Eleven OpenAI models switch off on one day — and three GPT-5.6 variants inherit all of them

OpenAI's deprecation page now lists a single date that removes more of the company's history than any before it. On 23 October 2026 eleven models stop answering: GPT-3.5 Turbo, GPT-4, the GPT-4 preview snapshot from November 2023, GPT-4 Turbo, the May 2024 snapshot of GPT-4o, GPT-4.1 nano, GPT Image 1, o1, o1-pro, o3-mini and o4-mini. Five fine-tuned lines go with them, among them ft-babbage-002 and ft-davinci-002 — the last names in the API that still came from the GPT-3 era. The notice ran exactly to policy. OpenAI announced the list on 22 April 2026, 184 days ahead, and the company's published commitment for generally available models is six months. What replaces them is the more striking number. Eleven models collapse into four: GPT-5.6 Sol takes over from seven of them (GPT-4, both GPT-4 preview and Turbo snapshots, GPT-4o from May 2024, o1, o1-pro and o3-mini), GPT-5.6 Terra from two (GPT-3.5 Turbo and o4-mini), GPT-5.6 Luna from one, and GPT Image 2 from GPT Image 1. A generation of specialised models — a cheap chat workhorse, two sizes of reasoning model, a vision-capable mid-tier — is being answered with a single family running at different reasoning efforts. Two details cut against the usual pattern. First, within the GPT-4.1 family the shutdown hits the cheapest member: nano, at $0.10 per million input tokens, is switched off, while the base model at $2 and mini at $0.40 are not on the list at all. Vendors normally retire the expensive legacy tiers first; here the entry tier goes and the flagship stays. Second, o3-mini and o4-mini die on the same day, which removes the whole small-reasoning-model line at once — and they are pointed in opposite directions. The older o3-mini is sent to GPT-5.6 Sol, the top-end reasoning variant, while its own successor o4-mini is sent to Terra, the volume tier. Developers moving off two models that cost the same $1.10 per million input tokens are being routed to replacements at different levels of the price list. What this does not tell us is how many applications are affected. OpenAI publishes no usage figures per model, and a deprecation list says what will stop working, not what anyone still runs on it. All eleven models keep their entries in this catalogue after the shutdown, with the date recorded.