On June 29, 2026, OpenAI retired four separate GPT chat-tuned model snapshots from Microsoft’s Azure AI Foundry in a single move, folding gpt-5-chat, gpt-5.1-chat, gpt-5.2-chat, and gpt-5.3-chat into one rolling alias called gpt-chat-latest. It’s the fourth distinct deprecation notice OpenAI has pushed through in June alone, following separate waves on June 2, June 3, and June 11, and it landed just two days after OpenAI quietly retired GPT-4.5 from ChatGPT itself on June 27.
The timing exposes an odd split between how fast AI companies retire models and how slowly the market actually stops using them. New research from cloud security firm Orca puts GPT-4o, a model OpenAI pulled from the ChatGPT consumer app back in February, inside 37.6% of AI-adopting cloud environments, more than any newer OpenAI release. That gap between announced retirement and real-world usage has become one of the more expensive planning problems for enterprise IT teams in 2026, part of a broader shift in how the industry chooses and retires AI models this year.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Happened on June 29: Four Chat Models Retire at Once
Microsoft’s Foundry retirement schedule, the reference document Azure customers use to plan migrations, lists four “chat” variants as retired effective June 29, 2026: gpt-5-chat (the 2025-08-07 snapshot), gpt-5.1-chat (2025-11-13), gpt-5.2-chat (2026-02-10), and gpt-5.3-chat (2026-03-03). All four point to the same replacement, gpt-chat-latest, a rolling alias rather than a fixed model version.
It’s the second time this year OpenAI has retired a chat snapshot on Azure. An earlier gpt-5.2-chat build (2025-12-11) and a gpt-5-chat build (2025-10-03) were already pulled on May 13, 2026. Between the two rounds, five pinned “chat” snapshots have gone dark in seven weeks, each replaced by the same moving target.
The pattern matters because “chat” variants are the conversational, personality-tuned counterparts to OpenAI’s core reasoning models (gpt-5, gpt-5.1, gpt-5.2, and later releases), which sit on far longer retirement timelines, some running into 2027. Developers who pinned a chat snapshot in production code rather than tracking the rolling alias are the ones scrambling today.
GPT-4.5’s ChatGPT Exit Lands Two Days Before the Azure Cut
The June 29 Azure purge wasn’t the only retirement that week. Two days earlier, on June 27, 2026, OpenAI’s Help Center confirmed that GPT-4.5 had been retired from ChatGPT itself, closing out the last GPT-4-generation model still available in the consumer app.
That timing matters more than a coincidence of dates. When OpenAI retired GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT on February 13, 2026, GPT-4.5 was the one GPT-4-series model left standing. It survived the February cull by a little over four months, only to be pulled the same week Microsoft’s Azure Foundry retired four separate “chat” variants of the GPT-5 line. Between the two dates, every GPT-4-generation model named in OpenAI’s 2026 retirement notices has now left ChatGPT: GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini in February, and GPT-4.5 in June.
OpenAI hasn’t published a detailed usage breakdown for GPT-4.5 the way it did for GPT-4o, whose daily active share had fallen to just 0.1% by the time of its cutoff (see below). But GPT-4.5’s exit fits the same pattern documented throughout this article: OpenAI keeps retiring legacy models on a compressed timeline once usage concentrates on a newer default, whether or not the retirement draws public attention. GPT-4o’s February exit generated headlines because of how many workflows had been built around it. GPT-4.5’s June exit barely registered, a sign that whatever consumer base it still had was already small by the time OpenAI pulled it.
For developers, the practical fallout follows the same playbook this article already lays out for GPT-4o, GPT-4.1, and o4-mini: any workflow still calling gpt-4.5 directly, whether through the ChatGPT interface or a pinned API model ID, needs to move to a current default now rather than after the next notice lands. The updated retirement timeline below reflects both events. It’s a reminder that ChatGPT consumer retirements, Azure API retirements, and platform-feature shutdowns each run on their own calendar, and a team tracking only one of those calendars can still be caught off guard by a change on another.
