Token cap per request
LiveA 200,000-token ceiling on any single model call, and a 150,000-character ceiling on the text sent for contract analysis.
- ✓200,000-token ceiling on any single model call, enforced by refusal
- ✓150,000-character ceiling on the text sent for contract analysis
- ✓Head-slice truncation — not section-aware, and not reported in the response
- ✓Output bounded at the model's wire limit
What it is.
No single model call may exceed 200,000 input tokens. A call estimated over that is refused rather than silently trimmed, which keeps a runaway input from becoming a runaway bill. Output is bounded separately at the model's own wire limit.
Contract analysis has its own ceiling: the text sent for analysis is capped at 150,000 characters, and the cut is a plain head slice — the first 150,000 characters, not a section-aware selection. On a very long contract that means late-document clauses can fall outside the analysed text, and governing law and dispute resolution are exactly the clauses that tend to live at the end. The truncation is logged on our side and is not currently reported back in the response. We would rather you knew that than assumed the whole document was read.
Three steps.
End to end.
A call estimated above 200,000 input tokens is refused, not trimmed.
Contract analysis sends the first 150,000 characters of the extracted text.
For a very long contract, split it into logical parts and analyse each, so nothing important is silently outside the window.
What you get.
- ✓200,000-token ceiling on any single model call, enforced by refusal
- ✓150,000-character ceiling on the text sent for contract analysis
- ✓Head-slice truncation — not section-aware, and not reported in the response
- ✓Output bounded at the model's wire limit
Quick answers.
The first 150,000 characters are analysed and the rest is not, without a notice on the report. Split it into parts, or use bulk analysis with the document already divided, so that every section is inside a window.
No. These are product-wide constants, not per-organisation settings. Your monthly token allowance is the thing that can be raised.
More in AI Controls & Guardrails.
We do not train models on customer content. There is no opt-in, because there is no training programme.
User-pasted text is sanitised before reaching the LLM.
A monthly AI token quota per plan, an 80% warning by email and in-app, and a hard stop at 100%.
Lower the model ceiling, switch AI features off, cap daily spend, and read the AI audit history and security events.