OpenAI Deprecated Models and API Retirement Schedule

An OpenAI model deprecation announces a planned retirement, but the model keeps answering calls until its final shutdown date. At Layer3Labs, we build AI workflows for businesses. Tracking model lifecycles prevents outages when old endpoints stop responding. This guide is a snapshot of the retirements OpenAI had announced as of September 8, 2026.

A deprecated model remains usable during its notice window. Once shutdown arrives, API requests return an error immediately. Because OpenAI publishes updates across different schedules, verify your exact model IDs against the official OpenAI deprecations page.

Deprecation, Shutdown, and Legacy Status Explained

OpenAI defines three distinct lifecycle stages for older models: deprecation, shutdown, and legacy status. Deprecation represents the public notice period. The model continues to process API requests normally while deprecated. Developers use this window to test replacement models and update their codebases.

Shutdown marks the day the model stops answering requests completely. Any request sent to a shut model fails with an invalid model error. Legacy status is different. A legacy model no longer receives updates and is likely to be deprecated later. Legacy models still answer requests, but they sit next in line for retirement.

• Deprecation: OpenAI announces a model will be retired, but the endpoint stays fully functional until the shutdown date.
• Shutdown: The endpoint ceases operations, and any subsequent request to that model ID fails.
• Legacy: still answering, no longer updated, and likely to be deprecated later.

Notice Windows Across OpenAI Model Tiers

OpenAI provides tiered advance notice periods before shutting down a deprecated model. Generally available models receive at least 6 months of notice before shutdown. That is enough time to migrate safely. OpenAI communicates these timelines through direct customer emails and documentation updates.

Specialized models get a shorter runway. Chat variants and codex models receive at least 3 months of advance notice. Preview models get as little as 2 weeks. That is why OpenAI does not recommend them for production. Safety or compliance concerns can also force a faster shutdown timeline on any tier.

• Generally available models receive at least 6 months of notice before complete shutdown.
• Specialized variants like chat models and codex models receive at least 3 months of notice.
• Preview models receive as little as 2 weeks of advance notice.
• Safety and compliance issues allow OpenAI to accelerate any retirement schedule.

Flagship and Reasoning Model Retirement Schedule

OpenAI scheduled eight prominent flagship and reasoning models for shutdown in late 2026 across two major announcement batches. The earliest batch shuts down on October 23, 2026. It takes original GPT-4 snapshots, legacy GPT-3.5 endpoints, and the early o1 reasoning model. Update those endpoint strings first. October is the nearer deadline.

The second major wave of chat model shutdowns lands on December 11, 2026. This batch removes early GPT-5 snapshots and the o3 reasoning snapshot. OpenAI recommends migrating workloads to the newer GPT-5.6 family. Checking your pinned IDs early prevents unexpected failures in December.

• Shutdown October 23, 2026 (announced April 22, 2026) - gpt-3.5-turbo-0125 -> gpt-5.6-terra
• Shutdown October 23, 2026 - gpt-4-0613 -> gpt-5.6-sol
• Shutdown October 23, 2026 - gpt-4-turbo-2024-04-09 -> gpt-5.6-sol
• Shutdown October 23, 2026 - o1-2024-12-17 -> gpt-5.6-sol
• Shutdown December 11, 2026 (announced June 11, 2026) - gpt-5-2025-08-07 -> gpt-5.6-sol
• Shutdown December 11, 2026 - gpt-5-mini-2025-08-07 -> gpt-5.6-terra
• Shutdown December 11, 2026 - gpt-5-nano-2025-08-07 -> gpt-5.6-luna
• Shutdown December 11, 2026 - o3-2025-04-16 -> gpt-5.6-sol

Audio, Realtime, and Speech Model Retirements

Voice, audio, and transcription models follow their own retirement timeline running into early 2027. These endpoints handle multimodal audio processing, streaming conversation, and speech-to-text workflows. Because these models serve latency-sensitive pipelines, testing replacements requires verifying audio payloads and response formats. Do not wait until the shutdown deadline to validate audio streaming.

