Prices,
in writing.
Every line below maps to a step on the AI Model Training page. Pilots are fixed price. Programs are quoted after the scoping week. Retainers are monthly. All prices in EUR, excluding VAT, valid until 31 March 2027.
Heights to scale with price. EUR, ex VAT.
How much does it cost to fine-tune an LLM on company data?
The training run itself is cheap: at public self-serve prices, a few dollars to a few hundred dollars for a typical company dataset. Most of the cost is the work around it: preparing the data, building a test set and measuring the model. With us that is in a fixed price: evaluation €1,900, Pilot S €3,200, Pilot M €9,000, programs from €15,000, excluding VAT.
Where the money goes
| Part of the cost | On a self-serve platform | With us |
|---|---|---|
| Training run | Billed per token processed (dataset size × epochs). See the worked example below. | Included in every pilot and in the retainer cadence. GPU time for continued pre-training on large corpora is estimated in the program quote. |
| Data work | Your team: exporting, cleaning, removing duplicates, anonymising, checking licences. Does not appear on the invoice. | Pilot S assumes the data is ready. Pilot M (€9,000) includes building, translating or distilling the training data. |
| Evaluation | Your team builds the test set; the platform reports training loss and bills evaluation tokens. | Frozen test set with gold answers, base vs tuned scoreboard, written recommendation. On its own: €1,900. Another use case: €2,500. |
| Deployment | Served on the platform and billed per call or per endpoint. Some platforms let you download the weights, others (OpenAI) do not; OpenAI is also closing fine-tuning to new users. | GGUF or vLLM build handed over, running on your server. EU hosting if you want it, from €600 a month. |
| Retraining | Each new run is billed again, plus your team's time for data and testing. | Watch €190 a month (monthly measurement) plus €1,490 per retraining cycle, or a retainer with a fixed cadence from €690 a month. |
Market reference: public per-token training prices
List prices for supervised fine-tuning, in USD per million training tokens, checked 3 Oct 2026 on together.ai/pricing and OpenAI API pricing. Worked example: 2,000 examples of 1,500 tokens each, three epochs, is 9 million training tokens.
| Provider and model class | Training, per 1M tokens | 9M tokens | After training |
|---|---|---|---|
| Together AI, 0.8B–9B | $0.34 (minimum charge $4.00) | $4.00 (minimum applies) | Download the weights or serve on a dedicated endpoint |
| Together AI, 27B–35B | $1.05 | $9.45 | Download the weights or serve on a dedicated endpoint |
| OpenAI, gpt-4.1-mini | $5.00 | $45.00 | Tuned model runs only on OpenAI: $0.80 input / $3.20 output per 1M tokens |
| OpenAI, gpt-4.1 | $25.00 | $225.00 | Tuned model runs only on OpenAI: $3.00 input / $12.00 output per 1M tokens |
These prices are fair for what they cover. They cover the run. The test set, the cleaned data and the answer to "is the tuned model better than what we had" are not in that number.
OpenAI is winding down its fine-tuning platform: new users can no longer start, existing users can train for a few more months. Tuned models stay available until their base model is retired.
What you keep
- Model weights, or the LoRA adapter plus the base it belongs to
- The training dataset in a documented format
- The frozen test set with gold answers, so you can measure the next version yourself
- Model card and a record of where every piece of data came from
When a self-serve platform is the better choice
- Your data is already clean, licensed and in one place
- Someone in your team can build a test set and judge the results
- The data may go to the platform provider, and the model may stay there
- Unsure? The break-even calculator compares the API path with your own model, and the sample evaluation report shows what the measuring part looks like.
Questions about the cost
How much does it cost to fine-tune an LLM on company data?
At public self-serve prices the training run itself costs a few dollars to a few hundred dollars for a dataset of a few million tokens. A complete pilot with us, with test set, evaluation and deployment on your server, costs €3,200 (2–3 weeks), or €9,000 if we also build the training data. EUR, excluding VAT.
