Fine-tuning, priced.
OpenAI is closing self-serve fine-tuning — no new training jobs from January 6, 2027 — so the questions are where to train next and whether it pays at all. Enter your dataset and traffic: every managed fine-tuning API is priced for its training runs plus a year of the tuned model, against simply prompting the base model.
Training bills dataset tokens × epochs, every run. The comparison is the same base model without tuning, sent the extra prompt tokens on every request — the trade fine-tuning actually makes. Set saved tokens to 0 to see the pure premium of a tuned endpoint.
- Amazon Nova Micro$6.00/run Bedrock $102/yr
- Amazon Nova Lite$12.00/run Bedrock $165/yr
- Google Gemini 2.5 Flash-Lite$9.00/run Vertex AI $192/yr
- OpenAI gpt-4o-mini$18.00/run OpenAI API $540/yr
- Google Gemini 3.1 Flash-Lite$18.00/run Vertex AI $837/yr
- Amazon Nova 2 Lite$22.68/run Bedrock $894/yr
- Google Gemini 2.5 Flash$30.00/run Vertex AI $900/yr
- Google Gemini 3.5 Flash-Lite$18.00/run Vertex AI $1,242/yr
Fine-tuning rates verified on vendor pages 2026-10-13; base-model rates from our catalog, dated per row in the table below. Batch inference, provisioned throughput and enterprise contracts not modeled.
Six fine-tuning profiles
Click one to load it. Dataset sizes, traffic and saved prompt tokens are scenario assumptions you can edit; the verdicts are computed from the current rate sheet.
Your workload on every managed model
The scenario configured above, priced on all 14 models. "vs prompting" is the first-year cost of the untuned base model with your saved prompt tokens added, minus the tuned model's first year — positive means fine-tuning wins. Rows keep the default order; numbers recompute live.
| # | Model | Train $ / 1M | Tuned in / out | Base in / out | Per run | Tuned / mo | Prompted / mo | First year | vs prompting | Run pays back | Src |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Amazon Nova MicroBedrock Custom models bill on-demand at the base Nova rate, plus monthly storage. US East. | $1.00 | $0.035 / $0.14 | $0.035 / $0.14 | $6.00 | $6.50 | $9.80 | $102 | +$15.60 | 1.8 months | ↗ |
| 2 | Amazon Nova LiteBedrock Custom models bill on-demand at the base Nova rate, plus monthly storage. US East. | $2.00 | $0.06 / $0.24 | $0.06 / $0.24 | $12.00 | $9.75 | $16.80 | $165 | +$36.60 | 1.7 months | ↗ |
| 3 | Google Gemini 2.5 Flash-LiteVertex AI Tuned endpoint bills the base price. | $1.50 | $0.10 / $0.40 | $0.10 / $0.40 | $9.00 | $13.00 | $28.00 | $192 | +$144 | 18 days | ↗ |
| 4 | OpenAI gpt-4o-miniOpenAI API Existing fine-tuning customers only; no new training jobs from 2027-01-06. | $3.00 | $0.30 / $1.20 | $0.15 / $0.60 catalog, verified 2026-10-13 | $18.00 | $39.00 | $42.00 | $540 | −$36.00 | 6.0 months | ↗ |
| 5 | Google Gemini 3.1 Flash-LiteVertex AI Tuned endpoint bills 1.5× the base model. | $3.00 | $0.375 / $2.25 | $0.25 / $1.50 | $18.00 | $63.75 | $80.00 | $837 | +$123 | 1.1 months | ↗ |
| 6 | Amazon Nova 2 LiteBedrock Custom models bill on-demand at the base Nova rate, plus monthly storage. US East. | $3.78 | $0.30 / $2.50 | $0.30 / $2.50 | $22.68 | $66.95 | $110 | $894 | +$426 | 16 days | ↗ |
| 7 | Google Gemini 2.5 FlashVertex AI Tuned endpoint bills the base price. | $5.00 | $0.30 / $2.50 | $0.30 / $2.50 | $30.00 | $65.00 | $110 | $900 | +$420 | 20 days | ↗ |
| 8 | Google Gemini 3.5 Flash-LiteVertex AI Tuned endpoint bills 1.5× the base model. | $3.00 | $0.45 / $3.75 | $0.30 / $2.50 | $18.00 | $97.50 | $110 | $1,242 | +$78.00 | 1.4 months | ↗ |
| 9 | OpenAI gpt-4.1-miniOpenAI API Existing fine-tuning customers only; no new training jobs from 2027-01-06. | $5.00 | $0.80 / $3.20 | $0.40 / $1.60 catalog, verified 2026-10-13 | $30.00 | $104 | $112 | $1,368 | −$24.00 | 3.8 months | ↗ |
| 10 | Amazon Nova ProBedrock Custom models bill on-demand at the base Nova rate, plus monthly storage. US East. | $8.00 | $0.80 / $3.20 | $0.80 / $3.20 | $48.00 | $106 | $224 | $1,463 | +$1,225 | 12 days | ↗ |
| 11 | Google Gemini 2.5 ProVertex AI Tuned endpoint bills the base price. Prompts over 200K tokens bill a higher tier. | $25.00 | $1.25 / $10.00 | $1.25 / $10.00 | $150 | $263 | $450 | $3,750 | +$1,650 | 24 days | ↗ |
| 12 | Google Gemini 3.5 FlashVertex AI Tuned endpoint bills 1.5× the base model. | $10.00 | $2.25 / $13.50 | $1.50 / $9.00 | $60.00 | $383 | $480 | $4,830 | +$930 | 18 days | ↗ |
| 13 | OpenAI gpt-4.1OpenAI API Existing fine-tuning customers only; no new training jobs from 2027-01-06. | $25.00 | $3.00 / $12.00 | $2.00 / $8.00 catalog, verified 2026-10-13 | $150 | $390 | $560 | $5,280 | +$1,440 | 26 days | ↗ |
| 14 | OpenAI gpt-4o (2024-08-06)OpenAI API Existing fine-tuning customers only; no new training jobs from 2027-01-06. | $25.00 | $3.75 / $15.00 | $2.50 / $10.00 catalog, verified 2026-09-06 | $150 | $488 | $700 | $6,450 | +$1,950 | 21 days | ↗ |
Rates per 1M tokens. Vertex base rates are Google Cloud's Global-endpoint list; OpenAI base rates come from our model catalog (dated per row, linked). "Run pays back" is one training run divided by the monthly saving against prompting — "never" when the tuned model costs more to run. Bedrock rows include $1.95/mo storage in "Tuned / mo".
