Fixed fees, written guarantees
The check is free. The migration is a fixed fee, refunded in full if the model doesn't match. Hosting is optional and priced against what you pay today.
Three steps, priced separately
- Cost check of every option on your volumes
- Fine-tune on your examples, test on a held-out slice
- PASS / FAIL / INCONCLUSIVE report
- We tell you if a cheap API already does the job
- Fee set by data size and task, agreed before work starts
- Half up front, half on delivery
- Full refund if the final model misses the agreed margin
- You keep the weights, eval set and scripts
- OpenAI-compatible endpoint, monitoring and retraining runs
- Flat price per call, fixed for 12 months
- Priced below your cheapest option that passes your eval
- Leave any time and take the model with you
Prices are in US dollars. Invoices go out through Wise or Payoneer. Larger or multi-model engagements are quoted separately.
Your options when a fine-tuned model is retired
You can stay until the deadline, rebuild in-house, or have us migrate it. Here is how they compare.
| ILLATE migration | Rebuild in-house | Stay on the OpenAI fine-tune | |
|---|---|---|---|
| Keeps working after OpenAI's dates | Yes | Yes | No, on their schedule |
| Retrain when labels change | Yes | Yes | No new jobs after 6 Jan 2027 |
| You own the weights | Yes | Yes | No |
| Parity proven before you switch | Written margin, intervals, refund | If you build the eval | n/a |
| Your engineers' time | About 2 h a week | Weeks of ML and infra work | None, until it stops |
| Serving and on-call | Us, or your cloud | Your team runs GPUs | OpenAI |
| Upfront cost | $1,500–3,000, refundable | Engineering salaries | $0 |
About the money
What counts as "doesn't match"?
Before the final run we agree in writing on a test set and a margin, usually 2 or 3 accuracy points against your current model. If the final parity report isn't a PASS against that margin, we refund the whole fee.
Why is the check free?
It costs us a few dollars of GPU time and tells both sides whether the work is worth doing. Most of what we learn in the check carries straight into the migration.
Will hosting save us money?
It depends on volume. Short classification calls are cheap on OpenAI, so at low volume the case is control and continuity, not savings. High-volume or long-output workloads usually save. The cost check gives you the number before you commit.
Can we host it ourselves?
Yes. We deliver the weights and a vLLM deployment for your cloud. You pay only the migration fee.
Get your number first
The cost check and parity check are free.