ` — unavailable` with a hint to reconnect the provider or choose another model, and “the agent never silently switches to a different model.” A turn that can’t reach the saved default stops with an error saying the model is not available. If the default fails partway through a turn instead, the reply names the model that took over, which is [how Insulin keeps failover visible rather than silent](/blog/who-picks-the-model/). After a retirement, reconnecting the provider won’t bring the model back, so choose another. Three details decide whether that change goes through: - **A save re-checks the model only when you change it.** Saving an edit to the prompt or the knowledge bases doesn’t re-check the model, so a successful save is no evidence the default is current. - **A replacement can be refused on save**, with the reason the runtime would give: _AI policy_ (your organization’s AI model policy doesn’t permit it), _Model unavailable_ (it can’t be reached for this agent’s scope) or _Nothing to resolve with_ (no provider could be resolved for the scope). - **Not everyone can make the change.** A Personal agent is private to its owner. On an Organization agent, EDITOR and ADMIN can edit the default model, while a USER sees the configuration read-only. ## What about Insulin’s own default, jobs and apps? **For these three, the documentation says where the model is set but not what happens when it retires, so change what you can now and look again after the date.** - **The built-in assistant’s Default row.** The pencil in a chat with Insulin, labelled _Insulin model settings_, opens a Models list where clicking a row makes that model the **Default**. Insulin runs that model first, then fails over to the others from there; how the panel shows a Default whose model has retired isn’t described. Re-pick it now, and the first model Insulin tries stays one the provider still serves. - **A manual job’s pinned model.** A model badge sits between the agent’s name and the timestamps on the job’s detail page, but only on a manual job that pins a model. A manual job without one inherits its agent’s default model, which the agent’s row covers; a cron or push job never shows a badge, so its page can’t tell you. The documentation describes no way to change a job’s pinned model, and no retirement behaviour for one. After the date, open each job you noted: its card shows the latest run’s **Result**, or an **Error** with the message the run ended with. - **The app builder’s model.** The builder’s model picker, the sparkle button near the composer, chooses which model generates the app, and the model you pick is pinned on the app, so saved apps keep building with the same model. What the builder does once that model retires isn’t described. Pick the replacement before the date. Only a job’s owner can see it, and being an organization administrator doesn’t change that, so each owner checks their own jobs. ## What happens to a knowledge base’s embedding model? **Insulin’s own precedent is that the knowledge base keeps working.** Insulin has stopped offering Gemini embedding 001 for new selections, in favour of Gemini embedding 2, but a knowledge base already indexed with it keeps working, can still be searched and re-indexed, and shows the model as _legacy_. That is Insulin withdrawing an option, not a provider shutting a model down. What an indexed knowledge base does in the second case isn’t described, so check after any such date. Moving a knowledge base to another embedding model is a rebuild: switching re-embeds every document, which takes provider time and spend, and a document stays out of search until it has been re-embedded. [Which embedding models Insulin offers](/blog/which-models-serve-insulin/) is the list to choose from, and [the runbook for changing an embedding model without breaking search](/blog/switch-embedding-models-without-breaking-search/) covers the move. ## How do you find which models you actually use? **Group the Usage panel by Model, then walk the pins anyway, because Usage shows what ran, not what is set.** On **Settings → Billing**, the Usage panel’s **Model** grouping answers “which model was used”. - **To see who ran a retiring model, and in which app,** pick it in the **models** filter while the Model grouping is showing, then switch to **User** or **App**. The selection stays applied as you switch, and scheduled runs show up as _No user — scheduled and system runs_ rather than under their owner. - **It covers one month at a time**, chosen from the last 12, and its filters offer only what that month’s results contain. An agent nobody used that month won’t appear, and a quarterly job’s model shows only in the month the job ran. - **Usage on your own key is included** whenever the provider reports how many tokens the request used. The breakdown is re-aggregated every couple of hours, so the newest turns may be missing. - **Anthropic has its own audit.