OpenAI lets devs fine-tune GPT-3.5 Turbo, claims it can beat GPT-4

OpenAI just handed developers the keys to one of its most popular models. Fine-tuning for GPT-3.5 Turbo is live, letting businesses inject their own data to create custom, more obedient versions of the model. The company says a fine-tuned 3.5 Turbo can match or even beat the base GPT-4 on certain narrow tasks — a claim that’s bound to raise eyebrows given the capability gap between those two models. For companies watching their inference bills, this is genuinely interesting.

In early tests, the results look practical rather than magical. OpenAI highlighted three core use cases: improved steerability so the model follows specific instructions like responding in German, reliable output formatting for things like generating clean JSON from user prompts, and injecting a custom brand tone. The real kicker is prompt compression. Some early testers slashed their prompt sizes by up to 90% by baking instructions directly into the model weights, which means faster API calls and lower per-request costs. For high-volume applications, that math adds up fast.

Pricing is split into training and usage. Training runs $0.008 per 1,000 tokens, while input and output during usage cost $0.012 and $0.016 per 1,000 tokens respectively. A modest fine-tuning job — say, a 100,000-token training file run for three epochs — clocks in around $2.40. The fine-tuned model also handles 4,000 tokens, double what previous fine-tuned versions could manage. Support for the 16k context window and function calling is coming later this fall, along with GPT-4 fine-tuning.

OpenAI is being careful to note that customer data stays put — it’s not used to train other models. Safety gets a nod too, with training data passed through the Moderation API and a GPT-4-powered system to catch anything unsafe. The old GPT-3 base models are still headed for retirement on January 4, 2024, with babbage-002 and davinci-002 now available as drop-in replacements via a new fine-tuning endpoint. For developers who’ve been waiting to make GPT-3.5 Turbo feel less generic and more like their own product, the door is finally open. Whether it truly rivals GPT-4 on specific tasks is something the community will stress-test soon enough.