OpenAI enabled supervised fine-tuning for GPT-5.5 mini on Tuesday, charging $6 per million training tokens with no hourly hosting fee for the resulting model and inference priced 15 percent above the base model rate. That structure is new. Previous fine-tuning offerings charged for keeping a customized model available, which meant a company with forty department-specific variants paid for forty idle deployments. Removing the hosting charge makes narrow, specific fine-tunes economical, and it targets the main reason teams choose open weight models: the ability to customize freely on infrastructure they already pay for.
What changed in the offering
OpenAI's previous fine-tuning was available on older models and carried a per-hour charge for a deployed customized model, which for a rarely used variant dominated the economics. The new arrangement uses a low-rank adaptation approach where the customization is a small set of weights loaded on demand onto shared base model capacity. Cold-start latency for an infrequently used adapter measured about 900 milliseconds in testing, dropping to normal after the first request.
Training runs are priced per token consumed rather than per hour, which makes costs predictable before starting. A typical customization on 5,000 examples averaging 800 tokens each runs 4 million tokens across a standard three epochs, costing $24. That is small enough that experimentation becomes reasonable, and the difference between $24 and the multi-hundred-dollar minimums of the previous offering is the difference between trying five approaches and committing to one.
What fine-tuning is actually good for
The persistent misconception is that fine-tuning teaches a model facts. It does not do that well, and retrieval remains the right approach for knowledge. What fine-tuning does well is behavior: consistent output format, adherence to a specific style, correct handling of domain-specific terminology, and reliably following instructions that a prompt struggles to convey. A model fine-tuned on 800 examples of a company's support tone will match it better than any system prompt.
The second strong use case is cost reduction. A fine-tuned mini model that matches a larger model's performance on a narrow task, which is achievable for classification and extraction, runs at a fraction of the price. A logistics company we spoke with replaced a full-size model doing shipping document extraction with a fine-tuned mini and cut that workload's cost by 82 percent with no measurable accuracy loss on their evaluation set.
The story is rarely the launch. It is what breaks, what ships, and who owns the mess at 2 a.m.
The competitive picture
Fine-tuning has been the standard argument for open weights. If you can adapt Llama or Qwen freely on your own hardware, you control the result, you can deploy anywhere, and the marginal cost of another variant is zero. OpenAI's new pricing does not match that, since you still cannot take the weights anywhere, and it removes the cost objection that made the open path obviously cheaper for teams without existing GPU capacity.
Google offers tuning on Gemini through Vertex with a similar adapter approach and comparable pricing. Anthropic still does not offer fine-tuning on Claude at all, positioning prompt engineering and its longer context as the alternative, a stance that increasingly looks like a gap rather than a philosophy. Amazon Bedrock supports fine-tuning for the models it hosts with a hosting charge that makes narrow variants uneconomical, which is the problem OpenAI just solved.
The practical workflow
OpenAI's tooling produces an evaluation report comparing the fine-tuned model against the base on a held-out split, which is more than the previous version offered and less than a serious evaluation requires. Teams should bring their own evaluation set representing actual production distribution, because the most common fine-tuning failure is a model that scores better on curated examples and worse on the messy real ones.
The other frequent mistake is too little data or too much. Fewer than roughly 200 examples usually produces no reliable improvement. More than about 20,000 for a narrow behavioral task produces diminishing returns and increases the risk of degrading general capability. The sweet spot most practitioners report is between 500 and 5,000 carefully curated examples, with curation mattering far more than volume.
What this does not change
Data residency and control remain the arguments for open weights that pricing cannot address. Fine-tuning through OpenAI means sending training data to OpenAI, which for a bank's internal documents or a hospital's records is a governance conversation regardless of the contractual terms. OpenAI's enterprise agreement covers data handling and does not change where the data goes.
There is also a durability question. A fine-tuned model is tied to a base model version, and base models get deprecated. OpenAI's policy gives twelve months of notice, which is generous by its historical standards and short compared to the life of a production system. Teams building on a fine-tune should plan for retraining as a recurring operational task rather than a one-time project.
The evaluation discipline matters more than the pricing here, and it is where most teams will waste money. A fine-tune should be compared against a well-engineered prompt on the same base model, not against a mediocre prompt, because a surprising share of the improvements teams attribute to training were available for free from better instructions. Several practitioners we asked run that comparison as a gate before approving any training run. The ones who skip it end up maintaining a customized model that beats a prompt nobody spent an afternoon improving.
Skarvonix will keep following this beat with reporting grounded in how systems behave outside the launch keynote.
- LLMs
