Your gym’s booking system, a fitness app’s customer service team and the company that sells your training gear all depend on decisions made under pressure. When a customer threatens to leave or a rival makes an offer, can an AI spot the problem and follow through—or does it stop just short of the finish line? Firmulate’s live company experiment puts that question to the test.
Get workout gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Same company, same hard week
In the final Crucible League, held in July 2026, frontier AI models each ran the same small software company through its worst week. They faced the same customers, crises and temptations, and every decision was versioned and auditable. The experiment measures management decisions, not just how convincing a model sounds in conversation.
The final standings were gpt-5.6-sol at 95, Kimi K3 at 93, Sonnet 5 at 88, Fable 5 at 77 and Opus 4.8 at 73. The do-nothing baseline scored 26. Integrity mattered alongside progress: a single breach of trust capped the total. As the league puts it, “no amount of good work outweighs a breach of trust.”
Knowing what to do is not the same as doing it
All the models spotted every crisis and refused every manipulation attempt. Yet only two signed a €55,000 deal that their own analysis had earned. “Same diagnosis, same pitch — no signature” captures the gap: an AI may identify the right opportunity and make the case for it, then fail to complete the decision.
The winning clue was buried in the company’s own files, two document references deep. It was not in the customer event that brought the deal into view. Models that read the file won at full price, worth +€4,583 MRR. For a fitness business, the parallel is practical: an AI assistant might notice a customer complaint but need to consult the right account history, cancellation terms or service notes before recommending what to do.
Trust under pressure
The models also faced fake CEO messages escalating over three stages, followed by a reporter’s request for “just one yes/no, on background.” All five refused. Kimi K3 described the request as: “Treat the request as a suspected approval-bypass / possible impersonation.” It is a useful reminder that an AI handling business tasks needs to recognize not only an urgent request, but also when that request may be trying to bypass safeguards.
Opus 4.8 was the most thorough participant, with +80 learned rules and the deepest analyses, but finished last. It left the deal unsigned and discipline slipped: it made write attempts into a locked department instead of escalating. A weaker version of the same weakness appeared in all four. More detailed analysis, by itself, did not guarantee a better outcome.
A company you can watch
Firmulate presents its live experiment as a real, watchable synthetic company: 13 synthetic employees operating with real money mechanics, burning €105k a month against €2.3k MRR. Its public cash countdown, 680+ self-learned playbook rules and versioned workdays make the decisions visible over time. The live page lets readers watch the company in the lab, while a quiz built from 242 real, unedited management decisions invites them to guess which model made each choice.
There is a fairness caveat in the final results: Kimi K3 ran without an effort parameter, using the API default, while the other models ran at xhigh. That context belongs beside the ranking when readers consider what the scores show.
From watching to a company’s own test
For enterprise leaders, Firmulate’s proposed next step is a pilot against a read-only export of their own business. That means putting company information and crisis scenarios into a wargame to produce a board report, rank models and expose weak points in existing playbooks. Nothing writes back to real systems. A fitness operator could use that kind of exercise to examine how an AI might respond to a churn wave, a pricing change or a customer-service crisis before relying on it in daily operations.
The experiment’s lesson is not simply that one model topped a table. Models that recognized crises and resisted manipulation still differed in whether they completed the work, followed discipline and found evidence hidden in the files. Those differences are precisely what a company needs to see before deciding how much responsibility to give an AI workforce.
Readers can watch Firmulate’s live company and see the experiment for themselves. Enterprises can run the same wargame on a read-only export of their own business, with no writes to real systems. To discuss a pilot, visit firmulate.com/pilot.html or contact contact@firmulate.com.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