Why is a pilot more expensive than the per-token training price?
The per-token price covers only the compute. Preparing the data, building a frozen test set and measuring the tuned model against the base is most of the work, and on a self-serve platform your team does it.
Is GPU time included in your prices?
Yes for pilots and for the retraining cadence of each retainer. Continued pre-training on large corpora is estimated separately in the program quote.
What do we own at the end?
The weights or adapter, the training dataset, the frozen test set with gold answers, the model card and the record of where the data came from.
When is a self-serve fine-tuning platform cheaper?
When your data is clean, someone on your team can build a test set and judge the results, and the data may go to the platform provider. Then you pay only for the run.
LLM evaluation on your documents. €1,900 fixed · one day
Before we train anything: does the model you have in mind (a public Slovene model, an open base, a foreign API) actually work on your documents? You send fifty anonymised examples and one task. We build a frozen set with gold answers, run up to five models through the same instruction and return a table of measured results, verbatim outputs and a written recommendation: which model, whether training is needed at all, and what it would buy you. The fee is credited against a pilot.
- Frozen set from your examples, gold by construction
- Up to five models: public Slovene, open bases, your candidate, a foreign API
- Deterministic scoring where possible; a judge for free text
- Results table, time per document, where each one runs
- Written recommendation: train or not, and what
- Under NDA, data stays where agreed
A one-week trial under NDA comes first.
You pick fifty anonymised examples and one task. Within a week we return a model running on your server, before-and-after numbers and a written recommendation. The trial is billed as Pilot S and credited against everything that follows. If the result is not good, it stops there.
Fixed scope, fixed price, one number at the end.
Both pilots end with a base-vs-tuned scoreboard on a frozen eval and a written recommendation. Pick by how much data you already have.
Every pilot includes all three: evaluation on your frozen set, deployment on your side (GGUF or vLLM, hand-over call) and knowledge transfer so you can run and re-train the model yourselves. The training run itself is the smallest part.
Pilot S
Supervised fine-tuning on data you already have. One task, one model, one frozen eval, one report.
- Scoping call + data and licence audit
- Frozen eval set built from your examples
- One fine-tuned open-weight model (4B–12B class)
- Base-vs-tuned scoreboard, GGUF build, model card
- Written go/no-go recommendation
- 2–3 weeks · prepaid
Pilot M
Everything in Pilot S, plus the dataset engineering: building, translating or distilling the training data, and an eval set from scratch.
- Scoping week + data audit on site or remote
- Dataset build: cleaning, dedup, quality scoring, licence gating
- Translation or teacher distillation where data is missing
- Base vs tuned vs the API or RAG setup you were considering
- Deployment recipe (GGUF / vLLM), hand-over call
- 4–5 weeks · 50 % on signing, 50 % on delivery
Starting from one of our published domain models (Slovenian, biomedical, cybersecurity) takes €2,000 off Pilot M.
Programs from €15,000, quoted after the scoping week.
Fine-tuning, evaluation and deployment for a production model, with continued pre-training where the domain needs it. Most programs land in the lower half of the range; the table shows what moves the price. The scoping week is billed as Pilot S and credited against the program.
| What moves the price | Lower end | Upper end |
|---|---|---|
| Price | from €15,000 | up to €90,000 |
| Duration | 6–8 weeks | 3–4 months |
| Training stages | SFT only, one or two re-tunes | CPT on hundreds of millions of tokens, then SFT, then re-tunes |
| Data readiness | Clean, licensed, in one place | Scattered, needs translation, distillation or OCR |
| Model size | 4B–12B, single GPU serving | 27B–35B MoE, multi-GPU or on-prem cluster |
| Evaluation | One objective track | Objective + multi-judge tracks, several use cases |
| Deployment | GGUF hand-over | vLLM on-prem install, cascade router, air-gapped delivery |
| Payment | 40 % on signing · 40 % at the mid-program eval gate · 20 % on delivery. GPU time for CPT billed separately, estimated in the quote. | |
The model stays current. The scoreboard is the contract.
