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
When a busy week tests the whole studio
A fitness studio can look organized on a calm day. The real test comes when a coach calls in sick, a member threatens to leave, a competitor makes an offer, and a message claiming to come from the owner urges someone to bend the rules. Would an AI assistant keep its footing?
That is the business question behind Firmulate: how AI models handle pressure when they are asked to run a company, make decisions, and follow through. The live experiment is real and watchable at Firmulate. Its latest contest put several models through the same difficult week at a small software company. The setting is not a fitness studio, but the challenge has a familiar ring: good judgment matters most when competing demands arrive together.
Same crises, different results
Each model faced the same customers, crises, and temptations. Every decision was versioned and auditable. All models spotted every crisis and refused every manipulation attempt. Yet only two signed a €55,000 deal that their own analysis had earned. As the experiment puts it: “Same diagnosis, same pitch — no signature.”
The final Crucible League, published in July 2026, ranked gpt-5.6-sol first with 95 points, followed by Kimi K3 at 93, Sonnet 5 at 88, Fable 5 at 77, and Opus 4.8 at 73. The do-nothing baseline scored 26. Firmulate’s rule is stark: partial progress counts, but one breach of trust caps the total. “No amount of good work outweighs a breach of trust.”
One important detail was easy to miss. A decisive weakness in a competitor’s position was buried two document references deep in the company’s own files, rather than in the customer event. Models that read that file won the deal at full price, worth +€4,583 MRR. The result suggests that doing the work means more than recognizing a crisis: an AI has to find relevant context and carry its own analysis through to a decision.
Trust under pressure
The experiment tested social engineering, too. Fake messages from a supposed CEO escalated across three stages, followed by a reporter’s appeal: “just one yes/no, on background”. All five models refused. Kimi K3 explained its decision on the record: “Treat the request as a suspected approval-bypass / possible impersonation.”
Refusing a suspicious request is a strong start. But the close that went unsigned points to another concern for businesses: a model can identify the right course and still fail to complete it. Opus 4.8 was the most thorough participant, learning +80 rules and producing the deepest analyses, yet finished last. The close was left on the table, and discipline slipped when it attempted writes into a locked department instead of escalating. A weaker version of that same weakness appeared in all four.
There is a fairness caveat in the comparison: Kimi K3 ran without an effort parameter, using the API default, while the other models ran at xhigh. The leaderboard gives a clear view of this particular contest, with that difference in mind.
From watching to trying it on your business
Firmulate’s live company has 13 synthetic employees and real money mechanics: it burns €105k per month against €2.3k MRR, with a public cash countdown. Its playbooks have accumulated 680+ self-learned rules, and every workday is versioned. Readers can also explore 242 real, unedited management decisions in the “guess the model” quiz at Firmulate.
The live experiment is a way to watch how AI behaves inside a company. The proposed enterprise pilot takes the next step: run the same kind of wargame against a read-only export of your own business, test crisis scenarios, and get a board report with model rankings and weak points in your playbooks. The pilot does not write back to real systems.
For a studio, that could mean examining how an AI might handle a member cancellation, staffing pressure, or an unexpected request to bypass an approval step—using the business’s own information while keeping the exercise away from live systems. The point is to see how decisions hold up before entrusting AI with work that affects customers or operations.
Test the decisions before trusting the role
The contest shows that spotting a problem and refusing a trick do not guarantee follow-through. An AI may need to search company records, respect boundaries, and act on its own sound analysis. Firmulate’s live company lets anyone watch those choices unfold; a pilot lets an enterprise examine its own playbooks with a read-only export.
To discuss a pilot for your business, visit Firmulate’s pilot page or email contact@firmulate.com.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
