Has fine-tuning been debunked? No. OpenAI and Mistral are winding down self-serve fine-tuning of their own models, and training open models on a team's own task data keeps producing measured gains, including the largest single gain in our published QMSum study.
The view that fine-tuning is over has a clear source. In "The End of Finetuning" (13 May 2026), Latent Space's AINews newsletter called fine-tuning "the next casualty of the 2026 Side Quest massacre" after OpenAI's deprecation, and argued that "the modal 80% of the AI Engineering industry was probably trending there anyway." The same piece added that the top tier, naming Cursor and Cognition, had "INCREASED open model RLFT and usage."
This post covers what the closed labs stopped and why that is a product decision, where trained open models win with our measured results, and where prompting is enough.
What did the closed labs stop, and why?
Three providers have pulled back from fine-tuning their own models, each on its own schedule. Every fact below was read from the provider's documentation on 28 October 2026.
| Provider | What changed |
|---|---|
| OpenAI | New fine-tuning jobs end 6 January 2027 |
| Mistral | Fine-tuning API deprecated, no longer actively supported |
| Gemini API dropped tuning in May 2025 |
OpenAI is winding down in stages, per its deprecations page. Organisations that had never fine-tuned lost access on 7 May 2026, organisations with no fine-tuned inference in 60 days lost it on 2 July 2026, and remaining customers can create jobs until 6 January 2027. Fine-tuned models keep answering until their base model retires, and GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini retire on 14 January 2027.
Mistral lists its fine-tuning API under deprecated features: "This feature is deprecated and is no longer actively supported."
Google dropped tuning from the Gemini API and AI Studio with the deprecation of Gemini 1.5 Flash-001 in May 2025. It still offers supervised tuning for several Gemini models through Gemini Enterprise Agent Platform.
OpenAI's notice gives no reason for the change. Latent Space read it as one of the labs' 2026 cuts to side projects, which a crunch in compute would explain. Either way it is a decision about what a provider wants to host. Every fine-tune is a separate model the provider has to serve, version and eventually retire for one customer's traffic, and a lab focused on its general models has good reasons to stop. Those are reasons about the provider's business. The evidence on whether training works comes from measuring trained models, which the next section does. The deprecations do show the cost of renting a trained model: when the base model retires, the fine-tune goes with it.
Where do trained open models win?
A trained model wins where the same task runs at volume, the output can be checked, and the instructions repeat in every prompt. Four results from our research show where.
Training on the task beat improving the inputs. In our QMSum study (arXiv 2609.25028), on query-focused meeting summarization, we held one base model fixed and changed one thing at a time. Giving it better inputs, retrieved transcript spans in place of the transcript's opening words, added 1.55 ROUGE-1 on the test split and 0.29 on validation. Training it on those same span-shaped inputs added a further 5.29 ROUGE-1, with a 95% interval of +4.02 to +6.56. A released specialist moved to the new input shape without retraining lost 6.30 ROUGE-1, and training it on the new shape recovered the loss.
A smaller trained model held the quality. A specialist much smaller than our main system, trained the same way, scored 36.33 ROUGE-1 against our system's 35.41, a difference the benchmark cannot statistically separate. It needed less than half the peak inference memory.
The trained model cost a fraction to run. A model trained for that one task had an estimated serving cost per query 99% below Claude Opus 5's and 96% below GPT-5.6 Luna's, and answered 23 times faster than Claude Opus 5, at September 2026 prices and excluding training cost. The paper limits its quality comparison with hosted frontier models, because they wrote longer answers and no human or factuality evaluation was run, so the results we build on are the cost, the speed and the size of the training gain.
The spend sits on repeatable work. In OpenRouter's public data for the 30 days to 12 October 2026, classification, extraction and transformation took 20.8% of classified model spend, at 1.50 to 1.84 times the average price per token, and closed frontier models took 75% to 87% of the top ten models' spend on those tasks. These are tasks with labels and records as outputs, which a trained model can be evaluated on precisely.
A trained model also changes the token count. The instructions, protocol and examples a frontier model needs in every prompt are trained into the model, so each task sends fewer tokens. Our note on AI startup gross margins works through both cost dimensions with prompt sizes from real workloads.
