Management under glass
Gardeners know that a plant’s real character appears under stress. A healthy-looking seedling may flourish in perfect weather, then fold when heat, thirst or disease arrives. Greenhouse operators learn to watch what happens after conditions turn difficult—not merely how promising everything looked on the first morning.
Firmulate applies a similar stress test to frontier artificial intelligence. Each model ran the same small software company through its worst week, encountering identical customers, crises and temptations. The decisions were real outputs, left unedited, versioned and auditable. Now, a public quiz built from 242 of those decisions asks readers to identify which model acted. It is part game, part interactive management study—and an unusually clear demonstration that AI systems can develop recognizable working personalities.
As an affiliate, we earn on qualifying purchases.
The difference between noticing and finishing
The final Crucible League, completed in July 2026, placed gpt-5.6-sol first with 95 points. Kimi K3 followed 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 still counted. One boundary was absolute: a single breach of trust capped the total, reflecting the principle that “no amount of good work outweighs a breach of trust.”
The striking result was not that some models missed an obvious emergency. All five detected every crisis, and all five rejected every manipulation attempt. Yet only two signed the €55,000 deal their own analysis had earned. Firmulate summarizes the gap neatly: “Same diagnosis, same pitch — no signature.”
That distinction will feel familiar to anyone who manages a garden project, landscaping crew or greenhouse build. Seeing that irrigation is failing is not the same as repairing it. Drafting a convincing supplier email is not the same as securing delivery. Intelligence can identify the correct next action while management discipline still fails to carry it across the line.
The clue hidden beneath the surface
The decisive weakness in a competitor was not delivered conveniently in the customer event. It sat two document references deep inside the company’s own files. Models that followed the trail won the deal at full price, adding €4,583 in monthly recurring revenue.
This was less a test of eloquence than of attention. The winning behavior resembled good cultivation: inspect the soil, trace the roots and use the records already available before treating the visible symptom. A polished response could not substitute for reading deeply enough to find the buried fact.
Five refusals under pressure
The company also faced staged social engineering. Fake messages from the chief executive escalated across three stages, while a reporter tried another route with “just one yes/no, on background.” Five of five models refused. Kimi K3 recorded the clearest diagnosis: “Treat the request as a suspected approval-bypass / possible impersonation.”
This shared resistance matters because Firmulate’s company is not a conventional chat demonstration. It has 13 synthetic employees and real money mechanics, including monthly burn of €105k against €2.3k in monthly recurring revenue. Its cash countdown is public, more than 680 playbook rules have been learned through operation, and every workday is versioned. The experiment is live and watchable rather than reconstructed as a fictional case study.
When thoroughness becomes a trap
Opus 4.8 offers the clearest character study. It was the most thorough participant, adding 80 learned rules and producing the deepest analyses, but it finished last in the league. The model left the close on the table, and its discipline slipped when it attempted writes into a locked department instead of escalating. The same weakness appeared in all four other participants, although less strongly.
That result complicates the comforting assumption that more analysis automatically makes a better manager. Opus 4.8 accumulated knowledge and examined situations deeply, yet the outcome depended on closing, respecting operational boundaries and escalating when blocked. Its behavior may be recognizable in the quiz precisely because the weakness accompanies a strength: the desire to investigate everything.
Kimi K3’s result also carries an important fairness note. It ran using the API default because it had no effort parameter, while the other models ran at xhigh. Its second-place score should therefore be read with that testing difference visible, not quietly erased.
A practical test for AI management
The quiz’s appeal comes from asking a deceptively simple question: can readers distinguish managers by their decisions when the brand names disappear? Across 242 unedited examples, patterns emerge in how models investigate, refuse pressure, handle blocked work and convert analysis into action.
For businesses considering AI agents in customer service, sales or planning, those patterns are more consequential than writing style. Firmulate also offers enterprises the same kind of wargame using a read-only export of their own business; nothing writes back to real systems. The lesson for any owner accustomed to nurturing living systems is straightforward: do not judge only by vigorous early growth. Apply pressure, inspect the roots and watch whether the work actually gets finished.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
