ai-modelsmodel-retirementmodel-managementai-costsapi

AI model retirement dates: the calendar to July 2027

November 19, 20269 min readPixel Management

This article is also available in Dutch

On 1 December 2026, an application calling gemini-2.5-pro gets a different model back, without anyone touching the code. The same day, the price of Gemini 4 Argon doubles.

An AI model retirement is the fixed end of a model version: the provider names a date, and after it your call either fails or lands on another model. Every provider keeps its own list, and if you work through Azure or Bedrock you already have two. This calendar puts the dates of seven providers in one place.

AI model retirement dates: the calendar to July 2027

These AI model retirement dates and price changes were checked on 6 November 2026 against the pages of OpenAI, Anthropic, Google, Microsoft, Amazon, Mistral and Moonshot. Anything before 20 November is left out, as are most image, preview and cyber variants. Updated on 12 November with the Gemini 4 Carbon price change and the new gpt-image-2 shutdown date.

Two caveats. Google calls its dates "earliest possible dates". And the Azure list is dated 21 September 2026, so confirm those dates in your own portal.

DatePlatformWhat stops or changesReplacement or effect
24 Nov 2026Azure Foundryclaude-opus-4-5No replacement listed
30 Nov 2026Claude APIclaude-sonnet-4-5-20250929claude-sonnet-5-5
30 Nov 2026Claude APIFast mode on Opus 4.8Returns an error, fast mode needs Opus 5.5
30 Nov 2026OpenAIEvals, reusable prompts, Agent BuilderPromptfoo, your own code, Agents SDK
1 Dec 2026Gemini APIgemini-4-argon, price doubles$2 / $10 becomes $4 / $20
1 Dec 2026Gemini APIgemini-4-carbon, introductory rate ends$2 / $10 becomes $4 / $20
1 Dec 2026Gemini APIgemini-2.5-proRoutes automatically to gemini-3.8-flash
1 Dec 2026Gemini APIgemini-2.5-flashRoutes automatically to gemini-3.5-flash-lite
1 Dec 2026OpenAIgpt-image-1-mini, gpt-image-1.5gpt-image-2.6-sunburst or -flare
3 Dec 2026Gemini APIgemini-omni-1.1-flashRoutes automatically to gemini-omni-1.2-flash
11 Dec 2026OpenAIgpt-5, gpt-5-mini, gpt-5-nano, gpt-5-pro, o3, o3-progpt-5.6-sol, -terra or -luna
15 Dec 2026Azurewhisper, tts, tts-hdNo replacement listed
End Dec 2026MoonshotKimi K2 seriesClosed to new users since 31 Oct
1 Jan 2027Gemini API3.8 Flash and 3.6 Flash, price doubles$0.75 / $3.75 becomes $1.50 / $7.50
6 Jan 2027OpenAItts-1, tts-1-hd, gpt-4o-mini-tts, new self-serve fine-tuninggpt-realtime-2.1-mini
8 Jan 2027BedrockClaude Opus 4.1 ends, Sonnet 4.5 enters extended accessHigher price expected for Sonnet 4.5
14 Jan 2027OpenAISnapshots of gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-minigpt-6-luna, fine-tunes stop too
14 Jan 2027Azuregpt-4o version 2024-05-13gpt-5.1
20 Jan 2027OpenAIOlder realtime and audio modelsgpt-realtime-2.1, gpt-audio-1.5
10 Feb 2027OpenAIgpt-image-2, gpt-image-2-2026-04-21gpt-image-2.6-sunburst or -flare
26 Feb 2027OpenAIwhisper-1, gpt-4o-transcribe familygpt-transcribe, gpt-live-transcribe
1 Apr 2027OpenAIgpt-5.1, gpt-5.3-codex, gpt-5.4-nanogpt-6-sol or gpt-6-luna
8 Apr 2027BedrockClaude Sonnet 4.5No automatic migration
14 Apr 2027Azuregpt-4o 2024-08-06 and 2024-11-20, gpt-4o-mini, gpt-4.1 familygpt-5.1 for gpt-4o
30 Apr 2027MistralLarge 3, Medium 3.5, Voxtral TTS, Voxtral Mini Transcribe 2Large 4, Voxtral TTS 3, Transcribe 3
13 Jul 2027Azuregpt-5.1No replacement listed yet

Same model name, a different end date per platform

An AI model has one end date at its maker and another on a cloud platform, and the nearer one isn't always the maker's. Amazon says so itself: for Bedrock, only the dates on the Bedrock page apply, even when the model maker publishes something else. Anthropic says the same about Bedrock and about the Google-managed option on Google Cloud.