That leaves five GPT-4-generation models fully retired from ChatGPT this year: GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini in February, and GPT-4.5 in June. Retiring an entire prior model generation from the flagship consumer product in a little over four months is a much faster clock than the multi-year gaps between model generations that were typical just a few years earlier. Whether GPT-4.5’s retirement followed the same six-month notice window OpenAI applies to generally available models, or the shorter three-month window reserved for specialized variants, isn’t stated in the Help Center update. What is clear is that its removal, combined with the June 29 Azure changes, means the last week of June 2026 did more to simplify OpenAI’s active ChatGPT lineup than any single stretch since the February consumer purge.
The Full OpenAI Model Retirement Timeline for 2026
Zoom out and June 29 is just the latest entry in a year-long sequence. Below is the retirement and deprecation calendar OpenAI and Microsoft have published so far for 2026, drawn from OpenAI’s developer deprecations page and Microsoft’s Foundry model retirement schedule.
| Date | What Retired or Was Announced | Platform | Replacement |
|---|---|---|---|
| 2026-01-29 | GPT-4o mini replaced in the model picker for paid users | ChatGPT | GPT-4.1 mini |
| 2026-02-13 | GPT-4o, GPT-4.1, GPT-4.1 mini, o4-mini retired from consumer app | ChatGPT | GPT-5.1 / GPT-5.2 family |
| 2026-03-11 | GPT-5.1 Instant, Thinking, and Pro retired | ChatGPT | GPT-5.2 family |
| 2026-04-03 | GPT-4o access inside Custom GPTs ends for Business, Enterprise, Edu | ChatGPT | N/A |
| 2026-05-13 | gpt-5-chat (2025-10-03) and gpt-5.2-chat (2025-12-11) retired | Azure AI Foundry | gpt-chat-latest |
| 2026-06-02 | Legacy GPT Image models deprecated | OpenAI API | Removal Dec 1, 2026 |
| 2026-06-03 | Reusable prompts, Evals platform, Agent Builder deprecated | OpenAI platform | Shutdown Nov 30, 2026 |
| 2026-06-11 | Older GPT-5 and o3 snapshots given removal notice | OpenAI API | gpt-5.6-sol (removal Dec 11, 2026) |
| 2026-06-27 | GPT-4.5 retired from consumer app | ChatGPT | GPT-5.x family |
| 2026-06-29 | gpt-5-chat, gpt-5.1-chat, gpt-5.2-chat, gpt-5.3-chat retired | Azure AI Foundry | gpt-chat-latest |
| 2026-10-01 | o3-mini retires | Azure AI Foundry | o4-mini |
| 2026-10-14 | gpt-4.1-nano retires | Azure AI Foundry | N/A |
| 2026-10-21 | o1, o1-pro, o3 retire | Azure AI Foundry | gpt-5.6-sol |
| 2026-10-23 | gpt-4o (2024-05-13 build) retires | Azure AI Foundry | gpt-5.1 |
Read across the year and the cadence is hard to miss. OpenAI has touched some part of its model lineup, consumer app defaults, API snapshots, or adjacent platform features, in at least eight separate months through the end of June. Few enterprise software vendors retire product versions this often.
Why GPT-4o Still Leads Enterprise Adoption at 37.6%
None of that retirement pressure has fully caught up with actual usage. Orca’s cloud security researchers examined 426 organizations running at least one AI model in production and found GPT-4o present in 37.6% of them, the single most-deployed model in the study, ahead of every model OpenAI has released since.
