Fine tuning in 2026: what it costs, what it fixes, and the platform shift nobody covered

OpenAI closed self-serve fine tuning to new customers in May 2026 and closes it to everyone in January 2027. Meanwhile 86% of a fine-tuned endpoint's monthly bill can be hosting. Here is when the technique still earns its place.

The economics of fine tuning changed in 2026, and the largest model provider quietly closed the door. OpenAI's deprecation documentation records that new customers lost access to fine-tuning on 7 May 2026 and that existing customers lose the ability to create new jobs on 6 January 2027.

That is on OpenAI's own pages and it has had almost no coverage in consulting content.

What changed, and when

OpenAI is winding down self-serve fine tuning

From OpenAI's own deprecation documentation. No rationale has been published and we do not speculate on one.

Source: OpenAI deprecations documentation, retrieved 25 September 2026. Google Vertex AI and AWS Bedrock continue to offer supervised fine tuning.

OpenAI has not published a rationale, so we will not invent one. The dates are the fact. Google Vertex AI and AWS Bedrock continue to offer supervised fine-tuning, so this is a narrowing of the field rather than the end of the technique.

The cost nobody models

Training is the cheap part, and almost every business case we see is built on it. The expensive part is the endpoint.

Training is the cheap part

A documented monthly example at 10,000 conversations, plus published training rates.

Source: Microsoft Foundry fine-tuning cost management documentation; Google Vertex AI and AWS Bedrock published pricing, retrieved September 2026.

Microsoft's own documented example is the clearest illustration: 10,000 conversations a month producing $22 of input cost, $176 of output cost, and $1,241 of hosting. The model sits there whether you use it or not, and 86% of the bill is that. On Google Vertex AI the picture is friendlier, because a tuned endpoint serves at the same price as the base model, so the structure differs enough by platform that the platform choice changes the answer.

The question to ask is not what fine tuning costs to train. It is what the endpoint costs per month at your actual utilisation, and whether that utilisation is high enough to justify a dedicated one.

What fine tuning does not fix

This is the part worth being blunt about, because it is the most common reason the technique gets funded.

Peer-reviewed work presented at EMNLP 2024 studied what happens when you fine-tune on facts the model does not already know. Two findings. Models learn those examples significantly more slowly than examples consistent with their existing knowledge. And as the new-knowledge examples are eventually learned, they linearly increase the model's tendency to hallucinate.

The authors' summary is that models acquire factual knowledge in pre-training, and fine-tuning teaches them to use it more efficiently. If your problem is that the model does not know something, fine tuning is the wrong instrument and will make the symptom worse.

What it is actually good for

When fine tuning earns its place, and when it does not

Establish evaluations, optimise prompts, fine-tune only if prompting is insufficient.

• The model will not hold your format or tone. Output consistency is what fine tuning is actually good at.
• Specialist vocabulary the model keeps getting wrong. Teaching usage of knowledge it already has, rather than adding knowledge.
• Matching a larger model with a smaller one on a narrow task. A real cost play, and underused.
• Shortening prompts at scale. Fewer examples and less context per request, which saves tokens at volume.
• The model does not know your facts. Fine tuning on new knowledge linearly increases the tendency to hallucinate. Use retrieval.
• The model hallucinates and you want it to stop. Peer-reviewed work is clear that this is the wrong instrument.

Source: Gekhman et al., EMNLP 2024, arXiv:2405.05904; OpenAI model optimization guidance, retrieved September 2026.

The vendor guidance and our own experience agree here. Fine tuning earns its place when the model will not hold your format, your tone or your specialist vocabulary, or when you want a smaller and cheaper model to match a larger one on one narrow task. That last case is a real cost play and it is underused.

The order that works: establish evaluations, optimise prompts, and fine-tune only if prompting is insufficient. The first two are reversible and cost days. The third is neither.

What most enterprises are actually doing

Prompting and retrieval. A survey of around 500 enterprise AI decision-makers found prompt design dominant, retrieval second, and fine tuning alongside tool calling and reinforcement learning described as niche and used primarily by frontier teams. Most production architectures are simpler than the conversation about them.

Before a fine-tuning business case, price the two cheap options properly. Prompt caching reads at a fraction of base input price and batch processing runs at roughly half. For a workload with a large stable context and any tolerance for latency, those two together often beat a dedicated endpoint, and they leave you free to change model next quarter.

What we would do first

Write the evaluation set before anything else. Fifty examples with the output you would accept. Then run them against a good prompt on a frontier model and count the failures by type. If the failures are format, fine tuning is a candidate. If they are facts, it is not.

Our AI readiness work covers the data prerequisites underneath either path, and MLOps consulting covers the evaluation harness that makes the comparison measurable. The companion piece on fine tuning and hallucination goes deeper on why the technique gets reached for in the wrong situations.

Frequently asked questions

Is fine tuning still available in 2026?

Less than it was. OpenAI's own deprecation documentation records that from 7 May 2026 fine-tuning jobs stopped being available to organisations that had not previously run fine-tuning, from 2 July 2026 to organisations that had not run inference on a fine-tuned model in the past 60 days, and that from 6 January 2027 active existing customers will no longer be able to create new fine-tuning jobs. OpenAI has not published a reason. Google Vertex AI and AWS Bedrock continue to offer it.

What does fine tuning actually cost?

Training is cheap and hosting is not. On Google Vertex AI, supervised fine-tuning of Gemini 2.5 Flash runs $5 per million training tokens and the tuned endpoint costs the same to serve as the base model. On Microsoft's platform the picture inverts: a documented worked example puts 10,000 conversations at $22 input, $176 output and $1,241 hosting, so 86% of the monthly bill is the endpoint sitting there. AWS Bedrock provisioned throughput runs at hourly rates that reach five figures monthly per model unit.

Does fine tuning fix hallucination?

No, and on new knowledge it makes it worse. Peer-reviewed work presented at EMNLP 2024 found that large language models learn fine-tuning examples introducing new facts significantly more slowly than examples consistent with what they already know, and that as those new-knowledge examples are eventually learned they linearly increase the model's tendency to hallucinate. The authors' conclusion is that models acquire facts in pre-training and fine-tuning teaches them to use those facts, not to add new ones.

When is fine tuning the right answer?

Format, tone and specialist vocabulary. OpenAI's own guidance points at training a smaller cheaper model to match a larger one on a narrow task, output consistency, classification, and shortening prompts to save tokens at scale. The recommended order is to establish evaluations first, optimise prompts second, and fine-tune only if prompting is insufficient. That order is worth following because the first two steps are reversible and the third is not.

Is RAG a substitute for fine tuning?

They solve different problems and the research suggests they compound. A Microsoft Research case study measured fine-tuning alone at over six percentage points of accuracy improvement and retrieval alone at five points, with the gains cumulative rather than overlapping. That study dates from 2024 and we found no equally rigorous 2026 replacement, so treat the magnitudes as indicative and the direction as sound.

What are enterprises actually doing instead?

Prompting and retrieval. Menlo Ventures' survey of around 500 enterprise AI decision-makers found prompt design dominates, followed by retrieval, while fine-tuning, tool calling and reinforcement learning remain niche and used primarily by frontier teams. Most production architectures are simpler than the discourse suggests.

How do we cut cost without fine tuning?

Caching and batching, both of which are published list features rather than engineering projects. Prompt caching reads at a fraction of base input price, and batch processing runs at roughly half price on the major providers. For a workload with a large stable context and tolerance for latency, those two together usually beat the economics of a fine-tuned endpoint without the lock-in.