| Item | Price |
|---|---|
A run that ends ready | 3 × the compute time at its price + $50 |
A run that ends no_gain | $50 |
A run that ends failed or refused | Nothing |
| A run you cancel, or whose model you delete while it runs | 3 × the compute used so far, without the $50; nothing if it had not started |
| Requests to a DecisionNode-1.0 fine-tune | $0.062 per million input tokens: $0.042 (proposed) + $0.02 |
| Requests to a DecisionNode-1.0 Flash fine-tune | $0.055 per million input tokens: $0.035 (proposed) + $0.02 |
| Output, storage, deploying | Free |
- The console shows the estimate before you start, from the dataset's size and the base. The charge is made when the run ends, from the time it took.
- The balance must cover the estimate to start a run; below it the run does not start. The fee is drawn when the run ends.
- Everything is drawn from the prepaid balance, like requests. See Pricing and billing.
A worked example#
A run whose compute costs $0.70 at its price is charged 3 × $0.70 + $50 = $52.10 if it ends ready, $50 if it ends no_gain, and nothing if it fails. Serving it on DecisionNode-1.0 Flash, a million requests of 118 input tokens each (118 million tokens) cost $2.36 more than on the base.
Limits#
| Limit | Value |
|---|---|
| Records per dataset | 200 to 50,000 |
| Dataset file | Up to 64 MiB of JSON Lines; images are uploads, counted apart |
| Images per record | Up to 16, as a live upload_id or inline data; not by url |
Record weight | 0.1 to 10, default 1 |
| Calibration and held-out splits | 10% of the records each, at least 50 |
| Question types | choice, score, truth, number; point and box are not trained |
| Model name | 3 to 40 of a-z, 0-9, -; unique in the workspace |
| Description | Up to 200 characters |
| Fine-tuned models per workspace | 10 |
| Versions deployed at once | 3 |
| Versions kept per model | 5; older ones are archived |
| Runs at a time per workspace | 1 |
| Longest run | 6 hours; longer is stopped, at no charge |
| Dataset retention | 30 days after its last run, unless kept |
| Requests kept for labelling | 30 days each, only while the workspace has the setting on |
| Who can train and deploy | Owners, Admins and Developers; every member can see |
The workspace limits can be raised: write to us with what you need. The request limits of a fine-tuned model are those of its base, listed on Limits and at GET /v1/models.