OpenAI retires GPT-4o and GPT-4.1 on Jan. 14 – Azure and Bedrock differ

What this is about

On October 14, 2026, OpenAI announced that it will shut down GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini in its API on January 14, 2027. This affects the August and November 2024 versions of GPT-4o. The May 2024 version has been retired since October 23. On October 28, ActuIA compared the deprecation pages of OpenAI, Anthropic, Microsoft Foundry (the Azure platform) and Amazon Bedrock. The result: the same models stay available for different lengths of time depending on the platform, with differences of 22 to 173 days. If you only follow the vendor's own page, you can be wrong in either direction.

On Microsoft Foundry, GPT-4o and GPT-4.1 remain available until April 14, 2027. With o3 it's the other way round: Azure retires it on November 19, 22 days before OpenAI. The Microsoft page is marked "updated September 23," so it predates OpenAI's announcement.

The November deadlines

• November 2: xAI redirects requests for grok-imagine-image-quality to grok-imagine-image-2.0 at low quality (announced September 2). The ID keeps working, so you don't need to change your code, and each image costs $0.01 less. The model field in the response shows the model actually used.
• November 13: Bedrock retires Claude 3 Haiku, Claude 3 Sonnet, Claude 3.5 Sonnet (v2) and Claude 3.7 Sonnet in some regions. In Europe that includes London (Claude 3 Sonnet, 3.7 Sonnet) and Zurich (Claude 3 Haiku). In Paris, Claude 3 Haiku already ended on September 10.
• November 15: Foundry takes codex-mini and Claude Haiku 4.5 offline with no replacement named.
• November 19: Foundry ends o1, o1-pro, o3, o3-pro, o3-deep-research, o3-mini and o4-mini. The replacements are gpt-5.6-sol and gpt-5.6-terra.
• November 24: Foundry retires Claude Opus 4.5.
• November 30: Claude Sonnet 4.5 stops responding on Foundry and in Anthropic's API (listed there as a "tentative" date). The same day, OpenAI closes Agent Builder, the Evals platform and the v1/prompts API for reusable prompts. Existing evaluations become read-only from October 31. OpenAI points to the Agents SDK or ChatGPT workspace agents, and to Promptfoo for evaluations.

Up to 173 days of difference

Fine-tuned models follow a different pattern. At OpenAI, a fine-tuned version ends with its base model. On Azure, fine-tuned GPT-4o and GPT-4.1 deployments keep running until October 1 or 14, 2027, which for GPT-4.1 is nine months after OpenAI's date. New training is available to existing customers "no earlier than" April 1 or 14, 2027. OpenAI also stops accepting new fine-tuning jobs from existing customers from January 6, 2027.

What happens on the deadline

At Anthropic, requests to a retired model fail. OpenAI writes that access to the October 23 models has been "cut off" but doesn't describe the form of the error. On Foundry, every request after retirement returns a 410 error, according to Microsoft, and the dates are "not extendable." OpenAI mentions dedicated capacity for extended access on request through sales, without terms or pricing.

Foundry migrates some customers without asking. Standard, Global Standard and Data Zone Standard deployments are upgraded automatically at retirement, region by region, even to a version not yet separately available in that region. For several models, including Claude Haiku 4.5, codex-mini and GPT-4.1, the page names no replacement. Provisioned deployments (reserved capacity) are not upgraded automatically. Customers migrate them on their own, and per the Microsoft FAQ either "in place" in 20 to 30 minutes with no interruption, or in parallel. If you chose NoAutoUpgrade, your deployment stops on the deadline.

Bedrock extends access, possibly at higher prices. A model first goes into Legacy status, then into public extended access until end of life. AWS expects a higher price set by the model provider during this phase. For Claude Sonnet 4.5, the page names none. New customers can't use Legacy models, and existing customers can lose access after 15 days of inactivity.

Contradictions with their own rules

ActuIA finds three discrepancies between stated rules and actual dates:

1. Anthropic: The page promises at least 60 days of notice for active deployments. Foundry retires Haiku 4.5 on November 15 and Opus 4.5 on November 24. Anthropic lists both as active on October 28 ("not before" October 15 and November 24 respectively) and has published no deprecation notice. For Haiku 4.5, 18 days before the deadline, there's no visible 60-day notice.
2. Bedrock: Models are supposed to stay in Legacy status for "at least six months." For the Claude 3, 3.5 and 3.7 models in the affected regions, Legacy starts on October 10, 13 or 15, 2026. That gives at most 34 days, with extended access immediately, even though the page otherwise sets a three-month minimum. No explanation is given.
3. OpenAI: GA models are supposed to get at least six months of notice, specialized variants three. GPT-4o and GPT-4.1 got 92 days. The New Stack noted this on October 15. No justification has been given.

A replacement model doesn't mean the same settings

In Anthropic API version 2026-10-21, temperature, top_p and top_k are removed. Requests that include them get a 400 error. Earlier versions already reject them for Claude 4.7 and newer once they deviate from the default. The replacement for Sonnet 4.5, Claude Sonnet 5.5, is newer than 4.7, so if you pin the temperature, you need to remove it when you migrate. On Bedrock, status can vary by region. AWS points to ListFoundationModels and GetFoundationModel (the modelLifecycle field).

Next up: on January 6, 2027, OpenAI ends new fine-tuning, and on January 14 it shuts down GPT-4o and GPT-4.1 (replacement: gpt-6-luna). On March 30, 2027, Bedrock retires 26 models, including Llama 3 and 4, Mistral 7B, Mixtral 8x7B, Devstral 2, DeepSeek-R1 and GLM 4.7.

Takeaways

If your team integrates AI models via API, the finding is practical: the shutdown date depends on the platform, and the pages are sometimes outdated or contradict their own rules. Azure gives you more room on GPT-4o but is earlier on o3. Migration details such as removed sampling parameters matter too. The worrying part is that several vendors don't visibly keep the notice periods they promised, and no reasons are given. It remains to be seen whether Bedrock's extensions will cost more. The details are based on ActuIA's comparison of the public pages on October 28, 2026, and may change.