GPT-3.5 Turbo

GPT-3.5 Turbo is the model that turned conversational AI from a demonstration into an ordinary line item in a software budget. OpenAI put it in the API on 1 March 2023, alongside a new chat completions format, at $0.002 per thousand tokens — a tenth of what text-davinci-003 had cost the week before. The same GPT-3.5 family had been powering ChatGPT since its launch in November 2022, so developers were buying a model that millions of people had already used. It was never the strongest model of its generation and was not meant to be. In OpenAI's own GPT-4 technical report it scored 70.0% on MMLU and 48.1% on HumanEval, well below GPT-4. What it had was price, latency and an interface that everyone copied: a list of role-tagged messages, a system prompt, and later function calling (June 2023), a 16,385-token context and JSON mode (gpt-3.5-turbo-1106, November 2023) and fine-tuning on customer data (22 August 2023). The January 2024 snapshot gpt-3.5-turbo-0125 cut the price again, to $0.50 per million input tokens and $1.50 per million output. For two years it was the default choice for classification, extraction, routing and every other task where a frontier model would have been an extravagance. Its retirement has been unusually slow. OpenAI removed it from ChatGPT in July 2024, when GPT-4o mini took over the free tier, but kept it in the API for developers who had built around it. On 22 April 2026 the company set an end date: gpt-3.5-turbo-0125 is to be shut down on 23 October 2026, together with gpt-4-0613 and the o1 snapshot, with GPT-5.6 Terra given as the migration target. That will close the model line that gave the industry its first idea of what an AI assistant should cost.

▸Videos

No videos yet.

▸Reviews