Training an open model on your dataset
What one training run on the dataset above costs on each platform, minimum charges applied. Serving the tuned model is billed differently on each: Fireworks at the base model's serverless price, Together on a dedicated endpoint by the minute — so these rows aren't ranked against the managed models. Verified 2026-10-13.
| Model | Together LoRA | Together full | Fireworks LoRA | Fireworks full | Vertex SFT |
|---|---|---|---|---|---|
| Llama 3.1 8B | $4.00 $0.34/1M · min $4 | $4.00 $0.38/1M · min $4 | $3.00 $0.50/1M | $6.00 $1.00/1M | $4.02 $0.67/1M |
| Llama 3.3 70B | $12.18 $2.03/1M · min $4 | $13.44 $2.24/1M · min $4 | $18.00 $3.00/1M | $36.00 $6.00/1M | $40.32 $6.72/1M |
| gpt-oss-120B | $15.00 $2.50/1M · min $6 | — | $36.00 $6.00/1M | $72.00 $12.00/1M | — |
| DeepSeek-V3.1 | $42.00 $7.00/1M · min $20 | — | $60.00 $10.00/1M | $120 $20.00/1M | — |
| Kimi K2.6 | $90.00 $15.00/1M · min $60 | — | $60.00 $10.00/1M | $120 $20.00/1M | — |
Fireworks prices by size band — up to 16B $0.50 LoRA / $1.00 full; 16.1B–80B $3.00 LoRA / $6.00 full; 80B–300B $6.00 LoRA / $12.00 full; over 300B $10.00 LoRA / $20.00 full per 1M tokens. Sources: Together AI, Fireworks AI, Vertex AI.
The math behind the ranking
Every rate is a pay-as-you-go list price from the vendor's own page, dated. The engine prices two paths for the same year and the same traffic:
training_tokens = examples × tokens_per_example × epochs per_run = training_tokens × training_rate ÷ 1M /* fine-tune */ tuned/mo = requests × (input × tuned_in + output × tuned_out) ÷ 1M + storage first_year = runs_per_year × per_run + 12 × tuned/mo /* keep prompting the base model */ prompted/mo = requests × ((input + saved) × base_in + output × base_out) ÷ 1M vs_prompting = 12 × prompted/mo − first_year /* > 0: fine-tuning wins */ run_pays_back = per_run ÷ (prompted/mo − tuned/mo)
Not modeled: batch inference discounts, cached-input rates, provisioned throughput, validation tokens and per-job minimums on managed APIs, reinforcement fine-tuning billed by the hour, and the quality difference between a tuned model and a prompted one — the reason you'd fine-tune in the first place. Amazon rates are the US East list, read from the AWS Price List API.
Full methodology lives at /methodology/. Found a stale rate? Report it.
Three decisions, real numbers
Computed from the current rate sheet on this page's default workload unless noted.
What the training price doesn't tell you
New organizations lost access on May 7, 2026; organizations without recent fine-tuned inference on July 2, 2026. From January 6, 2027 nobody can start a new training job. Tuned models keep serving until their base model is retired — so a model you train now has a shelf life, and the next retrain happens somewhere else.
Fine-tuning mostly buys back prompt tokens: the instructions and examples you'd otherwise send with every request. That's why the calculator prices the base model with those tokens added. At low traffic the prompt is cheaper than the training; at high traffic the tuned model wins — the "prompt tokens saved" field moves the line.
OpenAI bills a tuned model 1.5–2× its base input rate. On Vertex, tuned Gemini 3 endpoints bill 1.5× the base model while tuned Gemini 2.5 keeps the base price. Amazon bills custom Nova models at the base rate but adds $1.95 a month per model for storage.
Training tokens are dataset tokens × epochs, billed on every run. Three epochs triple the bill; retraining quarterly quadruples the year. A one-off training job is rare — set "training runs / year" to how often your data actually changes.
Training an 8B model on Fireworks starts at $0.50 per 1M tokens, and Fireworks serves LoRA fine-tunes at the base model's price. Together prices training per model, then hosts the tuned model on a dedicated endpoint billed by the minute for as long as it runs — a fixed monthly cost no per-token table shows.
Fine-tuning pricing — answered.
The questions this calculator exists for, answered with the current numbers. Every figure recomputes when a vendor reprices.
Q · 01 How much does it cost to fine-tune GPT-4.1? +
Q · 02 Is OpenAI shutting down fine-tuning? +
Q · 03 How much does Gemini fine-tuning cost? +
Q · 04 How much does it cost to fine-tune Llama? +
Q · 05 How much does fine-tuning on Amazon Bedrock cost? +
Q · 06 Is fine-tuning cheaper than prompting? +
Q · 07 How are fine-tuning training tokens counted? +
Q · 08 What does running a fine-tuned model cost? +
Put this calculator on your site
Free to embed — your readers always see current rates. The widget links back here for the full tables.
The cost toolbox
Every tool runs on the same live-pricing backbone. See the full index →