ModelAt the makerOn the cloud platformGap
GPT-4o (2024-08-06, 2024-11-20)OpenAI: 14 Jan 2027Azure: 14 Apr 2027Azure 90 days later
GPT-5 (2025-08-07)OpenAI: 11 Dec 2026Azure: 9 Feb 2027Azure 60 days later
o3OpenAI: 11 Dec 2026Azure: 19 Nov 2026 (passed)Azure 22 days earlier
WhisperOpenAI: 26 Feb 2027Azure: 15 Dec 2026Azure 73 days earlier
Claude Sonnet 4.5Anthropic: 30 Nov 2026Bedrock: 8 Apr 2027Bedrock 129 days later, higher price expected
Gemini 2.5 ProGemini API: routed on 1 Dec 2026Google Cloud: 20 Oct 2026 (passed)Same maker, different rules

GPT-4o is the best-known case. If you run directly on OpenAI, you lose the 2024-08-06 and 2024-11-20 snapshots on 14 January 2027, while the same versions keep running on Azure until 14 April. Check your version, though: gpt-4o 2024-05-13 also ends on Azure on 14 January.

The real trap is the replacement. Microsoft names gpt-5.1 as the successor to GPT-4o, yet retires gpt-5.1 itself on 13 July 2027, 90 days after GPT-4o's end date. OpenAI points the same models at gpt-6-luna. Under Microsoft's lifecycle policy, the official replacement is chosen only 90 to 120 days before retirement, which for 14 April means between mid-December and mid-January. So the named successor can still change. Don't wait for it: pick a model with a longer runway yourself.

Claude Sonnet 4.5 stays on Bedrock for longer. That time has a price. From 8 January 2027 the model is in public extended access, a phase in which Amazon says to expect higher pricing, set by the model maker. And in the Legacy phase, where Sonnet 4.5 has been since 8 October, existing customers may lose access after 15 days of inactivity. A quarterly report that runs once every three months walks straight into that.

At Google, the rules differ within one company. The Gemini API routes gemini-2.5-pro onwards on 1 December. The Google Cloud model page, where Vertex AI now goes by Gemini Enterprise Agent Platform, gives 20 October 2026 as the retirement date for the same model and says nothing about routing. If you run Gemini through Google Cloud, check in your own project what happens to those calls.

What breaks when an AI model is retired?

A retired AI model breaks in four ways, and only the first is visible straight away.

The call fails

This is the tidy version. At Anthropic, requests to a retired model fail, Azure returns 410 Gone, at OpenAI the model is no longer accessible, and on Bedrock migration doesn't happen automatically either, so your process stands still until someone changes the model name. At least your monitoring sees it at once.

A different model answers

From 1 December, Google sends all requests for gemini-2.5-pro to gemini-3.8-flash, at that model's rates. Azure does something similar: Global Standard and Data Zone Standard deployments of gpt-4o 2024-05-13 move to gpt-5.1 automatically at retirement. Provisioned deployments don't. You move those yourself. On Standard deployments you control this with versionUpgradeOption, and with NoAutoUpgrade the deployment stops at retirement.

No code change doesn't mean everything keeps working. In its changelog of 2 November, Google lists breaking changes itself: function calling and thought_signature validation, thinking_level replacing thinking_budget, and removed sampling parameters and prefilled model turns. The migration guide tells you to remove temperature, top_p and top_k, but not what happens if you keep sending them. Test it yourself. An AI agent that calls tools is the most exposed, because function calling is on that list.

A parameter disappears

Anthropic removed temperature, top_p and top_k in API version 2026-10-21. They now return a 400 error, and the new SDKs, such as Python 2.0.0 and TypeScript 1.0.0, send that version by default. On earlier API versions, a non-default value returns the same error on Claude 4.7 and later.

So if you replace Sonnet 4.5 with claude-sonnet-5-5 on 30 November and leave temperature: 0.2 in place, you've swapped a retired model for an error. Anthropic and Google are removing the same control. Consistency now comes from your prompt and your test set.

The tooling around it stops

OpenAI closes three products on 30 November that aren't models. Evals has been read-only since 31 October, reusable prompts (v1/prompts) have to move back into your own code, and Agent Builder shuts down. If your test set lives in Evals, you lose it in the very weeks you need it to judge the successor to GPT-5, because gpt-5 itself stops on 11 December. OpenAI points to Promptfoo.

Fine-tunes are the quiet version of this. At OpenAI they follow the schedule of their base model, so a fine-tune on gpt-4o-2024-08-06 stops on 14 January 2027. From 6 January, existing customers also can't start new self-serve fine-tuning jobs, so you won't be redoing that fine-tune in the final week. On Azure, such a deployment keeps running until 1 October 2027.