| Rank | Model | Share of AI-Adopting Organizations (of 426 studied) |
|---|---|---|
| 1 | GPT-4o | 37.6% |
| 2 | GPT-4.1 mini | 34.7% |
| 3 | text-embedding-ada-002 | 29.1% |
| 4 | GPT-4.1 | 28.9% |
| 5 | GPT-4o mini | 28.6% |
| 6 | GPT-5.1 | 22.3% |
| 7 | GPT-5 mini | 19.7% |
| 8 | o4-mini | 17.8% |
GPT-4.1 mini, the model OpenAI is nudging free-tier ChatGPT users toward, sits second at 34.7%. The newest generally-available flagship in the mix, GPT-5.1, trails at 22.3%, a reminder that enterprise migration typically lags a consumer rollout by months or years, not weeks. Code written against GPT-4o’s API behavior, prompt templates tuned to its response style, and internal evaluation pipelines built around its output format all cost real engineering hours to move, regardless of what OpenAI’s retirement calendar says.
Inside OpenAI’s Deprecation Policy: 6 Months, 3 Months, 2 Weeks
OpenAI’s published deprecation policy sets three notice tiers. Generally available models get at least six months’ warning before retirement. Specialized variants, the company’s own examples include chat-tuned builds like gpt-5.1-chat-latest, Codex variants, and deep research models, get at least three months. Preview models, flagged by the word “preview” in their name, can be retired with as little as two weeks’ notice. OpenAI reserves the right to move faster than any of these windows if safety or compliance concerns demand it, though it says it aims to give as much notice as reasonably possible even then.
The June 11 notice fits the policy almost exactly. OpenAI told developers that three snapshots, gpt-5-2025-08-07, gpt-5-mini-2025-08-07, and o3-2025-04-16, would be removed on December 11, six months to the day later, with gpt-5.6-sol named as the intended replacement. It’s a useful data point for any team trying to predict OpenAI’s next move: count six months forward from a GA model’s launch, and three months forward from a specialized variant’s, and that’s a rough retirement floor.
How the February ChatGPT Consumer Retirement Played Out
The Azure Foundry retirements getting attention today are really the second act of a story that started in the consumer app. On February 13, 2026, OpenAI’s help center confirmed the retirement of GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT itself, following a smaller change on January 29 that had already swapped GPT-4.1 mini in as the fallback model for free users who hit their GPT-4o limits.
Business, Enterprise, and Edu customers got a longer runway. GPT-4o remained available inside Custom GPTs for those plans until April 3, 2026, giving companies with internal GPT-based tools roughly seven extra weeks to migrate before it disappeared everywhere.
What stands out is how few people actually noticed. OpenAI itself said the vast majority of ChatGPT usage had already shifted to GPT-5.2 by the time of the cutoff, with just 0.1% of daily active users still choosing GPT-4o, a figure Dutch-Belgian tech outlet Techzine also reported when it covered the retirement. That’s a very different picture from the Orca cloud data above, where GPT-4o remains the single most common model in production systems. Consumers stopped noticing GPT-4o months before enterprises stopped depending on it.
Azure AI Foundry Becomes the De Facto Enterprise Retirement Clock
If ChatGPT’s consumer app runs on a fast clock, Azure AI Foundry runs on a much slower, tiered one, and that difference is now doing a lot of the work of keeping enterprise customers calm. Microsoft’s retirement schedule splits every model into two dates: a training retirement date, after which a business can no longer create new fine-tunes, and a later deployment retirement date, after which existing fine-tuned deployments finally stop serving traffic.
For gpt-4o and gpt-4o-mini, that deployment retirement date for existing customers stretches to October 1, 2027, roughly a year past gpt-4o’s own October 23, 2026 base-model sunset. The gpt-4.1 family gets until October 14, 2027, and o4-mini until October 16, 2027. In effect, Microsoft is selling enterprise customers extra time that OpenAI’s own consumer roadmap does not offer directly, and that gap has become part of Azure’s pitch to risk-averse IT departments choosing between Azure AI Foundry, AWS Bedrock, and Google Vertex AI as their model hosting layer.
The Hidden Cost of Migration: Latency, Pricing, and Rewrites
Retirement notices measure time. They don’t measure the engineering cost of actually moving off a model, and that cost is where most of 2026’s migration pain lives.
