RFT Cost Planning

RFT cost depends on the selected model and compatible training shape, rollout volume, generated tokens, evaluator latency, and total runtime. These inputs change as models, shapes, and pricing evolve, so this page does not provide a hardcoded dollar calculator.

Do not estimate cost from model size alone. Confirm that the live Training Shapes matrix supports RFT for the selected model and tuning mode, then use current pricing and the resolved shape.

Before estimating

Choose an RFT-enabled model on Models.
Resolve the shape and parameter mode used by the job.
Read current pricing, including any eligible managed RFT promotion.
Record dataset rows, epochs, rollout candidates, maximum output tokens, expected average output length, evaluator latency, and concurrency.
Set a cost ceiling and label any unknown line item rather than guessing.

Cost drivers

Driver	Why it matters	User control
Model and compatible shape	Determines available hardware and runtime characteristics	Choose from the live method-support matrix
Dataset rows	Determines prompt groups processed	Curate representative prompts
Epochs	Repeats the dataset	Start with one
Rollout candidates	Multiplies sampled completions per prompt	Use the minimum that gives useful reward variance
Output length	Drives rollout generation and evaluation time	Set a realistic maximum
Evaluator latency	Slow evaluation increases wall-clock runtime	Cache, batch, and remove unnecessary external calls
Trainer and deployment runtime	Dedicated resources are metered by runtime	Monitor progress and tear down promptly

Planning formulas

Use formulas to expose assumptions, not to replace current platform pricing:

rollout_count = prompts × epochs × candidates_per_prompt
estimated_output_tokens = rollout_count × average_output_tokens
runtime_cost = measured_or_estimated_runtime × resolved_resource_rate

For a first run, estimate a range using conservative and expected output lengths. If throughput or a resource rate is unavailable, mark the total as incomplete and keep a hard runtime or budget ceiling.

Reduce cost safely

Validate the evaluator offline and confirm that representative outputs receive different scores.
Start with a small RFT-compatible model when the live matrix and pricing support it.
Run a bounded dataset sample before the full dataset.
Start with one epoch.
Reduce rollout candidates only if reward variance remains sufficient.
Set output limits to the task’s real needs.
Stop runs that show no real progress or a constant reward.
Delete or scale down billable deployments after evaluation.

Managed versus Training API

Managed RFT: use the current managed pricing and any eligibility rules shown on the pricing and model-support pages.
Training API serverless: verify current per-token meter definitions and rates during private preview.
Training API dedicated: use the resolved trainer and deployment resources, current runtime rates, and measured or bounded duration.

For broader serverless-versus-dedicated economics, use Choose Serverless or Dedicated Training and the multi-turn cost comparison.

Get Started

Serverless

Deployments

Models & Inference

Training

Fire Pass

Fireworks Nexus

Administration

Security & Compliance

Integrations

Reference

Before estimating

Cost drivers

Planning formulas

Reduce cost safely

Managed versus Training API

​Before estimating

​Cost drivers

​Planning formulas

​Reduce cost safely

​Managed versus Training API

Before estimating

Cost drivers

Planning formulas

Reduce cost safely

Managed versus Training API