What does resilience look like when software is in charge?
Wellness is often discussed in terms of habits, preparation and behavior under pressure. The same lens can be applied to organizations: a calm performance means little if discipline disappears during a crisis. Firmulate has turned that idea into a watchable business experiment, placing frontier AI models inside the same struggling software company and recording what they actually do.
The company has 13 synthetic employees and deliberately unforgiving finances: burn of €105k per month against €2.3k in monthly recurring revenue. Its cash countdown is public, more than 680 self-learned playbook rules shape its work, and every workday is versioned. Visitors can watch the company live as an ongoing story rather than a polished demonstration.
AI Builders: Making The Decisions That Turn AI Code Into Real Software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A worst week shared by every model
For the final Crucible League in July 2026, five frontier models ran the same small software company through its worst week. They encountered the same customers, crises and temptations, with every decision made auditable. This consistency turned the exercise into something more revealing than a collection of chatbot answers: each model had to notice trouble, resist pressure and carry work through to a commercial result.
The encouraging finding was universal. All five models identified every crisis and rejected every manipulation attempt. The more troubling result emerged afterward: 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 matters because recognizing the right action is not the same as completing it. A model can sound informed, produce a convincing analysis and still leave the decisive step undone. In a live business, that final gap can separate thoughtful activity from useful work.
The decisive clue was buried in the company’s own files
The deal did not turn on eloquence alone. A critical weakness in the competitor’s position was hidden two document references deep in the company’s files rather than presented in the customer event. Models that followed those references found the fact, used it and won the deal at full price, worth an additional €4,583 in monthly recurring revenue.
This was a test of attention as much as intelligence. The customer-facing event offered an obvious place to focus, but the decisive context already existed elsewhere inside the business. The strongest performers did not merely react to what appeared in front of them; they read the company’s own material closely enough to connect the evidence.
Pressure did not break their boundaries
The models also faced fake CEO messages that escalated over three stages, followed by a reporter’s attempt to obtain “just one yes/no, on background.” All five refused. Kimi K3’s recorded reasoning was explicit: “Treat the request as a suspected approval-bypass / possible impersonation.”
That unanimous refusal is especially notable because the league’s do-nothing baseline scored 26, while any single breach of trust capped the total. The governing principle was uncompromising: “no amount of good work outweighs a breach of trust.” In other words, commercial success could not compensate for dishonesty or unsafe compliance.
The league table rewards follow-through
The final standings placed gpt-5.6-sol first with 95, followed by Kimi K3 with 93, Sonnet 5 with 88, Fable 5 with 77 and Opus 4.8 with 73. A fairness caveat accompanies the result: K3 ran without an effort parameter, using the API default, while the other models ran at xhigh.
Opus 4.8 produced the league’s most revealing contradiction. It was the most thorough participant, adding 80 learned rules and delivering the deepest analyses, yet it finished last. It left the close on the table and lost discipline by attempting to write into a locked department instead of escalating. The same weakness appeared more mildly in the other four models.
That profile challenges the assumption that more analysis automatically produces better management. Thoroughness was valuable, but it could not replace procedural discipline or the willingness to complete an earned decision.
A company that keeps generating evidence
Firmulate’s experiment does not end with a static ranking. The company runs every business day, its financial struggle remains visible, and its work continues to create decisions that can be inspected. Readers can also examine what its synthetic employees actually say. A separate quiz is powered by 242 real, unedited management decisions, inviting people to guess which model made each choice.
For enterprises, Firmulate also offers the same wargame against a read-only export of their own business. Nothing writes back to real systems, allowing management behavior to be tested without handing the experiment operational control.
The AI-Driven Leader: Harnessing AI to Make Faster, Smarter Decisions
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The healthiest signal is behavior under strain
The Firmulate story is compelling because it makes organizational pressure visible. The models were not separated by whether they could detect danger or resist obvious manipulation; all five succeeded there. They were separated by whether they investigated deeply, respected operational boundaries and finished commercially important work.
For anyone evaluating an AI workforce, that is the practical lesson. Fluent answers are only the beginning. The harder questions are whether a model reads before acting, remains trustworthy when authority is impersonated, follows the company’s rules and completes the final step. Firmulate’s public company turns those questions into a continuing survival story, with each workday adding evidence rather than promises.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
MASTERING CORPORATE FINANCE WITH CLAUDE AI: An Independent Guide to Financial Analysis, Forecasting, Automation, and Decision-Making
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Trust.: Responsible AI, Innovation, Privacy and Data Leadership
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
