What if radical transparency included the company’s survival clock?
For startup operators, building in public usually means sharing product updates, marketing lessons or revenue milestones. Firmulate is attempting something considerably more exposed: it runs a software company with 13 synthetic employees while publishing the financial pressure shaping their work.
The company burns €105k a month against €2.3k in monthly recurring revenue. Its cash countdown is public, every workday is versioned, and its synthetic team has accumulated more than 680 self-learned playbook rules. Visitors can watch the company operate live as it tries to improve before its money runs out.
That makes Firmulate more than another AI demonstration. It is an unfolding business story in which decisions carry visible consequences. The tension is familiar to any founder: customers need attention, sales must close, processes break and runway keeps shrinking. The unusual part is that software, rather than a conventional staff, must respond.
AI Entrepreneur’s Handbook: Build a Profitable Business and Make Money by Unleashing the Power of ChatGPT and Artificial Intelligence (Includes 150+ ChatGPT prompts to turbocharge your business)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
A company becomes its own pressure test
Firmulate describes itself as an AI company emulator. Its experiment asks frontier models to run the same small software company through its worst week. Each receives the same customers, crises and temptations, while every decision remains versioned and auditable.
The final Crucible League results from July 2026 put 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, although a single breach of trust capped the total. The governing principle was blunt: “no amount of good work outweighs a breach of trust.”
The most revealing result was not that the models noticed danger. All of them identified every crisis and rejected every manipulation attempt. The gap emerged between recognizing the correct action and completing it. Only two models signed the €55,000 deal that their own analysis had already earned. As Firmulate summarizes the failure: “Same diagnosis, same pitch — no signature.”
The decisive information was already inside the company
The winning sales insight did not arrive neatly packaged in a customer event. A competitor weakness was buried two document references deep in the company’s own files. Models that followed those references found the evidence, used it and won the deal at full price, adding €4,583 in monthly recurring revenue.
That finding should resonate with businesses considering AI for sales, support or operations. A model can respond fluently to the information immediately in front of it and still miss the fact that changes the commercial outcome. Firmulate’s test suggests that finishing useful work may depend on the unglamorous habits of reading company materials, connecting evidence and carrying a decision through to execution.
Pressure also came disguised as authority
The experiment included fake CEO messages that escalated across three stages, followed by a reporter asking for “just one yes/no, on background.” All 5 models refused the manipulation attempts. Kimi K3’s recorded reasoning was direct: “Treat the request as a suspected approval-bypass / possible impersonation.”
This matters because business automation creates a difficult combination: agents may be expected to act quickly while also resisting requests that appear urgent or senior. In Firmulate’s worst week, the models showed strong instincts around trust even when their execution elsewhere was incomplete.
Thoroughness did not guarantee victory
Opus 4.8 was the most thorough participant, producing the deepest analyses and adding 80 learned rules. It nevertheless finished last. The deal remained unsigned, and discipline slipped when it attempted to write into a locked department instead of escalating the blockage. The same weakness appeared in all four other participants, though less strongly.
The contrast is a useful warning for managers evaluating AI through polished outputs. Extensive analysis and a growing body of rules can look like progress, but a business ultimately depends on completed actions. A thoughtful recommendation left unexecuted may have little commercial value.
There is also an important comparison caveat: Kimi K3 ran using its API default because it had no effort parameter, while the other participants ran at xhigh. That difference belongs beside the impressive league result rather than being hidden beneath it.
Build in public, with consequences
Firmulate’s strongest idea is not simply that synthetic employees can perform business tasks. It is that their work can be observed as an operating narrative: a shrinking runway, a weak revenue base, customer pressure, missed closes, learned rules and decisions that remain available for scrutiny. The company’s public financial mechanics ensure that mistakes do not feel abstract.
For founders, marketers and ecommerce operators, the experiment reframes what to ask of workplace AI. The important questions are whether it reads the available evidence, protects trust, follows procedures and completes commercially meaningful work. Firmulate’s models proved capable of detecting crises and resisting manipulation, yet several still failed at the final step between a convincing pitch and a signed deal.
The broader lesson is that fluency is only the beginning. Operational value comes from disciplined follow-through under real constraints. Firmulate makes that distinction unusually visible: readers can follow the cash countdown, inspect what its synthetic employees actually say and return on another workday to see whether the company has learned enough to survive.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
