OpenAI shuts down GPT-4o and GPT-4.1 on January 14, three months before Azure

On October 14, 2026, OpenAI set a date of January 14, 2027 to discontinue GPT-4o, GPT-4o mini, GPT-4.1, and GPT-4.1 mini in its API. This includes the August and November 2024 versions of GPT-4o; the May 2024 version has been cut since October 23. On Microsoft Foundry, the Azure platform, the withdrawal schedule keeps these same versions until April 14, 2027: three more months. With o3, it’s the opposite: Azure withdraws it on November 19, 22 days before OpenAI. A company that only follows the publisher’s page can therefore get the date wrong in both directions: ignore three months of margin on GPT-4o, or lose o3 three weeks earlier than expected. Microsoft’s page is marked “updated September 23,” predating OpenAI’s October 14 announcement.

On October 28, ActuIA encountered the depreciation pages of OpenAI, Anthropic, Microsoft Foundry and Amazon Bedrock. For the same model, the difference noted ranges from 22 days to 173 days, almost six months. The rules displayed by the platforms do not always correspond to the published dates: three deviations are detailed below.

Six deadlines in November

The first change doesn’t seem to break anything. Starting November 2, xAI serves requests to grok-imagine-image-quality with another model, grok-imagine-image-2.0, set to low quality, according to its migration guide. xAI announced the change 60 days in advance, on September 2. The identifier continues to respond and the code does not have to change, xAI writes. Each image costs $0.01 less. The field model of the response indicates the model that served the request.

On November 13, Bedrock withdraws Claude 3 Haiku, Claude 3 Sonnet, Claude 3.5 Sonnet, its version 2 and Claude 3.7 Sonnet in part of its regions. In Europe, the page cites London (eu-west-2, Claude 3 Sonnet and 3.7 Sonnet) and Zurich (eu-central-2, Claude 3 Haiku). In the Paris region (eu-west-3), Claude 3 Haiku has already reached its end of life on September 10.

On Foundry, on November 15, codex-mini and Claude Haiku 4.5 were removed, with no replacement indicated. On November 19, the platform shut down o1, o1-pro, o3, o3-pro, o3-deep-research, o3-mini and o4-mini, with gpt-5.6-sol and gpt-5.6-terra as replacements. On November 24, she withdrew Claude Opus 4.5.

On November 30, Claude Sonnet 4.5 became unresponsive on Foundry and in the Anthropic API, which listed this date as the “attempted” removal date. The same day, OpenAI closes Agent Builder, its agent design tool, the Evals platform and the v1/prompts API which stores reusable prompts. Existing reviews become read-only on October 31. OpenAI links to its Agents SDK or ChatGPT workspace agents for Agent Builder, and to Promptfoo for reviews.

Up to 173 days apart for the same model

The table below reconciles the dates published by each platform, noted as October 28, 2026. “Absent” means that the model does not appear on the page of the platform concerned; “not noted” that the page was not read for this model. The Microsoft page does not list any Claude 3.

The refined models do not follow the same shift. At OpenAI, a refined version stops with its base model. On Azure, refined GPT-4o and GPT-4.1 deployments remain served until October 1 or 14, 2027, nine months after the OpenAI date for GPT-4.1, and training of new versions remains possible, for existing customers, “not before” April 1 or 14, 2027. OpenAI will also stop accepting new refinements from its existing customers on January 6, 2027.

On the day of withdrawal, the call fails, switches or costs more

At Anthropic, requests to a retired model fail, the page says. At OpenAI, the page states that access to the October 23 models “was cut off” and does not describe the form of the failure. On Foundry, Microsoft’s lifecycle policy states that any request returns a 410 error after removal. The dates “are not extendable”, specifies the page (translated from English). OpenAI, for its part, indicates that dedicated capacity for extended access can be offered by contacting its sales teams; the page gives neither the conditions nor the price.

Foundry also changes the model for some customers without asking them. Standard, Global Standard, and Data Zone Standard deployments are automatically upgraded upon retirement on a region-by-region basis, including to a version not yet separately available in the region. The page does not designate replacements for several models on the list, including Claude Haiku 4.5, codex-mini and GPT-4.1. Provisioned capacity deployments are not upgraded: their customers must migrate themselves, either “in place” in 20 to 30 minutes without interruption according to the Microsoft FAQ, or side by side. A team that has chosen the NoAutoUpgrade option will see its deployment stop on the day of removal.

Bedrock extends access, at a price which may increase. A model first moves to Legacy status, then to public extended access until its end of life. During this period, AWS indicates that you should expect a higher price, set by the model provider: the page predicts price changes “if any”. The lifecycle page does not give any pricing for Claude Sonnet 4.5. Two rules limit this extension. No new customers can use a Legacy model. An existing customer may lose access after 15 days of inactivity.

Three discrepancies between posted rules and dates

Anthropic’s page states that its dates apply to the Claude API and Microsoft Foundry. Only Bedrock and the Google-managed version of Google Cloud set their own schedule. It promises at least 60 days notice to customers with active deployments. Foundry retires Claude Haiku 4.5 on November 15 and Claude Opus 4.5 on November 24. As of October 28, Anthropic lists both models as active, with “not before” retirement dates of October 15 and November 24, and has not announced the depreciation of either. Foundry dates meet these floors. With 18 days until the deadline for Haiku 4.5, no 60 day notice is visible on the Anthropic page.

Bedrock writes that a model remains “at least six months” in Legacy status before its end of life. In the regions affected by the withdrawal of November 13, the same page enters the Claude 3, 3.5 and 3.7 into Legacy on October 10, 13 or 15, 2026. It therefore leaves them 34 days at most, and these lines open extended access on the same day, while the page sets the minimum in Legacy at three months before this access. The page doesn’t give an explanation.

OpenAI writes that general availability models receive at least six months’ notice, and at least three months for specialized variants (cat, Codex, deep research). The page also provides for faster withdrawals for security or compliance reasons. GPT-4o and GPT-4.1 were given 92 days, from October 14 to January 14. The American site The New Stack noted on October 15 this gap with the six months promised to developers. The OpenAI page does not justify this shorter deadline.

The replacement does not always adjust like the old model

Changing your identifier is not always enough. In Anthropic API version 2026-10-21, the temperature, top_p and top_k parameters are removed: a query that contains them in this version receives a 400 error. Earlier versions of the API already reject them with Claude 4.7 and later models, as soon as they deviate from their default value. The replacement designated for Claude Sonnet 4.5, Claude Sonnet 5.5, is later than Claude 4.7. An integration that fixes the Sonnet 4.5 temperature will need to remove this setting when migrating.

On Bedrock, the state of a model can also change from one region to another. AWS refers to its ListFoundationModels and GetFoundationModel APIs, whose modelLifecycle field gives the situation region by region.

After November, the next deadline is January 6, 2027, end of new refinements at OpenAI, then January 14: GPT-4o and GPT-4.1 are cut, with gpt-6-luna as a replacement. On March 30, 2027, Bedrock is retiring 26 models, including Llama 3 and 4, Mistral 7B, Mixtral 8x7B, Devstral 2, DeepSeek-R1 and GLM 4.7. Pricing schedules for replacement models are collected in the October 2026 AI API Price Statement.