OpenAI announced on 22 August 2023 that fine-tuning is now available for GPT-3.5 Turbo, alongside updates to the Chat Completions API and deprecation timelines for older models [source].

The fine-tuning capability allows developers to customise GPT-3.5 Turbo on their own datasets, with OpenAI stating early tests showed fine-tuned versions can match or exceed GPT-4 capability on certain narrow tasks. The company reported one customer achieved a 50% reduction in prompt length after fine-tuning. Fine-tuning is priced at $0.008 per 1,000 training tokens and $0.012 per 1,000 input tokens for inference.

OpenAI also updated the Chat Completions API to support new function calling capabilities, including the ability to call multiple functions in a single request. The `gpt-3.5-turbo` model identifier will now point to `gpt-3.5-turbo-0613` by default, with automatic upgrades planned for future versions.

The announcement included deprecation notices for several legacy models. The original GPT-3 base models—`ada`, `babbage`, `curie`, and `davinci`—will be retired on 4 January 2024. Applications using these models must migrate to newer alternatives. The older `gpt-3.5-turbo-0301` and `gpt-4-0314` snapshots will also be deprecated on 13 September 2023, replaced by `-0613` versions.

OpenAI stated that developers using pinned model versions will need to manually update their integrations before the shutdown dates. The company provided a migration guide for transitioning from Completions API to Chat Completions API.

The changes represent a shift in OpenAI's model lifecycle management, with fixed deprecation windows now established for snapshot versions.