Measured results and practical guides on fine-tuning small open models, moving off OpenAI fine-tuning, and pricing AI features.
OpenAI stops new fine-tuning jobs on 6 January 2027. Here is the timeline, what you keep, what you lose, and how to move your training file to a model you own.
Your ft: model stays on OpenAI's servers. Your JSONL, your prompts and your test cases do not have to. A practical migration checklist for teams leaving OpenAI fine-tuning.
Your app already speaks the OpenAI Chat Completions format. Point it at your own fine-tuned model by changing the base URL, the key and the model name.
We fine-tuned open models on a public customer-support dataset and tested them against Claude Haiku, Claude Sonnet and others on the same held-out tickets. Full table, cost and timings.
We fine-tuned an open model to turn receipts into JSON and scored it against GPT, Claude and Gemini on every field. Where each model breaks, and what it costs per 1,000 documents.
Moderation needs the same decision for the same input, at high volume and low cost. We fine-tuned an open model on a public toxicity dataset and compared it with prompted frontier models.
If a large model already does the job well, its answers are training data. How distillation works, what the provider terms allow, and what we measured when we tried it.
A five-question test for deciding between a prompted frontier model and a fine-tuned small one, with the research that backs each answer.
Hosted models change under your code. Here is the research on how much, the notice periods providers give, and a simple test-set habit that catches it.
gpt-oss, Nemotron, DeepSeek, GLM and Inkling: sizes, licences and what each is good at when you fine-tune it for one job, with results from the same tasks.
How an AI agency can sell a dedicated fine-tuned model to each client, prove it beats the frontier models before the contract, and keep the client's data separate.
Agencies have no standard way to price AI work. Here is the cost maths for one feature at 100,000 calls a month, and three pricing models that hold up when the model gets cheaper.
For EU and UK agencies: who is controller and who is processor, what the DPA must say, where data goes during fine-tuning, and what the AI Act asks of a narrow fine-tune.