Introducing improvements to the fine-tuning API and expanding our custom models program
OpenAI
OpenAI just gave developers more knobs to turn when fine-tuning its models, plus new paths into deeper custom-model work. Translation: less guessing, more control over how your AI actually behaves.
Based on reporting by OpenAI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI rolled out a batch of upgrades to its fine-tuning API this week, aimed squarely at developers who've been asking for more say in how their custom models get built. The pitch is straightforward: fine-tuning has been useful but somewhat opaque, and OpenAI wants to hand over more of the steering wheel.
The changes give teams tighter control over the training process itself, letting them shape model behavior with more precision than the previous one-size-fits-most setup allowed. That matters for companies running fine-tuned models in production, where small behavioral quirks can turn into real support tickets or compliance headaches. More control at the API level means fewer workarounds and less trial-and-error guessing.
Alongside the API tweaks, OpenAI is widening its custom models program, which up to now has been a fairly exclusive arrangement for a handful of partners willing to work closely with OpenAI's research teams to build bespoke versions of GPT models. Expanding that program signals OpenAI wants more companies, not just the biggest and best-funded ones, building genuinely specialized models rather than just prompting a general-purpose one and hoping for the best.
This is also a competitive move. Anthropic, Google, and a swarm of open-weight players have been chasing the same customers who want models tailored to their own data and workflows. OpenAI leaning into fine-tuning flexibility and custom builds is as much about retention as it is about capability. Enterprises that can shape a model to their exact needs are a lot less likely to shop around.
My take — AI-written commentary, not fact-checked reporting
I'll believe the 'more control' framing when developers stop complaining about black-box fine-tuning results, but this is clearly OpenAI reacting to the fact that closed models only win if they're also flexible. Open-weight models already give you full control by definition, so every feature like this is really OpenAI narrowing the gap it created for itself.
Read more about this at: OpenAI