On October 14, 2026, OpenAI set the retirement of GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini in its API, the interface through which software calls its models. On Microsoft Foundry, Azure’s platform, the same GPT-4o versions remain available until April 14, 2027, whereas o3 is removed there on November 19, 22 days before the date set by OpenAI. On October 28, ActuIA compared the deprecation pages of OpenAI, Anthropic, Microsoft Foundry and Amazon Bedrock: for the same model, the observed gap ranges from 22 to 173 days.
- OpenAI retire GPT-4o et GPT-4.1 de son API le 14 janvier, alors qu’Azure les maintient jusqu’au 14 avril 2027.
- Pour un même modèle, les délais de retrait varient de 22 à 173 jours selon les plateformes.
- Les effets du retrait diffèrent : Anthropic renvoie une erreur, Foundry une erreur 410, et Bedrock prolonge l’accès, souvent à un prix plus élevé.
- Anthropic, Bedrock et OpenAI ne respectent pas toujours les préavis qu’ils annoncent.
Six deadlines in November
The first change should go unnoticed. Starting November 2, xAI serves requests addressed to grok-imagine-image-quality with another model, grok-imagine-image-2.0, set to low quality, according to its migration guide. The change had been announced on September 2, sixty days in advance. The identifier remains valid and the code does not need to be changed, writes xAI, and each image costs $0.01 less.
On November 13, Bedrock removes Claude 3 Haiku, Claude 3 Sonnet, Claude 3.5 Sonnet, its version 2 and Claude 3.7 Sonnet in part of its regions. On Foundry, codex-mini and Claude Haiku 4.5 disappear on November 15, with no replacement indicated. On November 19, o1, o1-pro, o3, o3-pro, o3-deep-research, o3-mini and o4-mini cease to be served, with gpt-5.6-sol and gpt-5.6-terra as replacements. On November 30, Claude Sonnet 4.5 stops responding, a date Anthropic presents as “attempt”. The same day, OpenAI shuts down Agent Builder, its agent design tool, as well as the Evals platform.
What changes on the day of withdrawal
Behaviors diverge across platforms. At Anthropic, requests to a retired model fail. On Foundry, they return a 410 error. Deployments in reserved capacity, called Provisioned, are not upgraded automatically: their teams must migrate themselves, “in place” in 20 to 30 minutes without interruption according to Microsoft’s FAQ, or side‑by‑side. For other deployments, Foundry automatically switches to a newer version, sometimes even to a version not yet separately available in the region.
Bedrock extends access at a price that may increase. A model first moves to Legacy status, then to public extended access, and AWS indicates that a higher price should be expected, set by the model’s provider. No new customer can use a Legacy model, and an existing customer may lose access after 15 days of inactivity.
Displayed rules that do not always match the dates
Anthropic promises at least 60 days’ notice to customers with active deployments. Yet Foundry removes Claude Haiku 4.5 on November 15, while Anthropic’s page lists this model as active, with a removal “not before” October 15. At 18 days before the deadline, no 60‑day notice is visible on Anthropic’s page.
Bedrock states that a model remains “at least six months” in Legacy status. Yet, in the regions affected by the November 13 removal, Claude 3, 3.5 and 3.7 enter Legacy between October 10 and October 15, 2026 – at most 34 days – and the page provides no explanation. OpenAI, for its part, announces at least six months’ notice for models in general availability. GPT-4o and GPT-4.1 received 92 days, a timeframe that the US site The New Stack noted on October 15 and that OpenAI’s page does not justify.
The replacement does not always configure like the old model. In Anthropic’s API version 2026-10-21, the parameters temperature, top_p and top_k are removed, and a request containing them receives a 400 error. After January 6, 2027, OpenAI will stop accepting new fine‑tuning, i.e., adapting its models to proprietary data, for its existing customers. On March 30, 2027, Bedrock will remove 26 models, including Llama 3 and 4, Mistral 7B and DeepSeek‑R1.
