Pricing and limits
Which request uses each rate, how caching is billed, and how your credit balance works.
Pricing depends on what you submit and whether AI input extraction is needed. Both the console and API use the same rates and draw from the same credit balance. Prices are in USD.
| Request or work | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Full model definition plus matching JSON state | $0.042 | Free |
| Questions plus state, including cached model reuse | $4 | $20 |
| Create or revise a model using AI | $4 | $20 |
| AI extraction of inputs from text or other supplied data | $4 | $20 |
Use the lower execution rate
Keep the complete decision model returned by aityx. Submit that model with JSON that satisfies its input schema. The engine executes directly, without a reader or generator. System One input billing includes the model and input you send; answers and the execution receipt are returned with output free.
Sending questions again uses System Two pricing. This remains true when the service finds the model in its temporary cache. A cache hit can be faster without making the call cheaper.
Text and mixed inputs
If required facts need to be extracted from your text or JSON, the AI reader runs before the decision engine. That extraction is an additional System Two charge based on the reader’s actual provider input and output tokens. The underlying execution still uses the rate selected by its request: System One for a supplied model, System Two for questions. The receipt separates supplied inputs from extracted inputs.
Matching JSON alone is not enough to select the lower rate when you are sending questions rather than the complete model. The request format and input extraction both matter.
Read the charge, not just provider usage
usage counts API request and response tokens. Model content contributes to input; accounting metadata is
excluded from output token counts. provider_usage reports actual AI work, which can be zero on a cache hit
even though a questions-based call is billable. billing.items separates execution and extraction charges.
The recorded matching-JSON invoice call used 688 input tokens, no provider calls and free output. Its charge was $0.000028896 at $0.042 per million input tokens. The console may show tiny charges as less than $0.001; that display does not make the call free.
Manual validation (content and output sent to System Two without a prompt or files) is free. AI
generation and revision are billable.
Your available balance
Approved preview accounts receive credits. Billable console and API calls consume those credits. The console shows the available balance, usage and the charge for each call. Cached questions-based calls also deduct at the System Two rate.
Billable requests reserve funds before work starts, settle the completed charge and release unused
reservations. A positive balance may still be too small for a request’s reservation, returning
402 insufficient_balance. Failed work releases its reservation. Insufficient available funds prevent
further billable calls. Preview credits are granted through the access
programme; the preview does not require a public checkout flow.
The billing ledger retains credits, charges and adjustments. It is separate from models, inputs and execution receipts, which are returned for you to keep.
Working within the preview
Keep requests focused on the facts and rules needed for the decision. Download models and receipts before leaving the console session. Use your own saved inputs to test revised rules.
For preview capacity and volume pricing, contact [email protected].