Calculadora
Calculadora Costo Fine-tuning LLM
Calcula costo de fine-tuning — tokens training × tarifa por millón, más uplift por token en inference del modelo custom.
Precios actualizados:
El AITOT LLM Fine-tuning Cost calculator estima coste training + inference uplift para modelos fine-tuned en OpenAI (GPT-4o, GPT-4o-mini, o3), Anthropic Claude (invite-only), Google Vertex (Gemini), y Together AI (LoRA para Llama 4, Qwen, Mistral).
Training cost = training tokens × epochs × per-million rate. OpenAI GPT-4o-mini: $3/M tokens training. Together Llama 4 70B LoRA: $1.20/M. La mayoría de fine-tunes producción corren $50-$500 one-time training.
Toggle epochs (default 3) y volumen inference. Bajo 10M tokens/mes, fine-tuning raramente bate prompts crafted. Sobre 100M con task definition estable, fine-tuned smaller model bate larger model con prompts 3-10×.
Total año 1 · más barato
Fireworks · Llama 4 8B
$248
| Proveedor | Modelo base | Costo training | Inference mensual | Total año 1 |
|---|---|---|---|---|
| Fireworks | Llama 4 8B ≤16B LoRA SFT tier | $8 | $20 | $248 |
| Cohere | Command R | $30 | $48 | $606 |
| OpenAI | GPT-4o mini Stale — OpenAI moved to per-hour training 2026-05; verify pending | $45 | $48 | $621 |
| Mistral | Mistral Small 3 $2/mo hosting per deployed adapter | $45 | $58 | $741 |
| Fireworks | Llama 4 70B 16-80B LoRA SFT tier | $45 | $90 | $1,125 |
| Together | Llama 3.3 70B Legacy v3 line; verify pending 2026-05-18 — no longer top-listed on Together pricing | $75 | $88 | $1,131 |
| OpenAI | GPT-5 mini Stale — OpenAI moved to per-hour training 2026-05; verify pending | $60 | $96 | $1,212 |
| Together | Llama 4 Maverick (LoRA SFT) $16 minimum charge; Maverick = ~70B-class | $120 | $120 | $1,560 |
| OpenAI | o3-mini Stale — OpenAI moved to per-hour training 2026-05; verify pending | $75 | $136 | $1,707 |
| Together | Llama 4 Maverick (LoRA DPO) | $300 | $120 | $1,740 |
| AWS Bedrock | Claude Haiku 4.5 (custom) Provisioned throughput required | $120 | $303 | $3,756 |
| Mistral | Mistral Large 2 | $135 | $564 | $6,903 |
| OpenAI | GPT-4o Stale — OpenAI moved to per-hour training 2026-05; verify pending | $375 | $600 | $7,575 |
Costo training = tokens × epochs × tarifa por millón. Inference usa la tarifa elevada del modelo fine-tuned, siempre superior al base. Total año 1 = training único + 12 meses de inference.
Qué hace esta calculadora
Multi-proveedor
OpenAI fine-tuning, Together LoRA, Vertex tuning, plus estimados self-host.
Training + inference split
Coste training one-time separado de uplift mensual.
Slider epochs
Default 3. Calculadora multiplica training × epochs.
Modelado inference uplift
Fine-tuned cuesta 1.5-3× base. Captura impacto año-1.
Total año-1
Training + 12 meses inference = un número budget.
LoRA vs full fine-tuning
LoRA Together 10× más barato que full FT OpenAI.
Comparación rápida
Fine-tuning 5M training tokens, 50M inference/mes, 3 epochs
| Proveedor | Training Cost | Inference Uplift | Total Año-1 |
|---|---|---|---|
| Together Llama 4 70B (LoRA) | $18 | +$50/mes | $618 |
| OpenAI GPT-4o-mini | $45 | +$120/mes | $1,485 |
| Google Gemini 2.5 Flash tune | $75 | +$150/mes | $1,875 |
| OpenAI GPT-4o | $375 | +$1,200/mes | $14,775 |
| OpenAI o3 | $2,250 | +$3,500/mes | $44,250 |
Año-1 = training + 12 × uplift mensual. Uplift es coste sobre modelo base.
Cómo usar esta calculadora
Calcula coste training + inference uplift para LLMs fine-tuned.
- 1
Entra training tokens
Total tokens en dataset training. 100 examples × 500 tokens = 50k.
- 2
Set epochs
Default 3. Más de 4 típicamente overfittea.
- 3
Estima inference mensual
Cuántos tokens servirá fine-tuned model/mes.
- 4
Compara proveedores
LoRA Together más barato; OpenAI full FT más caro.
Por qué usar esta calculadora
- ✓5 proveedores refrescados mensualmente
- ✓Training + inference split
- ✓Comparación LoRA vs full FT
- ✓Número budget año-1
- ✓Modelado epoch + token
- ✓Sin login