AI Deprecation Watch: November 2026 Deadline Calendar

TL;DR: The next eight weeks are the busiest retirement window the AI industry has ever had. November 30 is the big date: Anthropic retires Claude Sonnet 4.5 (requests fail after that), Google ends Gemini 4 Argon's $2/$10 introductory pricing (it doubles to $4/$20 on December 1), and OpenAI shuts down reusable prompts (v1/prompts), the Evals dashboard and API, and Agent Builder. Before that, Gemini Gems convert to Skills automatically on November 17, and on December 11 OpenAI retires custom GPTs (Enterprise first) plus the entire GPT-5 and o3 snapshot family. All deadlines below verified against provider pages and current reporting, 16 October 2026.

Model retirements used to be rare events you read about after the fact. In 2026 they are a scheduled part of the industry: nine Claude models retired in the twelve months before this one, OpenAI swept its entire GPT-4 and GPT-3.5 catalog in a single October day, and Google attaches expiry dates to launch pricing. The deprecation calendar is now a product feature you should compare before choosing a provider — right next to price and latency.

This post is the running calendar. One page, every deadline that matters between now and the end of the year, what replaces what, and exactly what to do before each date. It starts next week: on October 23 OpenAI shuts down the last GPT-4, GPT-3.5, o1, and o3/o4-mini names in a single sweep. This is the first edition of our Deprecation Watch series; each edition covers a new window and links back to the deeper migration guides instead of re-covering them.

The deadline calendar#

All dates verified 16 October 2026. Dates are the provider's own published dates — which, as the fine print shows, is not always the date your requests follow (more on that below).

November 30: Sonnet 4.5 goes dark#

Anthropic deprecated claude-sonnet-4-5-20250929 on 30 September 2026 and set retirement for 30 November 2026 — 61 days of notice, one day over the 60-day minimum Anthropic commits to for publicly released models. The recommended replacement is claude-sonnet-5-5.

Three things that catch people out:

1. Deprecated ≠ retired. Deprecated means the model still answers but is no longer recommended and a retirement date is assigned; retired means requests fail. Anthropic also warns that deprecated models are "likely to be less reliable than active models" — so the last weeks before retirement are not the time to leave a production workload on 4.5.
2. The date is per-platform. Anthropic's November 30 date covers the Claude API, Claude Platform on AWS, and Microsoft Foundry. Amazon Bedrock and Google Cloud set their own retirement schedules for the same model — check those tables separately if that's where you call it.
3. The swap is not one line. Sonnet 5.5 introduces breaking API behavior: a non-default temperature now returns a 400, assistant prefills return a 400, and responses can open with thinking blocks, which breaks code that reads content[0]. Our Sonnet 4.5 migration checklist walks through the full switch safely, including the pricing change — Sonnet 5.5 is $2/$10 per million tokens at list, and on the tokens.bd catalog it is served under the stable ID anthropic/claude-sonnet-5.5.

November 30 → December 1: Argon's intro price expires#

Google's Gemini 4 Argon launched with a footnote that turned out to be the most important line in the announcement: the $2/$10 per-million-token introductory price is valid only through 30 November 2026, and from 1 December every tier doubles. This is not a retirement, but it breaks budgets exactly like one — our full pricing and benchmarks breakdown has the complete table.

Google's published pricing, checked 16 October 2026, USD per 1M tokens:

The arithmetic for a steady workload — say 30M input and 3M output tokens a month: at intro rates that's (30 × $2) + (3 × $10) = $90/month; from December it's (30 × $4) + (3 × $20) = $180/month. The difference is $90/month, about ৳11,075 at the mid-market rate of 1 USD = 123.06 BDT (checked 16 October 2026). Price anything that will run past November at the doubled rate.

Worth noting the pattern, not just the date: Google reportedly did the same with Gemini 3.8 Flash, whose intro price is expected to double on 1 January 2027. Introductory pricing is becoming the industry's standard launch tactic — treat every launch price as temporary until the provider says otherwise in writing.