Chat Completions vs. the Responses API
Independent technical analysis from AI infrastructure firm TensorOps found that shifting off GPT-4.1 typically requires more than swapping a model name in a config file. It means migrating from OpenAI’s older Chat Completions API to the newer Responses API, a change the firm describes as forcing substantial application rewrites across request handling, streaming, and tool-calling code.
// Legacy: Chat Completions API (GPT-4.1 and earlier)
const completion = await openai.chat.completions.create({
model: "gpt-4.1",
messages: [{ role: "user", content: prompt }],
});
const text = completion.choices[0].message.content;
// Current: Responses API (GPT-5.x family)
const response = await openai.responses.create({
model: "gpt-5.1",
input: prompt,
});
const text = response.output_text;The Reasoning Tax: Why Successor Models Cost More
The economics shift too. TensorOps clocked GPT-4.1, a non-reasoning model, at roughly 1.35 seconds of average latency against $2 per million input tokens and $8 per million output tokens. The reasoning-focused models OpenAI is steering customers toward average 4.26 seconds of latency and run about $10 per million output tokens, a meaningfully higher bill for workloads that don’t actually need step-by-step reasoning.
The firm’s recommended fix is traffic routing rather than a wholesale switch: send the roughly 80% of enterprise requests that are simple lookups or classifications to cheap, fast models, and reserve the expensive reasoning tier for genuinely hard tasks. Some teams are hedging further by self-hosting open-weight alternatives like Llama 4 Maverick, or routing part of their workload through AWS Bedrock to avoid being fully dependent on any one vendor’s retirement calendar.
It’s Not Just OpenAI: Grok, DeepSeek, Llama, and Claude Face the Same Clock
Model churn isn’t a uniquely OpenAI phenomenon this year, it’s an industry-wide operating condition. Microsoft’s Foundry marketplace, which hosts models from multiple vendors alongside OpenAI’s own, shows nearly every major lab running its own retirement schedule in parallel.
xAI retired grok-3, grok-3-mini, and two grok-4-fast variants on May 1, 2026, pushing customers toward grok-4-1-fast-reasoning and its non-reasoning counterpart. DeepSeek has three models, DeepSeek-R1-0528, DeepSeek-V3-0324, and DeepSeek-V3.1, scheduled to retire on July 13, 2026, in favor of the newer DeepSeek-V4-Flash and DeepSeek-V4-Pro. Meta pulled five Llama 3.1 and 3.2 vision and instruct models on June 13, 2026, just sixteen days before OpenAI’s latest cut. Even Anthropic’s models on Azure’s marketplace carry retirement dates, claude-opus-4-1 is due to retire on August 5, 2026 in favor of claude-opus-4-8, with claude-opus-4-5 and claude-sonnet-4-5 following on October 19.
The common thread: every lab is compressing the distance between generally available and legacy as release cycles speed up. A model that felt current in January 2026 can carry a deprecation notice by summer.
Enterprise Adoption Beyond OpenAI: What the Futurum Survey Shows
The reason so many organizations feel this pain at once is that few of them bet on a single vendor to begin with. Futurum Group’s AI Platforms Decision Maker Survey, which polled 820 people responsible for AI purchasing decisions, found OpenAI leading model adoption at 57%, with Azure OpenAI close behind at 56% and Google Gemini at 48%. Those figures overlap heavily since many organizations run more than one platform at once, and Futurum found 68% of respondents already at advanced stages of generative AI adoption rather than still piloting it.
That multi-vendor pattern is exactly why a single retirement notice, whether it’s OpenAI pulling GPT-4o or DeepSeek sunsetting R1, doesn’t fully derail most enterprise AI strategies anymore. It also means the same IT team can be managing three or four separate deprecation calendars simultaneously, which is its own kind of operational overhead that rarely shows up on a vendor’s pricing page.
