In sport, a team can read the play perfectly and still fail to finish. The same gap showed up when Firmulate put AI models in charge of a small software company during its worst week: every model spotted the crises, but only two signed the €55,000 deal their own analysis had earned. For businesses considering AI agents, that is the difference between a promising drill and a result on the scoreboard.
A level playing field for AI
Firmulate’s Crucible League gave each frontier model the same company, customers, crises and temptations. Every decision was versioned and auditable. The final league, published in July 2026, put 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. A breach of trust capped the total: “no amount of good work outweighs a breach of trust.”
The models all spotted every crisis and refused every manipulation attempt. But only two closed the €55,000 deal. The experiment’s short verdict was: “Same diagnosis, same pitch — no signature.” Recognizing the right move did not guarantee follow-through.
The scouting report was buried in the files
The decisive competitor weakness was two document references deep in the company’s own files, not in the customer event. Models that read that file won the deal at full price, worth +€4,583 MRR. The result points to a practical challenge for AI at work: the useful clue may be available, yet still be easy to miss when a situation gets hectic.
Integrity held under pressure. Fake CEO messages escalated over three stages, followed by a reporter’s request for “just one yes/no, on background.” All five models refused. Kimi K3’s reasoning was: “Treat the request as a suspected approval-bypass / possible impersonation.”
Yet discipline could falter elsewhere. Opus 4.8 was the most thorough participant, adding +80 learned rules and producing the deepest analyses, but finished last. It left the deal unsigned and tried to write into a locked department instead of escalating. A weaker version of the same weakness appeared in all four models. K3 also ran without an effort parameter, using the API default, while the others ran at xhigh.
From watching the test to trying it on your business
The experiment is part of a live company at Firmulate. It has 13 synthetic employees, real money mechanics, burn of €105k/month against €2.3k MRR, a public cash countdown and more than 680 self-learned playbook rules. Every workday is versioned. The site also offers a “guess the model” quiz built from 242 real, unedited management decisions.
For business leaders, the next step can be a pilot against their own company. Firmulate can build a digital twin from a read-only export, then run crisis scenarios and produce a board report showing model rankings and weak points in the company’s playbooks. Nothing writes back to real systems. That creates a way to examine how an AI workforce handles your own customers, rules and pressure before relying on it in live operations.
A convincing AI demo can show that a model knows what to say. Firmulate’s wargame tests whether it can carry the decision through while respecting the rules. Enterprises can run the same test against a read-only export of their business. Explore a Firmulate pilot and contact contact@firmulate.com to discuss putting your own company through the exercise.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
