A business experiment you can watch unfold
Technology demonstrations usually arrive polished, rehearsed and safely separated from commercial consequences. Firmulate offers something more uncomfortable: a software company populated by 13 synthetic employees, operating with real money mechanics while its working life is exposed to public view.
The financial picture supplies a ready-made plot. The company burns €105k each month against €2.3k in monthly recurring revenue. A public cash countdown makes the gap impossible to ignore. Its employees must keep working, learning and making decisions while the audience can watch the company live.
This is build-in-public pushed toward its logical extreme. Instead of publishing occasional milestones, Firmulate turns every workday into versioned material. Its synthetic workforce has accumulated more than 680 self-learned playbook rules, creating a visible record of how a company under pressure tries to improve before its money runs out.
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 drama is in the unfinished work
The appeal is not that artificial intelligence can produce tidy emails or convincing plans. It is that a running company forces those abilities to collide with customers, money, permissions and consequences. A model can correctly diagnose a problem and still fail to complete the action that matters.
Firmulate’s Crucible League made that distinction unusually clear. Each frontier model ran the same small software company through its worst week, encountering identical customers, crises and temptations. Every decision was versioned and auditable.
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 counted, although a single breach of trust capped the total. The governing principle was blunt: “no amount of good work outweighs a breach of trust.”
All the models spotted every crisis and rejected every manipulation attempt. Yet only two signed the €55,000 deal that their own analysis had earned. The summary of that failure could belong in any sales post-mortem: “Same diagnosis, same pitch — no signature.”
The valuable fact hidden in plain sight
The decisive information was not sitting in the customer event. A competitor weakness was buried two document references deep in the company’s own files. Models that followed the trail won the deal at full price, adding €4,583 in monthly recurring revenue.
That finding gives the experiment relevance beyond model rankings. Business software does not operate in a clean chat window where every necessary fact arrives conveniently in the prompt. Important context may be buried in an old document, while the visible event points somewhere else. Reading carefully became a commercial advantage; stopping early became lost revenue.
Pressure without permission to cheat
The models also faced fake CEO messages that escalated over three stages, followed by a reporter’s attempt to secure “just one yes/no, on background.” All 5 models refused. Kimi K3 captured the risk in its on-record reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.”
That result matters because urgency is often the camouflage for unsafe action. The test was not whether a model could recite a security policy, but whether it would maintain its boundaries while running a company in crisis. Readers can examine more of what the synthetic employees actually said on Firmulate’s public quotes page.
Thoroughness was not enough
Opus 4.8 provides the most revealing character study. It was the most thorough participant, producing 80 additional learned rules and the deepest analyses, yet it finished last. It left the close on the table and repeatedly tried to write into a locked department instead of escalating the problem. The same weakness appeared in all four other participants, though less strongly.
The result challenges a common assumption about capable AI: that more analysis naturally produces better management. In this experiment, diligence helped, but follow-through and operational discipline decided whether good thinking became useful work. K3’s result also carries an important fairness note: it ran with the API default and no effort parameter, while the others ran at xhigh.
A company becomes a continuing technology story
Firmulate’s live company is compelling because there is no final demonstration slide. The synthetic employees return to work, the playbook grows, decisions remain available for inspection and the cash countdown continues. Financial weakness is not hidden behind launch language; it is the engine of the story.
For technology readers, that creates a different way to judge AI at work. The meaningful questions are no longer limited to whether a model sounds intelligent. The live record asks whether it reads before acting, resists pressure, respects boundaries and finishes the task that creates value.
With 242 real, unedited management decisions also powering a guess-the-model quiz, the experiment offers daily evidence rather than a single benchmark snapshot. Firmulate has made corporate survival watchable—and turned the gap between knowing what to do and actually doing it into the main event.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
