Skip to main content
Together AI bills fine-tuning by the total number of tokens processed across training and validation. The per-token rate depends on three factors: the model size bracket, the training method (supervised or DPO), and the implementation (LoRA or full fine-tuning). For current rates, see the together.ai/pricing. After training, hosting on a dedicated endpoint is billed separately by the minute.

​
How tokens are counted

The total tokens for a job is: 
total_tokens = (n_epochs × tokens_per_training_dataset) + (n_evals × tokens_per_validation_dataset)
Tokenization runs as part of the job. The final token count and price are recorded once tokenization completes and appear on the fine-tuning jobs dashboard and in client.fine_tuning.retrieve(id=<JOB_ID>). If you disable packing, training tokens are computed as dataset_length × max_seq_length instead.

​
Estimate a job cost

To approximate the cost of a fine-tuning job before launching:
  1. Estimate training tokens: context_length × batch_size × steps × epochs.
  2. Add validation tokens: validation_dataset_size × n_evals.
  3. Multiply by the per-token rate for your model size, method, and implementation.
For example, fine-tuning a 16B-class model with context_length=8192, batch_size=8, steps=1000, and 1 epoch processes roughly 65.5M training tokens. A validation file of 1M tokens with n_evals=10 adds another 10M. Multiply the total by the LoRA SFT rate for the 16B bracket.

​
Cancelled jobs

When you cancel a running job, you pay for completed steps only. To check progress before cancelling, run: 
client.fine_tuning.retrieve("<JOB_ID>").total_steps

​
Hosting charges

After training, your fine-tuned model can be served on a dedicated endpoint that bills per minute based on the hardware attached. These charges are separate from your fine-tuning job cost and continue until you stop or delete the endpoint. See deployment for the full setup and teardown flow.

​
Minimum spend

There is no minimum charge for a fine-tuning job. You pay only for the tokens processed.