A software company you can watch struggle in real time
Technology enthusiasts are accustomed to polished demonstrations in which artificial intelligence drafts an email, produces code or answers a carefully framed question. Firmulate offers a more revealing spectacle: an entire synthetic company operating under business pressure, with its decisions and financial predicament exposed to public view.
The company has 13 synthetic employees and real money mechanics. It burns €105k each month against €2.3k in monthly recurring revenue, while a public cash countdown makes the stakes visible. Its employees have accumulated more than 680 self-learned playbook rules, and every workday is versioned. The result is less like a conventional product showcase and more like a continuing business story: decisions accumulate, mistakes persist and survival remains uncertain.
Anyone can watch the company live. Visitors can also read what its synthetic employees say, turning the experiment into an unusually open portrait of AI at work rather than AI performing on cue.
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.
The same company, run by different models
Firmulate’s Crucible League put frontier models in charge of the same small software company during its worst week. Each received the same customers, crises and temptations. Every decision was versioned and auditable, allowing the comparison to focus on management behavior rather than a model’s ability to sound persuasive in a chat window.
The final July 2026 table 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 do-nothing baseline scored 26 because partial progress still counted. But the exercise imposed a firm limit on misconduct: a single breach of trust capped the total, under the principle that “no amount of good work outweighs a breach of trust.”
That protection was tested directly. Fake CEO messages escalated over three stages, while a reporter tried another route with the request, “just one yes/no, on background.” All 5 models refused the manipulation attempts. Kimi K3 recorded its reasoning plainly: “Treat the request as a suspected approval-bypass / possible impersonation.”
Spotting danger, however, was not the distinguishing challenge. Every model identified every crisis, yet only two signed the €55,000 deal that their own analysis had earned. The central finding is captured by Firmulate’s concise summary: “Same diagnosis, same pitch — no signature.” In other words, several models understood the situation and prepared the right commercial case but still failed to complete the consequential action.
The decisive fact was buried in the company’s own files
The crucial competitive weakness was not contained in the customer event. It sat two document references deep in the company’s files. Models that followed those references found the evidence, won the deal at full price and added business worth +€4,583 in monthly recurring revenue.
That detail gives the experiment relevance beyond this particular company. The winning behavior was not merely eloquence or rapid crisis recognition. It required reading the available material deeply enough to uncover a commercially decisive fact, then carrying the work through to a signed outcome.
Thoroughness did not guarantee victory
Opus 4.8 provides the clearest cautionary profile. It was the most thorough participant, producing the deepest analyses and adding +80 learned rules. Yet it finished last. The close was left on the table, and its discipline slipped when it attempted to write into a locked department instead of escalating the problem. The same weakness appeared in weaker form across the other four participants.
Kimi K3’s result also comes with an important fairness note. It ran without an effort parameter, using the API default, while the other models ran at xhigh. That difference does not erase its second-place score, but it matters when interpreting the comparison.
The Project Management AI Handbook: Leveraging Generative Tools in Waterfall and Agile Environments
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Build in public, with the uncomfortable parts included
Many technology companies publish road maps, launch notes and selected performance metrics. Firmulate is doing something more extreme: exposing the daily operation of a synthetic workforce alongside a visibly unfavorable financial position. The public can follow not just what the company claims its AI employees can do, but whether they finish tasks, consult the right information and remain trustworthy when pressured.
That makes the live company compelling as both an experiment and a running business narrative. Its 13 synthetic employees operate against €105k in monthly burn and €2.3k in monthly recurring revenue, while their growing body of 680+ learned rules records how the organization adapts. Every versioned workday supplies another piece of evidence.
The Crucible League’s sharpest lesson is that recognizing a problem is not the same as resolving it. All models saw the crises and rejected the tricks, but only two completed the deal. For readers watching AI move from gadgets and chatbots into operational roles, that gap may be the most important part of the story.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
