Which AI model actually saves the lead and books the visitor?

Answering a question well is one job. Doing the next thing, saving the visitor’s details or recording the booking they asked for, is another, and the models differ on it far more than on plain answers. This is the lead-and-booking cut of our November 2026 edition: 24 cases in which a visitor left what was needed and the agent had a tool to act with.

Published by Crescendo Lab · Data from the run of 2026-11-07 · 150 cases, 7 models, two independent AI judges

Three things the table says

1. The gap is in acting, not in answering. On plain questions the seven models sit within a few points of each other. On these 24 cases the spread is 7 to 24. Every model had the same tool, the same instructions and the same visitor message.
2. The GPT models rarely save on the first turn. GPT-5.6 Terra called the save action in 8 of 24 cases and GPT-6.1 Sol in 7. They were not wrong about the facts; they asked for one more contact channel before saving. Bookings, which come with a time attached, they handled (7/8 and 7/8); leads with a single channel they held back on (1/16 and 0/16).
3. The models that act also add. Gemini 4 Argon called the action in every case but passed 21: in 3 cases it saved the lead and then added a claim the business never made, such as “Pro plan includes unlimited forms”. Claude Opus 5.5 (24/24) and Sonnet 5.5 (22/24) acted and stayed closer to the facts, at 4.0% and 4.7% unsupported claims across all cases.

One case, two replies

What this means when you pick a model

• If the point of the conversation is a lead or a booking, model choice matters more than any prompt tweak. The three models that saved more than 19 of 24 cases cost between US$10.32 and US$21.65 per 1,000 conversations at list price.
• If every sentence has to be exactly right, the GPT models made the fewest unsupported claims (0.7% and 1.3%) and were the fastest (2.0 s and 3.2 s median). Pair them with a rule that saves the lead as soon as one contact channel is in, which is what CAAC Go does in production.
• Cheap and eager is not the same as good. Gemini 3.8 Flash acted in 14 cases but passed 12; it returned an empty reply in several lead cases even after a retry.

How to read this

Method, cases and judges are those of the November 2026 edition: the same production agent code for every model, 150 cases across six fictional businesses and seven languages, two judges from different vendors who must both agree. The 24 cases here are the ones whose expected outcome was a saved lead or a booking request. In production CAAC Go also saves an email or phone number a visitor types even when the model does not call the action; that safety net was off for the test, so real capture with the GPT models is higher than the table shows. One run per model: treat differences of one or two cases as noise.