Back to stories
Models

OpenAI Retires Six Older Codex Models Including GPT-5 and GPT-5.1

Michael Ouroumis2 min read
OpenAI Retires Six Older Codex Models Including GPT-5 and GPT-5.1

OpenAI today completed the retirement of six older Codex models for ChatGPT sign-in users, the second stage of a deprecation timeline first signaled on April 7 when the affected models disappeared from the model picker. Today's removal pushes Codex users onto OpenAI's newer GPT-5.3 and GPT-5.4 families and narrows the set of coding models available without an API key.

What changed today

As of April 14, Codex no longer serves gpt-5, gpt-5.1, gpt-5.1-codex, gpt-5.1-codex-mini, gpt-5.1-codex-max, or gpt-5.2-codex to users who authenticate with their ChatGPT account. The supported lineup for ChatGPT sign-in is now gpt-5.4, gpt-5.4-mini, gpt-5.3-codex, and gpt-5.2. ChatGPT Pro subscribers retain exclusive access to gpt-5.3-codex-spark.

The timeline has been public for roughly a week: OpenAI pulled the older models from the in-product picker on April 7, giving developers a window to migrate or switch authentication methods before full removal.

Migration path for developers

Teams that still need the retired models have an escape hatch. According to OpenAI's guidance, users can run Codex with a personal API key or configure a custom model provider, which preserves access to older endpoints that remain on the API even after they leave the ChatGPT sign-in path.

That split — more aggressive consumer-side deprecation paired with continued API availability — mirrors a pattern OpenAI has used throughout the GPT-5 series, letting the company simplify the default product surface without stranding integrations that depend on specific model versions.

Why it matters

Codex has become one of OpenAI's most visible developer surfaces, and the April 14 cutover marks the cleanest break yet with the GPT-5 and GPT-5.1 generations. For individual developers, the change is mostly a one-time adjustment: pick a replacement model, re-run benchmarks, and move on. For teams that pinned automation to a specific model identifier, today is the deadline to either switch to an API-key-based flow or accept one of the currently supported Codex models as the new baseline.

The broader signal is that OpenAI is willing to move quickly on model retirement inside its consumer-facing tools. With GPT-5.4 and the GPT-5.3-codex variants now the default coding options in ChatGPT, older snapshots are being treated as legacy within months rather than years of release — a cadence that puts pressure on downstream tooling, agent frameworks, and educational content to keep pace.

What to watch

The next questions are whether the remaining supported models — particularly gpt-5.2, the oldest one still in the lineup — survive the next deprecation wave, and how quickly OpenAI extends the same consolidation to its API surface where the retired models are still reachable today.

Learn AI for Free — FreeAcademy.ai

Take "AI Essentials: Understanding AI in 2026" — a free course with certificate to master the skills behind this story.

More in Models

Thinking Machines Lab Debuts 'Interaction Models' — Mira Murati's First Step Into Frontier AI
Models

Thinking Machines Lab Debuts 'Interaction Models' — Mira Murati's First Step Into Frontier AI

Mira Murati's Thinking Machines Lab released a research preview of 'interaction models,' a new class of full-duplex multimodal AI that listens, sees and speaks at the same time, with turn-taking latency reported at about 0.4 seconds.

3 days ago2 min read
OpenAI Ships GPT-Realtime-2 With Live Translation and Streaming Whisper, Pushing Voice Agents Toward GPT-5 Reasoning
Models

OpenAI Ships GPT-Realtime-2 With Live Translation and Streaming Whisper, Pushing Voice Agents Toward GPT-5 Reasoning

OpenAI launched three new audio models on May 7 — GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper — adding GPT-5-class reasoning, a 128K context window, and metered live translation across 70 languages to its already-GA Realtime API.

6 days ago2 min read
Zyphra Releases ZAYA1-8B, the First Frontier-Class Reasoning MoE Trained Entirely on AMD
Models

Zyphra Releases ZAYA1-8B, the First Frontier-Class Reasoning MoE Trained Entirely on AMD

Zyphra's ZAYA1-8B is an Apache-licensed reasoning MoE with under 1B active parameters that matches Claude 4.5 Sonnet and Gemini 2.5 Pro on math benchmarks — and it was pretrained, midtrained, and fine-tuned on AMD Instinct MI300X GPUs.

1 week ago2 min read