benchrEvery model OpenAI shipped before August 2025, and every reasoning model before GPT-5 — gone in a single day.
Model IDs scheduled for retirement
| Retiring ID | What it was | OpenAI's replacement | benchr's note |
|---|---|---|---|
gpt-4o-2024-05-13 | The 2024 default | gpt-5.5 | Full migration page; GPT-5 is the cost match |
gpt-4-0613 | Original GPT-4, June 2023 snapshot | gpt-5.5 | If this is still in production, the swap is overdue by two model generations |
gpt-4-turbo + gpt-4-1106-preview | The 128K-context workhorse of early 2024 | gpt-5.5 | GPT-5's 400K context covers every Turbo use case |
gpt-3.5-turbo-0125 | The model that started the API gold rush | gpt-5.4-mini | GPT-5 Mini ($0.25/$2.00) is the tracked-and-priced alternative |
o1-2024-12-17 | First production reasoning model, $15/$60 | gpt-5.5 | GPT-5.5 reasons better at $5/$30, a two-thirds input cut |
o1-pro-2025-03-19 | OpenAI's priciest model ever: $150/$600 | gpt-5.5-pro | Whatever you migrate to costs less than this did |
o3-mini-2025-01-31 | Budget reasoning, early 2025 | gpt-5.5 | Re-evaluate whether you still need a reasoning tier at all |
o4-mini-2025-04-16 | Last pre-GPT-5 mini reasoner | gpt-5.4-mini | Fine-tuned o4-mini variants die the same day |
gpt-4.1-nano | Cheapest 4.1-era tier, April 2025 | gpt-5.4-nano | Only the nano goes — gpt-4.1 and 4.1-mini have no announced date |
gpt-image-1 | First GPT-branded image model | gpt-image-2 | Image 1 mini and 1.5 get a longer runway, to December 1 |
The notice is generous: six months, per OpenAI's policy for generally available models. The scope isn't. This isn't pruning one stale snapshot. It's everything from before the GPT-5 line, including the entire o-series reasoning family, retired simultaneously. The longer a model lived, the more places its ID hides: SDK defaults, notebook experiments that became cron jobs, vendor integrations you don't own the code for.
Related retirements before and after October
- Jul 23Codex, deep research, computer use + chat snapshots
Five Codex IDs (gpt-5-codex through gpt-5.2-codex) retire into GPT-5.5, o3- and o4-mini-deep-research into GPT-5.5 Pro, computer-use-preview into GPT-5.4 mini, with gpt-5-chat-latest and gpt-5.1-chat-latest going the same day.
- Aug 10More chat snapshots
gpt-5.2-chat-latest and gpt-5.3-chat-latest retire in favor of GPT-5.5.
- Aug 26Assistants API sunset
Replaced by the Responses API + Conversations API. Threads, runs, and tool orchestration need rebuilding.
- Sep 24Sora 2 and the Videos API
Both go, with no listed replacement. The video market consolidated elsewhere.
- Oct 23The ten-model wave
GPT-4o, GPT-4, GPT-4 Turbo, GPT-3.5 Turbo, GPT-4.1 nano, GPT Image 1, o1, o1-pro, o3-mini, o4-mini all fail from this date.
- Nov 30Platform trio
The Reusable Prompts API, the Evals platform, and Agent Builder shut down (announced June 3, 2026). Agent Builder's official path is the Agents SDK.
- Dec 1GPT Image 1 mini / 1.5
Older image models leave the API; GPT Image 2 is the path (announced June 2, 2026).
Fine-tuning dates and constraints
Buried in the same announcement: OpenAI is phasing out open access to fine-tuning. Since May 7, 2026, organizations with no fine-tuning history can't create new jobs. From July 2, 2026, you also lose access if none of your fine-tuned models served inference in the past 60 days. From January 6, 2027, job creation closes to all but active existing customers.
Pair that with the base-model retirements and the implication is sharp: every fine-tune built on a retiring model dies with it, and your ability to re-tune on a newer base is no longer guaranteed. If a fine-tuned GPT-4o or GPT-3.5 model carries production traffic, the decision isn't "migrate the tune." It's whether prompting a current model replaces it entirely. Our RAG vs fine-tuning breakdown runs that math.
Replacement planning
OpenAI's documented paths send general GPT-4o and GPT-4 Turbo traffic toward GPT-5 ($1.25/$10), high-volume simple traffic toward GPT-5 Mini ($0.25/$2.00), and reasoning workloads from o1 or o3-mini toward GPT-5.5 ($5/$30). Treat those as migration candidates, not drop-in equivalents. Run your volumes through the calculator and compare the same workload with Gemini 3.5 Flash and DeepSeek V4-Pro before cutover.