How to pin a version, run a test set and keep a way back is covered in running AI agents in production. What a test set is worth is in evaluating AI benchmarks.

What does the 1 December price change cost?

According to the Gemini API pricing page, Gemini 4 Argon goes from $2 / $10 to $4 / $20 per million tokens on 1 December. Gemini 4 Carbon, available to all paying businesses since 10 November, loses its introductory rate the same day, with the same doubling. On 1 January 2027, Gemini 3.8 Flash and 3.6 Flash go from $0.75 / $3.75 to $1.50 / $7.50. If you're routed over from Gemini 2.5 Pro, that means two rate changes in 31 days.

A worked example, indicative only. Assumptions: 500 million input tokens and 100 million output tokens a month, standard rate, no caching, prompts under 200,000 tokens, and the same token use on every model.

Monthly billNovember 2026December 2026January 2027
Gemini 4 Argon$2,000$4,000$4,000
Gemini 3.8 Flash$750$750$1,500
Gemini 2.5 Pro, routed to 3.8 Flash$1,625$750$1,500

The sum for Argon in December: 500 × $4 + 100 × $20 = $4,000. That's $2,000 a month more, or $24,000 a year, at equal usage.

The 2.5 Pro routing looks like a windfall, but that depends on your ratio of input to output. From January, 3.8 Flash costs $0.25 more per million input tokens than 2.5 Pro and $2.50 less per million output tokens. So the break-even sits at ten times as much input as output, and above that, 3.8 Flash costs more than the model you had. A knowledge-base application that packs document text into the prompt and returns short answers can easily land there. The example above runs at five to one.

Gemini 2.5 Flash and 3.5 Flash-Lite both cost $0.30 / $2.50 for text. Your bill stays the same. Your model doesn't.

The weakest assumption in the table is equal token use. Google's output figure includes thinking tokens, and those differ per model. So measure on your own traffic, as described in comparing AI models on cost per task. These providers also invoice in dollars, so for a Dutch company the euro exchange rate moves the bill as well.

How much warning do AI providers give?

AI providers' notice periods range from six months to no commitment at all, so don't wait for an email.

ProviderMinimum noticeOn the end dateHow to check
OpenAI6 months, specialised variants 3 months, preview sometimes 2 weeksNo longer accessibleShutdown date in the List models endpoint
Anthropic60 daysRequests failUsage export by API key and model
Gemini APINo period statedOff, or routedDeprecations page and changelog
Azure Foundry60 days, preview 30 days410 Gone, or automatic upgradeModels API, field deprecation.inference
Bedrock6 months Legacy for models launched before 7 September 2026Requests failField modelLifecycle in the API

Google put the 2.5 Pro routing in its changelog on 2 November, 29 days before it takes effect, and a fixed period like OpenAI's or Microsoft's appears nowhere on the page. On Azure, extensions don't exist: "Retirement dates aren't extendable." Watch the wording in the Azure Models API too, because it differs from the docs: Deprecating there means the model still runs, Deprecated means it's gone.

"Not sooner than" on Anthropic's page is a floor, not an end date. On 6 November, claude-opus-4-5-20251101 was listed there with 24 November 2026 as that floor, but still as Active and without a deprecation notice. Microsoft lists 24 November as the retirement date for the same version on Foundry. If you use Claude through Foundry, plan for the stricter of the two.

If you buy AI inside a software package, these dates reach you through your vendor. Ask which model runs underneath, who tests the migration and who pays a higher token price. Those terms belong in the contract, see buying AI software.

Which end dates can you ignore?

Not every line in this calendar means work, and a rushed migration is a risk of its own.

Subscriptions. These dates cover APIs and cloud platforms. If you only use ChatGPT or Microsoft 365 Copilot as a product, the vendor picks the model and there's nothing to migrate. At most you'll notice the behaviour shift.

GPT-4o on Azure. There's no rush until 14 April, provided you've checked the version. Move once. Skip gpt-5.1.

Self-hosted open weights. On your own server, nothing changes unless you change it. Mistral retires its API models on 30 April 2027, and that says nothing about weights you run yourself. One end date is no reason to build your own environment, though: moving to the successor is nearly always less work than the full trade-off in local AI vs cloud AI.

The recurring task is small: one owner who goes through these lists monthly for the models you have in production. If you want to switch models more often, keep the model choice in one place in the code, the approach from avoiding vendor lock-in.

Service

Learn more about AI agents?

View service →

Curious how much time you could save?

Request a free efficiency audit. We'll analyze your processes and show you where the gains are, no strings attached.