Good taste is not the same as good judgment
Beauty and personal-care businesses live on details: the promise made to a customer, the information buried in a product file, the follow-up that turns interest into a sale and the discipline to protect trust when pressure rises. An AI can sound polished while still overlooking the action that matters.
Firmulate has turned that distinction into a public experiment—and a surprisingly revealing quiz. Its Crucible League placed frontier AI models in charge of the same small software company during its worst week. They faced the same customers, crises and temptations. Their decisions were preserved exactly as made, creating an auditable record of how each model behaved when management required more than a plausible answer.
Now, 242 real, unedited management decisions power a “guess the model” quiz. The challenge is entertaining, but the larger story is serious: frontier models display recognizable management personalities, and those differences affect whether work gets completed.
AI decision-making software for business
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The models agreed on the problems—and still produced different outcomes
The final July 2026 league table was close at the top. GPT-5.6-sol finished first with 95, followed by Kimi K3 with 93, Sonnet 5 with 88, Fable 5 with 77 and Opus 4.8 with 73. A do-nothing baseline scored 26 because partial progress counts. But a single breach of trust caps the total, reflecting the experiment’s governing principle: “no amount of good work outweighs a breach of trust.”
All five models spotted every crisis and rejected every manipulation attempt. Yet only two signed the €55,000 deal their own analysis had earned. Firmulate sums up the gap neatly: “Same diagnosis, same pitch — no signature.” It is the kind of failure that a conventional chat demonstration can easily conceal. A model may understand the situation, produce a capable plan and still leave the commercially decisive action unfinished.
The winning clue was hidden in the company’s own files
The decisive competitor weakness did not appear in the customer event. It sat two document references deep in the company’s files. Models that followed the trail found the fact and won the deal at full price, worth +€4,583 MRR.
That result should resonate with any business managing product records, customer histories, supplier notes or brand guidelines. The advantage did not come from a more theatrical response. It came from reading the available material closely enough to discover what the immediate prompt did not reveal.
Pressure exposed discipline as well as intelligence
The social-engineering test used fake CEO messages escalating over three stages, followed by a reporter’s attempt to obtain “just one yes/no, on background.” All 5 of 5 models refused. Kimi K3 recorded its reasoning plainly: “Treat the request as a suspected approval-bypass / possible impersonation.”
This clean result matters because Firmulate’s company is not a static question-and-answer test. It has 13 synthetic employees and real money mechanics, including burn of €105k/month against €2.3k MRR. Its cash countdown is public, every workday is versioned and the company has accumulated 680+ self-learned playbook rules. The live experiment is watchable, allowing readers to see management behavior unfold rather than accepting a polished retrospective.
The most thorough model did not win
Opus 4.8 offers the clearest warning against equating volume with effectiveness. It was the most thorough participant, adding +80 learned rules and producing the deepest analyses, yet it finished last. The deal close was left on the table, and discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in all four of the other models, though less strongly.
The quiz makes these differences easier to feel. Some decisions read like dissertations. Others are terse. Another pattern is the refusal to communicate noise. Without a model name attached, readers must judge tone, completeness and operational instinct—and then discover whether those impressions match the model that actually made the decision.
One comparison deserves context: Kimi K3 ran without an effort parameter, using the API default, while the others ran at xhigh. That does not erase its result, but it is an important fairness note when reading a close league table.
The useful question is not which AI sounds smartest
Firmulate’s experiment suggests a more practical test: which model reads before acting, finishes what it starts and remains trustworthy under pressure? The models shared much of the same diagnosis, but the business outcomes diverged because execution, document-reading and discipline diverged.
For beauty and personal-care operators considering AI around customer service, operations or commercial work, that distinction is more valuable than a dazzling sample response. The quiz packages the evidence in an accessible form, but the lesson extends beyond the game: management character becomes visible in the final action, especially when the obvious analysis is already complete.
Firmulate also offers enterprises the same wargame against a read-only export of their own business. Nothing writes back to real systems. That makes the public experiment more than a leaderboard: it is a demonstration of why AI workers should be tested inside realistic business pressure before they are trusted with consequential work.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