Market Impact: What Accelerating Retirements Mean for Cloud Vendors
The faster foundation model vendors retire their own products, the more valuable the hosting layer underneath them becomes. Microsoft, AWS, and Google are all positioning their AI platforms as the stable buffer between a chaotic model release schedule and a business that can’t afford to rewrite its AI stack every quarter.
Azure’s extended deployment-retirement windows, up to a year beyond OpenAI’s own base-model sunset, are a direct product of that positioning. So is the growing interest in routing layers and multi-model gateways that let a company swap the model underneath an application without touching the application itself. Cloud FinOps teams are already tracking AI spend as its own budget line, and cloud waste tied to AI spend has become a measurable line item in its own right this year. Model retirement cadence is quickly becoming another input into that same cost-governance conversation, alongside compute waste and idle GPU capacity.
For OpenAI specifically, the pace of change cuts both ways. It keeps competitors from catching up on capability, but it also pushes some of the most price- and stability-sensitive enterprise customers toward vendors that promise a calmer upgrade path.
Historical Context: From GPT-3.5’s 79% Monopoly to Today’s Fragmentation
Orca’s data includes a comparison that puts the current 37.6% figure in perspective. In 2024, GPT-3.5 alone commanded 79% adoption among the same kind of cloud-deployed AI workloads, roughly double GPT-4o’s current lead. Orca’s researchers note plainly that no single model dominates the way GPT-3.5 once did.
That shift lines up with a broader consumer-side trend. ChatGPT’s own share of the consumer chatbot market has fallen to 46.4% as rivals gain ground, down from a period when it faced essentially no serious competition. Two years ago, choosing an AI model mostly meant choosing OpenAI. In 2026, it means choosing between OpenAI, Anthropic, Google, xAI, DeepSeek, and Meta, often several at once, and living with each of their separate release and retirement schedules.
The practical result is that model retirement has stopped being a rare, disruptive event and become a routine planning input, closer to how IT teams already treat operating system end-of-life dates or database version support windows.
Analysis: Why OpenAI’s Release Cadence Keeps Accelerating
Line up the GA dates from Microsoft’s own retirement schedule and the acceleration is easy to see. GPT-5 launched August 7, 2025. GPT-5.1 followed November 13. GPT-5.2 arrived December 11, barely a month later. GPT-5.3-chat was already live by March 3, 2026, GPT-5.4 landed two days after that, and GPT-5.5 followed April 24. By late June, OpenAI was previewing GPT-5.6 directly, with its Sol, Terra, and Luna variants set to reach general availability on Azure on July 9.
That’s seven major version bumps in eleven months, before counting the specialized chat, Codex, and audio variants layered on top of each. Competitive pressure from Google, Anthropic, and xAI, all shipping their own frequent updates, is the most obvious driver. But the Futurum adoption numbers suggest a second one: with three platforms all sitting within nine points of each other’s enterprise adoption share, no single lab can afford to let its lineup look dated for long. The cost of that pace shows up exactly where this article started, in retirement notices landing faster than most engineering teams can act on them.
5 Predictions for AI Model Lifecycles Through 2027
None of this cadence looks likely to slow down. Based on the patterns in OpenAI’s and Microsoft’s own published schedules, here’s where model lifecycles are headed next.
- Notice windows keep compressing. Expect specialized and chat-variant retirements to move from today’s roughly three-month notice toward six-to-eight-week windows by 2027, as OpenAI leans harder on rolling aliases like gpt-chat-latest instead of pinned snapshots.
- “Model exit” clauses become standard in AI vendor contracts. Enterprises caught by the February and June 2026 waves will start negotiating guaranteed minimum support windows the way they already do for database and OS end-of-life.
- Cloud marketplaces keep out-running the model makers on runway. Expect Azure, Bedrock, and Vertex to keep extending fine-tuned deployment windows a year or more past a model’s own retirement date, turning hosting choice into a bigger risk-management decision than model choice.
- Routing middleware becomes a real product category. The gap between a retired snapshot and its replacement is exactly the problem third-party model-routing and gateway tools are built to solve, and demand for them should grow alongside every new retirement wave.