New data in, retrain, re-score, ship. Base-model upgrades when a better open model wins on your eval. Every version of the weights is yours. If there is no new data, Watch is enough: we measure the model every month and train only when needed, at €1,490 per cycle.
| Watch | Lite | Core | Scale | Enterprise | |
|---|---|---|---|---|---|
| Per month | €190 | €690 | €1,690 | €4,990 | from €13,000 |
| Models under retainer | 1 | 1 | 1 | up to 3 | as scoped |
| Retrain cadence | none; a cycle on demand €1,490 | quarterly | monthly | monthly, or on data threshold | weekly possible |
| Base-model upgrade | notice when a better base wins on your eval | once a year | next cycle after release | within days of release | within days, with rollback plan |
| Evaluation | frozen set, re-scored monthly, regression alerts | frozen set, re-scored each retrain | frozen set + regression alerts | + new eval slices as use cases grow | + multi-judge track, compliance re-run |
| Cascade router | — | — | — | maintained (small model first, frontier fallback) | maintained + cost reporting |
| Deployment support | e-mail, 5 business days | e-mail, 5 business days | e-mail, 2 business days | e-mail + chat, next business day | on-prem / air-gapped, response SLA |
| Documentation refresh | model card with every cycle | model card per version | model card per version | model + data card per version | + AI Act record pack per version |
| Term | 12 months, billed monthly. 10 % off when the year is prepaid. Included GPU time covers the retrain cadence above; CPT-scale runs are quoted separately. | ||||
Priced separately, so the retainer stays about the model.
Most customers run the model on their own hardware. For those who don't, we host it on capacity we control inside the EU. Sized by traffic, never by hardware.
Hosted endpoint · EU
- Light€600 / month — internal tools, a few users
- Standard€1,200 / month — customer-facing, business hours
- High€2,000 / month — 24/7, sustained traffic
- IncludesvLLM or llama.cpp serving, TLS, monitoring, monthly usage report
- ExcludesFrontier-API fallback calls in a cascade (passed through at cost)
GPU time & add-ons
- GPU timeOn request, blocks of 100 GPU-hours, for teams with their own pipeline
- Extra eval set€2,500 — a new frozen benchmark for an additional use case
- On-site day€1,500 + travel — install, workshop or air-gapped hand-over
- Dataset onlyQuoted per corpus — build, translate or distil without training
- SpeechASR / TTS fine-tunes priced as Pilot M or program, by hours of audio
Own model or API?
First: may the data go to an API provider at all? If it may, volume decides. The break-even calculator puts your requests, tokens and API price next to the packages on this page, accounts for the cascade, hosting or your own hardware, and shows the month in which the two paths cross. It runs in your browser; nothing is stored or sent.
Terms
- All prices in EUR, excluding VAT. Valid until 31 March 2027; this page carries a version, quotes reference it.
- NDA and data-processing agreement signed before any data is transferred. Data stays inside the EU.
- You own the weights, adapters, datasets, eval sets and provenance record of everything delivered, as described on the training page.
- Pilots are prepaid (S) or 50/50 (M). Programs 40/40/20 on milestones. Retainers monthly in advance, 12-month term.
- Open-weight bases only; we flag any base or data licence that is not commercially usable before work starts.
How to read this page
- Not sure which pilot? Send fifty real examples of the task. We answer within two business days with S or M and why.
- Already paying for a frontier API? Bring the monthly bill. The retainer should sit well under what the small model saves.
- Regulated data? Ask about the air-gapped rung. Same prices, plus on-site days.
- Slovenian version of this page: slovensko.
Pick a pilot. Get a number.
Thirty minutes to scope it. A dated schedule after that. Bring fifty examples and, if you have one, your current API bill.