The GPT-3.5 that OpenAI switches off on 28 September charged three times the chat model's input price for a quarter of the context

On 28 September 2026 OpenAI shuts down gpt-3.5-turbo-instruct, the last model in its catalogue built for the completions endpoint rather than for chat. The shutdown was announced on 26 September 2025, together with gpt-3.5-turbo-1106, and names gpt-5.6-terra as the replacement. The model is worth a second look on the way out, because its price never matched its specification. OpenAI's own model card lists it at $1.50 per million input tokens and $2.00 per million output tokens. The chat model of the same generation, gpt-3.5-turbo, costs $0.50 and $1.50. That is three times as much on input — for a context window of 4,096 tokens against the chat model's 16,385, with no fine-tuning, no function calling and no batch mode. What kept it in the catalogue was not capability but compatibility: code written against the legacy completions endpoint, which could not be rewritten as a conversation. That is also why the two models die separately. gpt-3.5-turbo-instruct goes on 28 September 2026; the chat model, deprecated in a separate notice on 22 April 2026, follows on 23 October 2026. When the second date passes, the GPT-3.5 generation is gone from the API. Both dates come from OpenAI's deprecations page, and both prices from the models' own cards.