OpenAI announced retirements for realtime and speech models across two separate dates in summer 2026. The July 20 announcement sets a January 20, 2027 cutoff for voice pipelines. Audio workflows must switch to gpt-audio-1.5. The August 26 announcement sets a February 26, 2027 shutdown for transcription systems. It catches Whisper and the mini transcribe models alike.

• Shutdown January 20, 2027 (announced July 20, 2026) - gpt-realtime -> gpt-realtime-2.1
• Shutdown January 20, 2027 - gpt-audio -> gpt-audio-1.5
• Shutdown January 20, 2027 - gpt-4o-audio -> gpt-audio-1.5
• Shutdown February 26, 2027 (announced August 26, 2026) - whisper-1 -> gpt-live-transcribe or gpt-transcribe
• Shutdown February 26, 2027 - gpt-4o-transcribe -> gpt-live-transcribe or gpt-transcribe
• Shutdown February 26, 2027 - gpt-4o-mini-transcribe -> gpt-live-transcribe or gpt-transcribe

Assistants API and Sora 2 Specialized Deprecations

Some OpenAI deprecations involve architectural API migrations rather than simple model ID swaps. The Assistants API was announced for retirement on August 26, 2025. Its final shutdown date was August 26, 2026. Because that date has passed, any unmigrated calls to the Assistants API fail. OpenAI directs developers to rebuild those workflows using the Responses API and the Conversations API.

Specialized video generation tools also undergo planned retirements. OpenAI announced the shutdown of Sora 2 video generation on March 24, 2026. Its shutdown date arrives on September 24, 2026. OpenAI has not listed a direct replacement model for Sora 2 on its deprecations page. Teams using Sora 2 must prepare for this endpoint to stop answering requests.

• Announced August 26, 2025 (Shutdown August 26, 2026): The Assistants API is retired. Migrate to the Responses API and Conversations API.
• Announced March 24, 2026 (Shutdown September 24, 2026): Sora 2 video generation shuts down with no replacement model listed.

Integration Failures and Codebase Audit Steps

Sending an API call to a shut down model returns an immediate client error. Your application receives an HTTP 400 or HTTP 404 response. The payload states that the requested model does not exist or has been retired. Check the OpenAI API error codes for the exact payload.

Locating deprecated model identifiers requires inspecting your repositories and configurations. Search for explicit model strings across your configuration files and environment variables. Review OpenAI project settings to see which keys trigger deprecated calls. Check the OpenAI community if a third-party plugin fails.

The failure mode we hit most often is teams assuming that a global search for model strings will catch every dependency. Hardcoded strings inside utility scripts are easy to spot. Dynamic model variables passed through third-party workflow tools often escape static detection entirely. Audit runtime traffic directly. Inspecting live gateway logs reveals deprecated requests that repository scans routinely overlook.

ChatGPT Interface Removals Compared to API Availability

Removal from the ChatGPT consumer interface and retirement from the OpenAI API operate on completely separate schedules. A model can disappear from ChatGPT while remaining fully functional in the API. The consumer web interface updates its picker according to product priorities. The API follows formal deprecation policies with published advance notice windows. Do not assume an API endpoint is gone just because ChatGPT no longer lists it.

Understanding the difference between undated model aliases and dated snapshot IDs protects production stability. An undated alias points to whichever model OpenAI designates as current. Aliases can change behavior overnight without changing your code. A dated snapshot ID never changes its underlying weights or behavior. Pinning a dated snapshot converts silent behavioral drift into a predictable, scheduled retirement date. Read our guide on ChatGPT model not found errors to troubleshoot missing model names.

When to Rely on Retirement Schedules and When to Verify

This guide serves developers and system administrators managing automated pipelines connected to the OpenAI API. It is not for casual ChatGPT users wondering about chat interface updates. Consumer ChatGPT features change without formal deprecation notices. If you only use the web chat interface, check the ChatGPT help center instead.

Our recommendation to pin specific snapshot models changes if your application requires automated daily updates. Teams with continuous automated evaluation suites can safely use undated aliases. They absorb upstream model changes automatically. For all other production environments, audit your pinned model strings against the OpenAI deprecations page before October 2026.