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

On October 28, ActuIA cross-checked the deprecation pages of OpenAI, Anthropic, Microsoft and AWS. The finding is simple to state and awkward to manage: for one and the same model, the retirement date can differ by 22 to 173 days depending on where you call it. GPT-4o and GPT-4.1 stop in the OpenAI API on January 14, 2027, while Microsoft Foundry, Azure's model service, keeps them until April 14. In the other direction, Azure retires o3 and o3-pro on November 19, 22 days before OpenAI does on December 11.

For a team whose acquisition or monetisation runs partly on these models, this goes well beyond infrastructure. Product descriptions, ad variants, onboarding assistants, support bots, lead scoring and email sequences increasingly depend on a specific model identifier. When that identifier stops answering, the funnel step it powers stops with it.

What actually happens on the day

The consequences differ by provider, according to ActuIA's reading of the pages.

On Anthropic's API, requests to a retired model fail. On Azure Foundry, they return a 410 error, and the dates cannot be extended. Standard deployments are upgraded automatically region by region, sometimes to a version not yet offered separately in that region; Provisioned deployments are not upgraded, and deployments set to NoAutoUpgrade simply stop. OpenAI's page says access is cut without describing the failure, and mentions possible dedicated capacity through its sales team, with no terms given.

AWS Bedrock takes another route: access can be extended, usually in Legacy status, at a price set by the model provider. New customers cannot use Legacy models, and existing customers can lose access after 15 days of inactivity.

So the same deadline can mean an error, a silent switch to another model, or a higher bill. Each has a different cost for a live funnel. An error is visible. A silent switch can change tone, answer length or conversion without anyone noticing. A price rise erodes the margin of any feature priced on the old cost.

Where the published rules do not match

ActuIA flags three gaps. Anthropic promises at least 60 days' notice and says its dates apply to its API and to Foundry, yet Foundry retires Claude Haiku 4.5 on November 15 and Claude Opus 4.5 on November 24, while Anthropic still listed both as active on October 28. Bedrock says models stay at least six months in Legacy before end of life, but the Claude 3, 3.5 and 3.7 models leaving on November 13 in some regions get at most 34 days. OpenAI promises six months' notice for generally available models; GPT-4o and GPT-4.1 received 92 days.

The practical lesson: a provider's stated policy is a weak planning tool. The dated entry for your model, on the platform you actually call, is the one to trust.

Replacements are named for some models: gpt-6-luna for GPT-4o and GPT-4.1, gpt-5.6-sol and gpt-5.6-terra for several o-series models on Foundry, Claude Sonnet 5.5 for Claude Sonnet 4.5. ActuIA also notes that Anthropic's API version of October 21 removes temperature, top_p and top_k, which now trigger a 400 error, so integrations migrating from Sonnet 4.5 must drop those settings.

What to do this week

• Inventory every model identifier in production, with the platform and region it runs on, including those inside no-code tools and your agencies' workflows.
• For each one, record the retirement date published by the platform you actually use, not the model maker's announcement.
• Flag the features that touch revenue or acquisition directly: checkout assistants, ad copy generation, lead qualification, retention emails.
• Test the named replacement on a sample of real inputs and compare outputs before the switch, not after.
• Check deployment settings on Azure and Bedrock, since auto-upgrade, Provisioned and Legacy status change what happens on the day.
• Put the November dates in the calendar first: ActuIA counts six deadlines that month.