November 30: OpenAI's platform triple-shutdown#

Models get the headlines, but November 30 is also when OpenAI removes three platform features — announced June 3, 2026, with roughly six months of notice. What breaks here is not your model calls but your tooling:

• Reusable prompts (v1/prompts). If your code calls the Responses API with prompt: { id: "pmpt_…" }, that prompt text lives in OpenAI's dashboard, not your repository. Copy it into your code and send it as input before the shutdown.
• Evals dashboard and API. Existing evals become read-only on 31 October and the platform shuts down on 30 November. OpenAI points teams to Promptfoo as the migration path. If your evaluation history lives in Evals, export it now — a provider-hosted eval store is not a durable record.
• Agent Builder. The low-code UI for composing Responses API agents goes away; the migration paths are the OpenAI Agents SDK (developer teams) or ChatGPT Workspace Agents (no-code enterprise use). Note the Responses API itself is not deprecated — only the builder UI and saved-agent objects.

The deeper lesson: the layer you built to make model migrations safe — stored prompts, hosted evals — can itself be on a countdown. Keep the artifacts that matter (prompt text, eval datasets, workflow definitions) in your own version control.

November 17: Gemini Gems become Skills#

Google is retiring Gems — its custom-assistant feature — and replacing them with Skills, reusable instruction blocks you invoke with a slash command inside any conversation. For personal Google Accounts the transition starts in November 2026 (automatic conversion from November 17); Workspace business, enterprise, and nonprofit accounts follow in March 2027, and education accounts in June 2027.

For most free users this is an upgrade: Google recreates your Gems as Skills automatically and they stay free. The audit item is the exceptions: Skills currently don't work with Canvas, Deep Research, video and music creation, or Guided Learning — a Gem built around one of those tools won't behave the same after the move. If you have one, rebuild its workflow by hand and test it before the 17th. Also note the transition coincides with Google retiring the Opal and Google Labs Gems the same month.

Watch list: December 11 and the rest#

• 11 December 2026 — custom GPTs retire (Enterprise first). OpenAI's migration FAQ sets December 11 for custom GPTs becoming inaccessible in Enterprise workspaces, with other plans expected to follow the same timeline (workspaces with an approved deferral keep theirs until February 11, 2027). Instructions, knowledge files, and connected apps move with a one-button migration; custom actions, the model choice, conversations, and sharing settings do not — a migrated plugin starts private. If a GPT calls your own API through an action, that rebuild needs a developer, and OpenAI's current docs point to an MCP server.
• 11 December 2026 — the GPT-5 and o3 snapshot sweep. gpt-5-2025-08-07, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3-2025-04-16, and o3-pro all retire, replaced by gpt-5.6-sol / the GPT-6 family. This is the one most likely to catch a codebase that migrated off GPT-4 only a year ago.
• 13 November 2026 — Google AI Plus loses Pro models. AI Plus ($4.99/month) subscribers lose Gemini Pro models in the app and AI Mode, keeping only Flash-Lite and Flash — the fix is a 4x jump to AI Pro at $19.99/month. Worth knowing if your "cheap" Gemini workflow runs on a Plus subscription.
• Already happened: GPT-5.5 left ChatGPT, ChatGPT Work, and Codex on 14 October (API access unaffected). Still to come this month: see the 23 October row at the top of the table.

Find every retiring identifier in your code today#

The single most useful thing you can do this week is an inventory. This one-liner finds every identifier retiring in this window across a codebase (adjust the file globs to your stack):

grep -rEn "claude-sonnet-4-5-20250929|gpt-4o-2024-05-13|gpt-4-turbo|gpt-3.5-turbo-0125|o1-2024-12-17|o3-mini-2025-01-31|o4-mini-2025-04-16|gpt-4.1-nano-2025-04-14|gpt-image-1|gpt-5-2025-08-07|gpt-5-mini-2025-08-07|gpt-5-nano-2025-08-07|gpt-5-pro-2025-10-06|o3-2025-04-16|o3-pro-2025-06-10|gpt-image-1-mini|gpt-image-1.5|chatgpt-image-latest|pmpt_" \
  --include="*.py" --include="*.js" --include="*.ts" --include="*.go" --include="*.env" .

