OpenAI upgrades fine-tuning API with checkpoints and side-by-side evals

Prompt engineering gets messy, and basic RAG setups often flood your context window with useless noise. When you genuinely need an LLM to master deep domain knowledge, fine-tuning is usually the next frontier—except it has historically been clunky and expensive.

OpenAI just rolled out a major double-barrel update to fix that. The company upgraded its self-serve fine-tuning API with better developer tooling, while officially expanding its Custom Model program to let enterprise teams build bespoke models alongside OpenAI researchers.

More control over the training run

Since OpenAI launched self-serve fine-tuning for GPT-3.5 in August 2023, thousands of teams have trained hundreds of thousands of models. But catching overfitting early or comparing performance across checkpoints has remained an annoying chore.

The new API update introduces epoch-based checkpointing, automatically saving a snapshot after every training cycle so teams can roll back without retraining from scratch.

Developers also get a side-by-side Comparative Playground to evaluate two model outputs against a single prompt, dashboard-level hyperparameter tuning, full validation dataset metric calculations, and an official integration with Weights and Biases.

Job platform Indeed showed off what this looks like in practice. By fine-tuning GPT-3.5 Turbo to generate job recommendation explanations, they slashed prompt token usage by 80%, scaling their monthly outreach from under 1 million to roughly 20 million messages.

White-glove models for enterprise giants

For companies that outgrow the basic API, OpenAI is expanding its Custom Models program into two dedicated tiers: Assisted Fine-Tuning and Fully Custom-Trained Models.

Assisted fine-tuning pairs enterprise teams directly with OpenAI researchers to apply advanced PEFT techniques and custom evaluation pipelines. SK Telecom used this setup to fine-tune GPT-4 for Korean telecom customer service, boosting intent recognition by 33% and lifting customer satisfaction scores from 3.6 to 4.5 out of 5.

The fully custom tier goes even deeper, altering mid-training and post-training steps to absorb massive proprietary datasets.

Legal tech firm Harvey partnered with OpenAI to train a custom model on 10 billion tokens of legal text and case law. With OpenAI modifying the core training process and baking in feedback from practicing attorneys, the resulting model scored an 83% jump in factual accuracy, beating base GPT-4 in attorney preference 97% of the time.

Why it matters

For developers, the API polish removes friction where it hurts most. Features like automated checkpointing and side-by-side evaluation turn fine-tuning from a blind gamble into a repeatable engineering process. If you are burning tokens on massive prompt instructions, fine-tuning smaller models is now much easier to justify.

From an industry perspective, we think OpenAI is playing a brilliant playbook by bifurcating its product strategy. Long-tail developers get an automated self-serve platform that keeps them inside the ecosystem, while lucrative enterprise accounts get high-touch consulting wrapped in proprietary mid-training access.

The trade-off is clear: lock-in. Sinking proprietary data into a closed-source custom training run yields incredible accuracy, but it marries your core workflow to OpenAI indefinitely. Companies sitting on valuable proprietary corpora will have to weigh whether the immediate performance gains are worth giving up architectural independence.