Usage quotas are separate from per-minute and per-hour request limits. See API rate limits for the default limits on uploads, fits, and predictions.
Token cost
Tokens measure computation. They are usage units, not a count of words or characters in your data. For TabPFN-3.5, token usage depends on the selected model and the work needed for your request:- Dataset size: The training rows, rows to predict, and feature columns all contribute to the workload. Rows contribute more than columns, and column-row interaction also contributes to the end token charge.
- Number of estimators:
n_estimatorscontrols how many model passes contribute to the prediction. More estimators use more tokens and can improve prediction quality. - Type of work: Thinking fits and predictions require additional computation. Successful KV cache reuse lowers the prediction charge.
- Minimum charge: Each billable TabPFN-3 or TabPFN-3.5 operation has a minimum charge of 10,000 tokens to cover request overhead.
What consumes tokens
Estimate and monitor usage
Withtabpfn-client, estimate a prediction’s token cost before running it. Given your training and test feature tables, X_train and X_test:
tabpfn-client:
Usage pools
Predictions and Thinking fits share a monthly token budget and a daily cap. Every charged operation counts toward both. There is no separate Thinking fit counter.
Your account may have custom limits; check the Usage page for the values that apply to you. If a request’s estimated cost exceeds either remaining budget, the API returns HTTP 429 with a message indicating which limit was hit and when it resets. This applies to predictions using existing fitted models as well as Thinking fits.
A warning email is sent when either daily or monthly usage crosses 70% of the limit.
When tokens are charged
The API reserves the estimated token cost before starting a billable request. This reservation counts toward both budgets while the work is running. Afterward, the charge is adjusted using the actual model settings and execution details, and any unused reservation is returned. For TabPFN-3.5-Thinking, the reservation covers the selected effort’s maximum fit charge, reduced by a shorter explicit fit budget when one is set. The final charge uses measured fit time, subject to a minimum charge and the reservation cap. Unused tokens are returned. A Thinking fit estimate fromestimate_cost() is the maximum charge for that effort, not the expected cost of the fit.
If the API confirms that computation did not start, the reservation is refunded. A failed or timed-out request can still consume tokens if work started or its execution status is uncertain.
HTTP error codes
Higher limits
Request higher token budgets through the Usage page. Include your expected daily and monthly usage and the types of workloads you plan to run.API rate limits
Per-minute and per-hour limits for uploads, fits, and predictions.
Thinking mode
Configure fit-time optimization and understand thinking parameters.
REST quickstart
Full upload → fit → predict walkthrough.
TabPFN-3 changelog
TabPFN-3 release notes and migration guidance.
Security
Encryption, data isolation, and access controls.