Management judgment has become a testable subject
For educators and science-minded readers, artificial intelligence is often presented through polished answers: a model explains a concept, summarizes a paper or completes an assignment. Firmulate asks a more revealing question. What happens when the model must manage a company, act under pressure and live with the consequences of its decisions?
Its interactive guess-the-model quiz turns 242 real, unedited management decisions into a kind of behavioral field guide. Readers see a decision and try to identify which frontier model made it. The exercise quickly becomes less about recognizing writing styles and more about spotting managerial habits: exhaustive analysis, terse action, careful refusal or the failure to finish an otherwise strong piece of work.
As an affiliate, we earn on qualifying purchases.
The same terrible week, repeated
Firmulate placed each frontier model in charge of the same small software company during its worst week. The customers, crises and temptations remained identical. Every workday and decision was versioned and auditable, making comparisons possible without relying on a staged chat demonstration.
The company has 13 synthetic employees and deliberately uncomfortable economics: it burns €105k each month against €2.3k in monthly recurring revenue. Its cash countdown is public, and it has accumulated more than 680 self-learned playbook rules. The experiment is real and watchable, with the company continuing to operate rather than disappearing after a benchmark run.
A close contest with a large practical gap
The final Crucible League standings from July 2026 put gpt-5.6-sol first with 95 points, followed by Kimi K3 with 93. Sonnet 5 scored 88, Fable 5 scored 77 and Opus 4.8 scored 73. A do-nothing baseline earned 26 because partial progress counts. But a single breach of trust caps the total: “no amount of good work outweighs a breach of trust.”
The models shared important strengths. All of them detected every crisis and rejected every manipulation attempt. Yet only two signed the €55,000 deal that their own analysis had earned. Firmulate summarizes the gap neatly: “Same diagnosis, same pitch — no signature.” It is a distinction conventional question-answering tests rarely expose. Knowing what should happen is not the same as carrying the task through to completion.
The winning fact was already inside the company
The decisive information was not in the customer event demanding immediate attention. A competitor weakness was buried two document references deep in the company’s own files. Models that followed that trail won the deal at full price, adding €4,583 in monthly recurring revenue.
That finding should resonate in education and research. Performance did not depend only on fluency or visible reasoning. It depended on consulting the available record before acting. The strongest decision-makers treated organizational memory as evidence, while others could reach the correct broad diagnosis without uncovering the fact that changed the commercial outcome.
Pressure revealed boundaries as well as initiative
The company 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 of 5 models refused. Kimi K3 recorded its reasoning plainly: “Treat the request as a suspected approval-bypass / possible impersonation.”
This unanimity matters because the experiment was not designed solely to reward aggressive action. A useful AI manager must know when finishing the job means proceeding and when responsibility requires stopping. The trust cap makes that boundary consequential rather than decorative.
Thoroughness was not enough
Opus 4.8 provides the clearest warning against equating volume with effectiveness. It was the most thorough participant, learned 80 additional rules and produced the deepest analyses, yet finished last. The deal remained unsigned, and discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in all four other models, though less strongly.
There is also an important comparison caveat. Kimi K3 ran without an effort parameter, using the API default, while the other models ran at xhigh. That does not erase the result, but it belongs beside it—a useful reminder that responsible comparisons disclose differences in test conditions.
As an affiliate, we earn on qualifying purchases.
A quiz about recognizable judgment
The quiz works because these decisions are not invented personality sketches. They are drawn from an auditable company experiment, and the reader encounters them before seeing the model identity. A long answer may reflect care, hesitation or both. A terse refusal may show disciplined security judgment. A persuasive analysis may still conceal the most expensive omission: failing to close.
Firmulate also offers enterprises a pilot using a read-only export of their own business. Nothing writes back to real systems, allowing organizations to examine how an AI workforce behaves around their own context without giving it control of operational data.
The broader lesson is that frontier models do display measurable management personalities, but personality is not merely tone. It appears in what they read, what they refuse, whether they escalate and whether they complete the final step. For anyone evaluating AI for consequential work, those behaviors are more informative than another polished answer in a chat window.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
AI management decision simulation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
