LLM Fine-Tuning Costs: What Each Provider Actually Charges (2026)

Fine-tuning gets pitched as a way to shrink your model bill — swap a fine-tuned small model in for a big general-purpose one on a narrow task. That can genuinely work, but the pricing structure varies enormously by provider, and two of the labs whose base-model pricing dominates comparison charts don't offer public fine-tuning at all. Here's what's actually available and what it actually costs, provider by provider.

What fine-tuning is (and isn't)

Fine-tuning updates a model's weights on your own labeled examples, so the behavior is baked into the model itself rather than re-supplied on every call. That makes it fundamentally different from a long system prompt, few-shot examples, or RAG — those all re-send context on every request (and can be cached for a discount); a fine-tuned model has already "learned" the pattern and needs a much shorter prompt to reproduce it. That's the real economic case: if your task involves an expensive few-shot prompt (thousands of tokens) repeated on every call, fine-tuning can shrink that prompt to a few hundred tokens on a smaller model — at the cost of an upfront training bill and ongoing complexity.

Provider by provider

OpenAI — training cost varies by model (roughly $0.80–$3.00 per million training tokens for GPT-4.1-class models, or an hourly rate for o-series reasoning models), and fine-tuned inference typically runs at a real markup over the equivalent base model's standard rate. The bigger story here isn't the price, though: since May 7, 2026, organizations that had never previously run a fine-tuning job can no longer start a new one — OpenAI is winding the self-serve fine-tuning API down through 2027, citing cheaper GPU rental as the more cost-effective path for most use cases. If you're evaluating this today, check current availability before planning around it.

Google (Gemini / Vertex AI) — supervised fine-tuning is billed per training character, and critically, the fine-tuned model then serves at the same per-token rate as the base model — no inference markup — plus a flat hourly hosting fee per deployed model copy (reported around $1.20/hour). This is a meaningfully different economic model from OpenAI's: your ongoing inference cost doesn't change, so the fine-tuning decision reduces to "is the one-time training cost plus the hosting fee worth it," not "will I be paying a permanent per-token surcharge forever."

Mistral — a flat, low-cost structure: roughly $2/month in model storage plus a small minimum per training job. The real value case Mistral itself highlights: a fine-tuned Small-class model can match a general-purpose Large-class model's quality on a narrow task while billing at the Small model's much lower per-token rate — a genuine order-of-magnitude cost reduction if your task is narrow enough. Mistral also offers a separate, license-fee-based path (Forge) for teams that want to run full training on their own infrastructure rather than paying Mistral for compute.

Anthropic (Claude) — no public fine-tuning API in 2026. The only supported fine-tuning path is Claude Haiku via Amazon Bedrock, not through Anthropic's own API, and Anthropic has been explicit that this is a deliberate product decision rather than a gap they intend to close — their published model-behavior priorities (safety and constitutional constraints ranked above raw helpfulness) are described as something fine-tuning is not meant to override. If your workflow assumes you can eventually fine-tune whichever model you start with, don't assume that about Claude.

DeepSeek and xAI (Grok) — neither publishes an official native fine-tuning API as of this writing. Third-party inference platforms (e.g. Nebius) offer to host and fine-tune DeepSeek's open-weight models, which is a real option since the weights are open — but that's a different arrangement (self-hosted or third-party-hosted training) from a first-party fine-tuning API, and the platform's pricing, not DeepSeek's, is what applies.

The hidden costs beyond the headline training price

The training bill is rarely the whole story. Budget for: the iteration cost of getting a fine-tune right (most teams retrain multiple times before the result beats a well-prompted base model, and each run bills again), evaluation cost (you need a real held-out test set and a way to score it, which is engineering time, not API spend, but it's still cost), storage/hosting fees where a provider charges to keep your custom model deployed (Google's per-model hourly fee, for instance), and data preparation — cleaning and labeling a training set well enough to actually move the needle is usually the single largest line item, and it doesn't show up on any provider's pricing page at all.

When it's not worth it

For most teams, the cheaper path is exhausted before fine-tuning is even considered: a well-structured prompt with a handful of examples, retrieval-augmented generation over your own documents, or simply routing to a stronger base model, all avoid the training bill, the iteration cost, and the operational complexity of maintaining a custom model. Fine-tuning tends to pay off specifically when you have (a) a narrow, repetitive task, (b) a large volume of calls where the per-token savings compound, and (c) a training set good enough that the result reliably beats prompting — which is a real bar, not a given. If you haven't tried prompt engineering and caching first, that's usually the higher-leverage place to start; see the full cost-optimization playbook for the math on each lever.

Practical takeaway

Fine-tuning pricing isn't one thing — Google's zero-markup-plus-hosting-fee model, Mistral's flat low-cost structure, and OpenAI's per-model training rate with a real inference markup (now being phased out for new users) are three genuinely different economic bets, and Anthropic and (as first-party offerings) DeepSeek/xAI aren't on the table at all. Check current availability and terms directly on the provider's own docs before committing a training budget — this space is moving fast enough that "OpenAI is winding fine-tuning down" would have sounded implausible a year earlier.