Then check the platform each one runs on — Anthropic's November 30 for Sonnet 4.5 doesn't automatically apply to Bedrock or Google Cloud, and OpenAI's dates don't apply to Azure. The deprecation date that matters is the one on the platform you actually call.

How to stop getting surprised#

Four habits that survive contact with the 2026 release pace:

1. Don't pin dated snapshots in long-lived code. A pinned snapshot like gpt-5-2025-08-07 is a hard expiry date in your source. Pin only where you need reproducibility (evals, benchmarks); use stable aliases everywhere else.
2. Route model calls through a gateway. When the model string lives in one config behind one API, a retirement is a config change, not a code change across five services. Our guide to using any LLM with the OpenAI SDK shows the one-line base_url version; the 60+ model catalog is the managed version, priced separately in taka and dollars so the exchange rate never ambushes you.
3. Calendar the dates and route the emails. Deprecation notices go to the account owner — if that's a departed contractor's inbox, the 60-day window expires unread. Verify who receives them this week.
4. Watch this series. Deprecation Watch runs monthly from here: each edition covers the next window and links the deeper guides. The retirement pace isn't slowing — the inventory habit is the durable fix.

Deprecation Watch FAQ#

What actually happens when a model is retired — do requests fail or just get worse? They fail. Anthropic's docs are explicit: requests to retired models fail outright. OpenAI's shutdowns work the same way — on the shutdown date, the model name stops answering. There is no grace mode and no degraded fallback; a retirement you missed is a production outage, which is why the inventory grep above matters more than any migration guide.

What's the difference between "deprecated" and "retired"? On Anthropic's lifecycle: Active means fully supported; Deprecated means still functional but no longer recommended, with a named replacement and a retirement date assigned; Retired means requests fail. OpenAI uses the same shape with its own dates. The deprecation period is your migration window — free and unhurried — and it is always shorter than it looks.

Does running through an AI gateway like tokens.bd protect me from retirements? It doesn't change the provider's retirement date — nobody's gateway can keep serving a model the provider switched off. What it changes is the blast radius: stable model IDs and one config to change instead of a model string scattered through five services, plus taka-denominated pricing so a December price doubling doesn't come with a currency surprise. Check the catalog for current availability.

Why does the Argon intro-price deadline matter if I don't use Argon? Because it sets the pattern you'll budget against all year. Introductory pricing — cheap launch rates with a fixed expiry — is now how frontier models launch: Argon's ends November 30, Gemini 3.8 Flash's is expected to double on January 1, 2027. Any annual budget built on a launch price is wrong from December on.

How much notice do providers actually owe me? It varies and it's getting shorter. Anthropic commits to at least 60 days for publicly released models (Sonnet 4.5 got 61). OpenAI gave roughly six months for the November 30 platform shutdowns but has retired product-surface models with far less. Google publishes model dates on its Gemini API deprecations page. Treat the notice period as a product feature to compare — and never assume one provider's date applies on another's platform.

Deadlines and prices verified 16 October 2026. Retirement dates apply to the provider's own platform unless stated otherwise — check Bedrock, Google Cloud, and Azure separately if you call models through them. The authoritative sources are each provider's deprecations page, linked below.

Sources: Anthropic model deprecations · End of Life — Claude Sonnet 4.5 retirement · Lindy — Gemini 4 Argon pricing · DataNorth — Gemini 4 Argon release · TechFeatured — Argon price doubling · GrN.dk — OpenAI retiring Agent Builder · gpt5-course deprecations research · Digital Matters — custom GPT retirement · dev.to — Gemini Gems become Skills · The Outpost — Google AI Plus loses Pro models · dev.to — 2026 shutdown calendar · CosmicJS — Sonnet 5.5 vs Sonnet 5 pricing · USD/BDT rate