ENFR

AI agent

Most of what your AI feature spends goes on requests it should have refused.

A typed check in front of the expensive call costs $0.018 per 1,000. The call it guards costs $13.159 per 1,000. Both measured on our own stack in September 2026, and both sources are named further down this page.

Two squares drawn to scale: a tiny blue square labelled the guard at $0.018 per 1,000, beside a square 730 times its area labelled the call it guards at $13.159 per 1,000.

Every figure on this page is either something we measured on our own systems or something the vendor publishes, and each one says which. There are no industry averages here, and no customer is named who has not agreed to be.

The bill arrives, and so do the tickets

You shipped the feature and it works. Then two things happen at once: the spend is higher than the estimate, and the model answers when it should have stopped. Swapping in a cheaper model usually trades one problem for the other. Here is what that looked like when we measured it.

The cheap model is not cheap. glm-5.3-flash writes the same artifact for $1.498 per 1,000 calls against gemini-3.5-flash-lite's $13.159 — and obeys a constrained JSON schema one time in three. Measured, first-party, three trials per cell.

Some models do not answer at all. minimax-m3 stopped mid-document after 77.7 seconds. deepseek-v4.1-flash returned an empty string after 71.0 seconds. Three times each. Measured, first-party.

You cannot tell which answers to trust. No calibration curve has been published for these models by the vendor or by anyone else, so any confidence threshold has to be earned on your own data. Primary, vendor documentation.

A flow diagram: requests arrive at a blue block labelled the guard, 70 to 500 milliseconds, $0.018 per 1,000. A thick path leaves it marked stopped here, never reaches the model. A thin path continues to a grey block labelled the expensive call, 14.8 seconds, $13.159 per 1,000.
Fig.
[ Enlarge ]

Step 1 — Name the decision

The question is written once, when the system is designed, not generated per request. Fixed question, varying input. That is what makes an answer comparable to the last one.

Step 2 — Put it in front of the spend

The check runs first and answers in 70 to 500 milliseconds in a single pass, with no text generation. It returns one of three shapes: a choice, a score, or a probability.

Step 3 — Earn the threshold on your data

No published calibration exists, so the cut-off is found on a set of your own cases before anything ships behind it. Skipping this is how a guard becomes confidently wrong.

What each call actually costs

The guard bar is a hairline because the guard costs 730 times less than the call it protects. That ratio is the whole argument.

Cost per 1,000 calls
Measured on our own stack, September 2026, three trials per cell
Table6 rows

What we measured, including what did not work

QuestionWhat we foundGrade
Does a guard pay for itself?$0.018 per 1,000 guard calls against $13.159 per 1,000 for the call it guards — 730 times cheaper.measured, first-party
Is the confidence signal usable?Right answers separate from wrong ones by +0.324 on our own set of 30 cases.measured, first-party
Does asking twice help?Accuracy rises from 92.2% to 96.7%, and the confidence signal collapses from +0.324 to +0.009. We do not ask twice.measured, first-party
Does the wording of the question matter?Asking about tone caught 9 of 10 bad cases but wrongly flagged 2 good ones at 0.84 and 0.73. Asking about provenance flagged none of 20.measured, first-party
Can it write the answer for you?No. It picks among candidates you supply. It cannot generate text and cannot pull a value out of prose.primary, vendor
Is it calibrated out of the box?No. The vendor documents nine named failure modes for this model and publishes no calibration curve.primary, vendor
Questions5 answered

Before you book

The five questions that come up most often on this call.

01Is this just prompt engineering?
No. The questions are written once, at design time, and typed — a choice, a score or a probability. The model cannot return anything else, which is what makes one answer comparable to the next.
02Will a second model call just add latency?
The check answers in 70 to 500 milliseconds in a single pass, with no generation loop. Batching 13 questions into one call measured 10 times faster than 13 separate calls, and 12.2 times cheaper, with no change in the answers.
03What happens when the guard is wrong?
It will be. The vendor documents nine named failure modes for this model, including counting, dates and literal reading. That is exactly why the threshold is earned on your own cases before anything ships behind it.
04Do I have to rebuild my stack?
No. It goes in front of a call you already make. On our own systems it is four call sites, all on the main branch.
05What does it cost to find out?
A 20-minute call. We will tell you whether your spend has a guardable shape — some do not, and we will say so.