Cost a fine-tuning run from published rates: Vertex AI, Together and Fireworks per 1M training tokens, or your own GPU-hours on rented hardware.
Scenarios
Every rate below is read from the vendor's own pricing page.
Vendors bill the training set once per epoch, and the evaluation set once per evaluation.
Providers that publish no per-token rate
Rates read from each vendor's pricing page on August 23, 2026. Fine-tuning prices move; check the vendor before committing budget.
You might also find these calculators useful
Estimate machine learning model training time and cost
Pick your GPU and see which LLMs actually fit
Compare self-hosted GPU vs API inference costs
Parameters, VRAM and GPU fit from a model's real architecture
Fine-tuning is billed on tokens, not on time — and the token count is not the size of your dataset. Every major vendor multiplies the training set by the number of epochs, so a 2-million-token dataset trained for three passes is billed as six million tokens. This calculator applies each vendor's published rate to that number, and covers the other route too: renting GPUs and running the job yourself, where the bill is GPU-hours × cards × hourly rate.
Google, Together and Fireworks all state the same rule: you pay for the tokens the trainer processes, which is your training dataset multiplied by your epoch count. The evaluation set is billed separately and — this is the part most estimates get wrong — it is multiplied by the number of evaluations, not by the number of epochs. Together also applies a minimum charge per job, so a very small experiment costs the floor rather than the arithmetic. Self-hosting swaps the whole model out: you rent accelerators by the hour and pay for every card the run occupies, for the whole run, whether or not the job saturates them.
Billed tokens
Turn a dataset size and an epoch count into a number you can put in a plan, on rates read from the vendor's own page.
The same job priced at each provider that publishes a rate for it, so the spread is visible rather than assumed.
Hosted vendors price the two differently, and self-hosting changes the memory footprint enough to change the card count.
Price the same run both ways — per-token on a managed platform, or GPU-hours on rented accelerators.
If a supplier's figure differs from the published rate applied to your billed tokens, you have a question to ask.
The self-host mode derives how many cards the model needs at your precision, which is what actually multiplies the hourly rate.
Vendors quote a price per million training tokens, and people apply it to the dataset. Three epochs triples the bill before anything else happens.
Charging the eval set once per epoch overstates it on almost every real run. Together states the rule explicitly: validation dataset size × number of evaluations.
Together applies a $4.00 minimum per job, with higher floors on its specialized models. Below that threshold, more data is free.
On Vertex AI, tuning Gemini 2.5 Flash costs $0.005 per million tokens; tuning Qwen 3 14B on the same platform costs $8.46. That is a factor of roughly 1,700 for the same operation.
Google charges 1.5× the base price for tuned Gemini 3 endpoints and the same price for Gemini 2.5 and earlier. Fireworks serves tuned models at base price. There is no single multiplier.
Not as a new user. OpenAI's pricing page states that it is winding down the fine-tuning platform and that it is no longer accessible to new users; existing users can create training jobs for the coming months, and fine-tuned models stay available for inference until their base models are deprecated. The only training price still published is $100 per hour for o4-mini reinforcement fine-tuning, which is an hourly rate rather than a per-token one.
Anthropic publishes no per-token fine-tuning price. Rather than invent one, this calculator lists the providers that publish rates and says plainly which ones do not. Cohere and Mistral are in the same position — neither pricing page carries a fine-tuning rate.
Roughly 1,000 tokens per 750 English words is a serviceable first pass, but tokenizers differ by a wide margin between model families. The token count calculator applies each provider's published characters-per-token ratio, which is the number your bill is actually computed from.
Two to four is the usual range for supervised fine-tuning, and the cost scales linearly with the count. More epochs raise the risk of overfitting, so the right move is usually to watch validation loss rather than to buy more passes.
It can be, because a tuned model needs fewer instructions in every request, and that saving repeats on every call. Whether it pays back depends on request volume and on what the tuned model costs to serve — which varies by vendor, so check the serving multiplier alongside the training cost.
That depends on the model, the dataset and how well the job saturates the hardware, so this calculator takes it as an input rather than guessing. The training time estimator works it out from parameter count, dataset size and GPU throughput; bring its figure back here to turn hours into money.
Full fine-tuning holds optimizer state and gradients alongside the weights, which is several times the memory of inference. When that total exceeds one card, the run is sharded and you pay for every card for the whole job. LoRA avoids most of this by freezing the base weights.