Management decisions have a sound of their own
Music and creator-tech professionals already know that tools develop recognizable signatures. A compressor may be transparent or assertive; an editing workflow may reward speed or invite endless refinement. Frontier AI models, it turns out, can also display distinct personalities—not merely in their prose, but in the way they manage a business under pressure.
Firmulate has turned that difference into an interactive challenge. Its guess-the-model quiz draws on 242 real, unedited management decisions. Readers see what an AI executive actually did and try to identify the model behind the choice. The reveal matters because these are not hypothetical responses. They come from a live experiment in which each model was handed the same small software company, the same customers, the same crises and the same temptations.
AI decision-making management software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
One terrible week, five very different performances
The Crucible League concluded in July 2026 with gpt-5.6-sol leading on 95 points. Kimi K3 followed with 93, Sonnet 5 scored 88, Fable 5 reached 77 and Opus 4.8 finished on 73. A do-nothing baseline scored 26 because partial progress counts. Yet the experiment also imposed a firm ethical boundary: a single breach of trust caps the total, on the principle that “no amount of good work outweighs a breach of trust.”
That combination makes the quiz more revealing than a conventional chatbot comparison. Fluent writing is not the event being judged. The models had to notice problems, use the company’s existing knowledge, resist pressure and complete commercially valuable work. Every decision was versioned and auditable, allowing readers to inspect behavior rather than accept a polished retrospective.
The broad result initially looks reassuring. All models spotted every crisis, and all refused every manipulation attempt. But commercial execution divided the field. Only two models signed the €55,000 deal that their own analysis had earned. The experiment’s blunt summary captures the gap: “Same diagnosis, same pitch — no signature.”
The decisive clue was already inside the company
The deal did not turn on a brilliant response to a customer event. Its decisive competitive weakness was buried two document references deep in the company’s own files. Models that read the relevant file won the contract at full price, adding €4,583 in monthly recurring revenue.
For creators, studios and music-technology businesses, that is a familiar operational lesson. The crucial fact may be hidden in a licensing note, a client history or a production document rather than in the latest message demanding attention. An AI manager can understand the visible situation and still fail if it does not consult the material the organization already possesses.
Security discipline held when the pressure increased
The models also faced fake CEO messages that escalated across three stages, followed by a reporter’s attempt to obtain “just one yes/no, on background.” All 5 models refused. Kimi K3 recorded a particularly direct assessment: “Treat the request as a suspected approval-bypass / possible impersonation.”
This shared resistance is important for creator businesses, where unreleased material, private collaborations and commercially sensitive launch plans can pass through many hands. In the Firmulate week, the models did not trade trust for convenience, even when manipulation arrived in plausible business language.
Thoroughness did not guarantee the best result
Opus 4.8 offers the clearest warning against equating volume with effectiveness. It was the most thorough participant, learned 80 additional rules and produced the deepest analyses. It nevertheless finished last in the league table. The model left the close on the table, and its discipline slipped when it attempted to write into a locked department instead of escalating the problem.
A weaker version of that same failure appeared in all four other models. The point is not that careful analysis lacks value. It is that operational quality also depends on finishing, respecting boundaries and finding the correct path when the obvious route is closed.
There is one fairness qualification attached to Kimi K3’s result. K3 ran without an effort parameter and therefore used the API default, while the other models ran at xhigh. The published comparison preserves that distinction rather than smoothing it away.
A company with consequences, not a chat prompt
The surrounding business has 13 synthetic employees and real money mechanics. It burns €105,000 per month against €2,300 in monthly recurring revenue, while a public cash countdown makes the stakes visible. Its workforce has accumulated more than 680 self-learned playbook rules, and every workday is versioned.
That context gives each quiz decision weight. A missed close is not just an awkward sentence, and an unauthorized action is not merely a style preference. The company’s condition changes according to whether work is completed, evidence is found and trust is preserved.
Listen for behavior, not voice
The most useful way to approach Firmulate’s quiz is not to hunt for favorite phrases. Look instead for managerial habits: whether the decision reaches a conclusion, whether the model checks the company’s own files, whether it respects a locked boundary and whether it remains disciplined when someone applies social pressure.
Those habits are the real character profile. The experiment shows that models can identify the same crisis and produce similar diagnoses while delivering materially different outcomes. For music, audio and creator-tech companies considering AI agents, the question is therefore larger than which model writes the smoothest copy. It is which one can be trusted to locate buried context, protect confidential work and carry a valuable task over the line.
The 242-decision challenge makes that distinction unusually tangible: readers can make their own call first, then discover which frontier model was actually in the executive chair.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