** **Export** on the Claude Console’s Usage page downloads a CSV of usage by API key and model. Usage can tell an admin a model was used; only the owner of a personal agent or a job can open it to see where. ## What do you press when a turn stops on a retired model? **Switch model.** A failed turn carries its own buttons: **Retry** re-sends the message that failed, and **Switch model** opens the agent’s model settings so you can pick a different model before trying again. A failure classified _Model unavailable_, meaning the model itself could not be reached, offers both. Retry changes nothing about the model, so after a retirement Switch model is the button that fixes the turn. [Which stop cards offer Retry, Switch model or both](/blog/usage-limit-or-billing-pause/) covers the other classifications. ## Frequently asked questions ### What happens to an Insulin agent when its model is retired? A custom agent’s picker marks its saved Default model unavailable, with a hint to reconnect the provider or choose another model. The agent never silently switches to a different model: a turn that can’t reach the saved default stops with an error saying the model isn’t available. ### When does Claude Sonnet 4.5 retire? Anthropic’s notice of September 30, 2026 gives November 30, 2026 as the retirement date for claude-sonnet-4-5-20250929, with claude-sonnet-5-5 as the recommended replacement. Its status table calls the date tentative. Anthropic says requests to models past the retirement date will fail. ### Does Insulin warn you before a model is retired? The documentation describes no warning while a deprecated model is still answering. It describes what you see afterwards: a custom agent’s Default model marked unavailable, and a turn that stops with an error. Until the date, check each pinned model’s name against the provider’s own deprecations table. ### How do I find which models my organization uses? Group the Usage panel on Settings → Billing by Model. It shows what ran in one month, chosen from the last 12, not what is configured, so check each pin too. Usage on your own provider key appears whenever the provider reports the request’s token count. ### Can I change the model a job has pinned? The documentation describes no way to change a job’s pinned model, and no retirement behaviour for one. A manual job that pins a model shows a model badge; a cron or push job never shows one. Organization administrators can’t see other members’ jobs, so owners check their own. ### What happens to a knowledge base when its embedding model is retired? The documented case is Gemini embedding 001, which Insulin no longer offers for a new selection: a knowledge base already indexed with it keeps working, can still be searched and re-indexed, and shows as legacy. Switching to another embedding model re-embeds every document. ## Takeaways - **Check the provider’s table, not only its newest notice.** Claude Sonnet 4.5 retires on November 30, 2026 and `gpt-5.1`, `gpt-5.3-codex` and `gpt-5.4-nano` on April 1, 2027, while some older OpenAI snapshots go sooner. - **A custom agent stops rather than switching**, so change its Default model before the date. - **Nothing in the documentation warns you while a deprecated model still answers**, so the check is yours to run. - **Three pins have no documented retirement behaviour:** Insulin’s Default row, a manual job’s model and the app builder’s model. Re-pick the two you can, note the jobs, and look again after the date. - **An embedding-model switch is a rebuild**, so make one only on purpose. - **Usage shows what ran, not what is set**, and only owners can open personal agents and jobs, so everyone who owns one has to look. Retirement dates are the providers’ to set; where each model is pinned is yours to check. To see how an Insulin agent is scoped and given its Default model, start with [AI agents for business workflows](/agents/). ## Sources Primary sources for the platform rules cited above. Last verified October 7, 2026. Cloud providers change fees, eligibility, and program terms without notice — check the source before relying on a figure. - [Claude Platform Docs: Model deprecations](https://platform.claude.com/docs/en/about-claude/model-deprecations) — A retired model is no longer available for use and requests to models past the retirement date fail; at least 60 days' notice before retirement for publicly released models; the advice to test new models well before the retirement date; the September 30, 2026 notice retiring