Where is prompting enough?
Prompting a frontier model is the right starting point for most new tasks, and for some it stays the right answer.
| Situation | Start with |
|---|---|
| A few thousand tasks a month | Prompting a frontier model |
| Task definition still changing weekly | Prompting, with examples logged |
| Open-ended work that differs every request | A frontier model |
| Facts that change often | Retrieval |
| The same task at high volume | A trained model |
| Long instructions repeated in every prompt | A trained model |
| Outputs you can check against a schema | A trained model |
The split follows one principle we use in every build: behaviour, format and task skill go into the model, and facts that change go into retrieval. A prompt is the fastest way to find out what the task needs. Once the task is stable and the volume is there, the logged examples become the training set, and the evaluation set you built while prompting becomes the bar the trained model has to pass. For the trade-offs between the two techniques, see fine-tuning vs RAG.
What should you do if you fine-tuned on OpenAI?
Keep your training data, your evaluation set and your production logs, because the fine-tuned model itself stays on OpenAI's API and retires with its base model. Retrain the task on an open model, serve it behind an OpenAI-compatible API, and compare it with the old fine-tune before 14 January 2027, while the original is still there as a baseline. Our migration guide walks through the four steps, and the cost-per-task calculator shows what the switch does to your bill.
Ertas runs the adaptive model endpoint for this: a model trained on your task, served through an OpenAI-compatible API, with each new version checked against your evaluation set and approved by your team before it takes traffic. It runs Ertas-managed, in your own cloud or on-premise. The endpoint is currently invitation only, and we are developing a more streamlined version for wider access.
Frequently Asked Questions
Has fine-tuning been debunked?
No. OpenAI and Mistral are winding down self-serve fine-tuning of their own models, a product decision about what they host. Training open models on task data keeps delivering measured gains: in our QMSum study it added 5.29 ROUGE-1 within one base model, more than three times the gain from improving the inputs.
Is fine-tuning still used?
Yes. Teams train open models every day to make a model reliable on one task, and Google still offers supervised tuning for Gemini through Gemini Enterprise Agent Platform as of October 2026. Latent Space, the newsletter that called fine-tuning's end, also reported that leading AI companies had increased their reinforcement fine-tuning and use of open models.
How much does LLM fine-tuning cost?
As of 28 October 2026, OpenAI bills supervised fine-tuning at $5 per million training tokens for GPT-4.1 mini and $25 for GPT-4.1 and GPT-4o, for existing customers until 6 January 2027. Training open models yourself in Ertas Studio is free to start, with plans from $10 a month as of October 2026. A managed build is a fixed fee scoped after a free audit.
Which LLM is best for fine-tuning?
The smallest open-weight model that holds your quality bar on a held-out evaluation set, with a licence that allows commercial use. Compare a few sizes on your own examples. A model trained for one task can be much smaller than a general model and still match it on that task.
When should you fine-tune instead of prompting?
When the task is stable, runs at high volume, and carries long instructions in every prompt, or when its output can be checked against a schema. Prompting stays the better choice for low volume, open-ended work and tasks still being defined.
Does Gemini allow fine-tuning?
Through Gemini Enterprise Agent Platform, yes: as of 28 October 2026 Google lists supervised tuning for Gemini 3.5 Flash, Gemini 3.5 Flash-Lite, Gemini 3.1 Flash-Lite and three Gemini 2.5 models. The Gemini API and AI Studio have offered no tunable model since May 2025.
The Short Version
The closed labs have stepped back from hosting fine-tuned copies of their models. Training a model on your own task works as well as it did, and a trained open model runs for as long as you need it. Use prompting while the task is new or the volume is low, and train a model once the task is stable, repeated and expensive. If the money question comes first, start with where AI startup inference costs go.
Sources, read 28 October 2026: Latent Space, "The End of Finetuning" (13 May 2026); OpenAI deprecations and API pricing; Mistral fine-tuning API (legacy); Google Gemini API fine-tuning and Gemini Enterprise Agent Platform supervised tuning. Our research: the QMSum study and the OpenRouter structured-task report.