Compute
Fine-tuning
Managed training of models on your data.
What it is
Managed fine-tuning trains a base model on your examples so it follows your format, tone or domain better, without you managing GPUs. Methods include full fine-tuning, LoRA, and preference optimisation (DPO).
How pricing works
Billed per training token (dataset tokens × epochs). Serving the tuned model may cost more than the base model.
We compare everything in one unit: USD per 1M training tokens.
Typical uses
- Consistent output format or brand voice
- Domain-specific classification/extraction
- Distilling a big model into a cheaper one
What to compare
- Supported base models
- Methods (LoRA, full, DPO)
- Hosted inference price after tuning
Watch out for
- Fine-tuning rarely adds knowledge; use RAG for facts.
- Check inference price of the tuned model, not only training price.
All 5 APIs, cheapest first
Free to browseTick 2 to 4 APIs to compare them side by side.
| API | $ / 1M train tokens | Methods | Download weights | Quality | Price data | Compare |
|---|---|---|---|---|---|---|
| Together fine-tuning (LoRA, ≤16B) Together AI | $0.48 | lora, full, dpo | Yes | 75 | Unverified sample | |
| Fireworks fine-tuning Fireworks AI | $0.50 | lora, rl | Yes | 74 | Verified Oct 4, 2026 | |
| Amazon Bedrock fine-tuning (Nova) Amazon Web Services | $1.00 | lora | No | – | Unverified sample | |
| Mistral fine-tuning Mistral AI | $4.00 | lora | No | – | Unverified sample | |
| GPT-4.1 mini fine-tuning OpenAI | $5.00 | full, dpo | No | 80 | Verified Oct 4, 2026 · aging |