OpenAI announced on 20 August 2024 that fine-tuning is now available for GPT-4o, its flagship multimodal model [source]. The capability allows developers to customize the model's behaviour using their own datasets, a feature previously limited to older models including GPT-3.5 Turbo and the original GPT-4.

According to the announcement, fine-tuning for GPT-4o supports both text and vision inputs, enabling developers to improve performance on domain-specific tasks. OpenAI stated that early tests showed fine-tuned GPT-4o models outperformed GPT-4 Turbo with prompt engineering on certain structured generation and instruction-following tasks.

The company is offering 1 million training tokens per day at no cost through 23 September 2024. After that period, training will be priced at $25 per million tokens, with inference costs of $3.75 per million input tokens and $15 per million output tokens for fine-tuned models.

Fine-tuning introduces a new vector for model drift. When developers train custom versions of a base model, the resulting behaviour becomes specific to that training data. If the base model is updated or deprecated, fine-tuned versions may behave unpredictably or cease to function. OpenAI has not specified how long fine-tuned GPT-4o snapshots will remain available or how updates to the underlying model will affect existing fine-tunes.

The announcement did not address version pinning, rollback mechanisms, or notification protocols for changes to fine-tuned model behaviour. Developers relying on fine-tuned models for production systems may face continuity risks if OpenAI updates or retires the GPT-4o base model without maintaining compatibility with existing fine-tunes.