Claude Sonnet 4.5 (claude-sonnet-4-5-20250929) on November 30, 2026, with claude-sonnet-5-5 as the recommended replacement; the status table's Tentative retirement date column; the Claude Console usage export by API key and model - [OpenAI API: Deprecations](https://developers.openai.com/api/docs/deprecations) — At shutdown a model or endpoint is no longer accessible; minimum notice of at least six months for generally available models and at least three months for specialized variants such as gpt-5.3-codex, unless safety or compliance concerns require faster, and much shorter notice for preview models; the October 1, 2026 notice removing gpt-5.1, gpt-5.3-codex and gpt-5.4-nano on April 1, 2027, with gpt-6-sol and gpt-6-luna as replacements; the April 22, 2026 notice shutting older snapshots including o4-mini and gpt-4.1-nano down on October 23, 2026, and the June 11, 2026 notice shutting older GPT-5 and o3 snapshots such as gpt-5-2025-08-07 down on December 11, 2026 - [Insulin Agents — Fours Doc](https://doc.fours.com/insulin/agents/) — Anthropic, OpenAI and Gemini as bring-your-own-key providers; a saved Default model that becomes unavailable because its provider was disconnected or the model was retired is marked unavailable in the picker, with a hint to reconnect or choose another model, and the agent never silently switches; a turn that cannot reach the saved default stops with an error, and a default that fails partway names the model that took over; a save re-checking the model only when you change it, and the AI policy, Model unavailable and Nothing to resolve with refusals; Personal agents private to their owner, EDITOR and ADMIN editing an Organization agent's default model, and a USER read-only; Insulin model settings, the Default badge, and running that model first before failing over; Retry, Switch model and the Model unavailable classification - [Insulin Jobs — Fours Doc](https://doc.fours.com/insulin/jobs/) — The model badge only on a manual job that pins a model, none on a manual job inheriting its agent's default model, and never on a cron or push job; no edit form, with the documented edits covering a job's name, schedule, filter and assigned agent but not its model; a job visible only to its owner, organization administrators included; a manual job's Result and Error cards - [Insulin Custom Apps — Fours Doc](https://doc.fours.com/insulin/custom-apps/) — The builder's model picker, the sparkle button near the composer, choosing which model generates the app; the model you pick pinned on the app, so saved apps keep building with the same model - [Insulin Knowledge Bases — Fours Doc](https://doc.fours.com/insulin/knowledge-base/) — Gemini embedding 001 no longer offered for a new selection, in favour of Gemini embedding 2, while knowledge bases already indexed with it keep working, can still be searched and re-indexed, and show it as legacy; the confirmation that switching the embedding model re-embeds every document, which takes provider time and spend, and documents staying out of search until re-embedded - [Insulin Inbox — Fours Doc](https://doc.fours.com/insulin/inbox/) — The AI model section of Inbox Settings → Account showing Connected or Connect for each provider and only reporting state; Inbox's work outside a chat walking the same chain of candidate models the chat agent uses, the next connected provider taking over, and Fours' hosted models as the last resort - [Insulin Reading Your Usage — Fours Doc](https://doc.fours.com/insulin/billing/usage/) — The Usage panel on Settings → Billing grouped by Model; filter choices coming from the grouping shown and staying applied across groupings; one month at a time from the last 12; filters built from the period's own results; re-aggregation every couple of hours - [Insulin What Is Metered — Fours Doc](https://doc.fours.com/insulin/billing/metering/) — Usage on your own provider key appearing in the breakdown, attributed to the provider whose key was used, whenever the provider reports how many tokens the request used ## Keep reading - [IntegrationsAI Agent for BigQuery: Let It Query, Not ChangeOct 7, 2026](/blog/ai-agent-for-bigquery/) - [IntegrationsCloudflare Workers AI BYOK and DigitalOcean ModelsSep 28, 2026](/blog/cloudflare-workers-ai-byok/) - [IntegrationsEWS Retirement and Your AI Email AssistantOct 8, 2026](/blog/ews-retirement-ai-email-assistant/) - [SecurityRead-Only Access for AI Agents: Where It's SetOct 8, 2026](/blog/read-only-access-for-ai-agents/) [Browse every post on the Insulin Blog](/blog/) ### Stay Updated New posts, product updates and marketplace strategy are shared on LinkedIn as they publish. [Follow Fours on LinkedIn](https://www.linkedin.com/company/suger-inc)