Fine-tuning now available for GPT-4o

Source: OpenAI Date: 2024-08-20 URL: https://openai.com/index/gpt-4o-fine-tuning

Summary

Summary

OpenAI made fine-tuning available for GPT-4o in August 2024, allowing API customers to train GPT-4o on their own data to improve performance on specific tasks and domains. Fine-tuning GPT-4o extended a capability previously limited to GPT-3.5-turbo and older models, enabling domain-specific optimization of the flagship model.

Implications

Platform/enterprise thread. GPT-4o fine-tuning is a significant enterprise offering because it enables organizations to get GPT-4o quality with task-specific optimization — better than base GPT-4o on narrow domains, without the cost of training a custom model from scratch. The competitive implication is that fine-tuning erodes the advantage open-weight models offer for customization: if enterprises can fine-tune the frontier model, the argument for running local fine-tuned Llama becomes weaker. The pricing of fine-tuned GPT-4o inference (higher than base GPT-4o) also creates a new revenue tier for OpenAI. This capability rollout tracks the pattern of each major platform feature flowing from smaller to larger models as infrastructure and economics allow.