- Multi-model procurement keeps rising. With OpenAI, Azure OpenAI, and Google Gemini already sitting within nine points of each other in Futurum’s survey, expect that gap to narrow further as CIOs treat single-vendor dependency as a retirement risk, not just a cost or capability one.
What Developers and IT Teams Should Do Now
A few habits separate the teams calmly reading this article from the ones scrambling through a retirement notice late on a Friday.
- Audit every pinned model ID in production code and replace dated snapshots with rolling aliases wherever the workload can tolerate silent upgrades.
- Bookmark both OpenAI’s deprecations page and Microsoft’s Foundry retirement schedule, and put the dates that affect your stack directly on an engineering calendar, not just in a vendor email.
- Budget for the reasoning-model cost and latency tax before migrating, not after, since a like-for-like model swap can raise output costs several times over.
- Test the Responses API migration path early if you’re still on Chat Completions, since that rewrite tends to take longer than the model swap itself.
- Treat multi-model support as a resilience feature, not just a cost optimization, given how many separate retirement clocks are now running at once across vendors.
Frequently Asked Questions
What exactly retired on June 29, 2026?
Four OpenAI chat-tuned model snapshots hosted on Microsoft’s Azure AI Foundry: gpt-5-chat (2025-08-07 build), gpt-5.1-chat, gpt-5.2-chat (2026-02-10 build), and gpt-5.3-chat. All four were replaced by a single rolling alias, gpt-chat-latest.
Did anything else retire the same week as the June 29 Azure changes?
Yes. OpenAI retired GPT-4.5 from ChatGPT on June 27, 2026, two days before the Azure chat-variant purge. It was the last GPT-4-generation model still available in ChatGPT after GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini were retired from the app on February 13, 2026.
Does GPT-4o still work today?
Yes, through the API and through Azure, though existing fine-tuned deployments are scheduled to retire between October 2026 and October 2027 depending on the specific version. It was removed from the ChatGPT consumer app on February 13, 2026.
Why does GPT-4o still show 37.6% enterprise adoption if it’s being retired?
Because cloud deployment lags consumer product changes by months or years. Orca Security’s analysis of 426 organizations found GPT-4o still the most common model in production AI systems, even though OpenAI stopped offering it to everyday ChatGPT users back in February.
How much notice does OpenAI give before retiring a model?
At least six months for generally available models, at least three months for specialized variants like chat-tuned builds, and as little as two weeks for anything still labeled preview.
Is this only happening to OpenAI models?
No. Azure’s marketplace schedule shows xAI, DeepSeek, Meta, and Anthropic models on comparable retirement timelines this year, including Grok-3 variants retired in May and several Llama 3.1/3.2 models retired in June.
What replaces the retired chat models?
gpt-chat-latest, a rolling alias that Microsoft and OpenAI update over time rather than a fixed, dated snapshot.
Will migrating to a newer model cost more?
Often yes for reasoning-focused successors. Independent testing from TensorOps found non-reasoning GPT-4.1 running about $2 and $8 per million input and output tokens versus roughly $10 per million output tokens and triple the latency for newer reasoning models.
What should I do if my app still calls a deprecated model ID directly?
Move to a rolling alias where your workload allows it, and if you’re still on the Chat Completions API, budget time to test the newer Responses API before your specific snapshot’s retirement date arrives.
Related Coverage
- ChatGPT Market Share Falls to 46.4% as Rivals Gain [2026]
- Claude Mythos 5 Blackout: GPT-5.6 Capped at 20 Firms [2026]
- Bedrock vs Azure AI Foundry vs Vertex AI: 17x Gap [2026]
- Best AI Model for Coding: DeepSeek Costs 20x Less [2026]
- Cloud Waste Hits 29% as AI Spend Breaks Budgets [2026]
- Opus 4.8 vs GPT-5.6 vs Gemini 3.1 Pro: $18 Price Gap [2026]
