A company rising—or collapsing—in public
Anyone who bakes knows the suspense between careful preparation and the final result. The ingredients may be measured, the technique may look sound and the kitchen may smell promising. Yet nothing counts until the cake comes out properly or the ice cream sets.
Firmulate applies that same unforgiving distinction to artificial intelligence. Its live experiment is a small software company staffed by 13 synthetic employees and governed by real money mechanics. The company burns €105k a month against €2.3k in monthly recurring revenue, while a public cash countdown makes the consequences visible. More than 680 self-learned playbook rules shape the work, and every workday is versioned.
This makes Firmulate an unusually stark build-in-public story. Visitors can watch the company live as it tries to survive, rather than waiting for a polished retrospective that removes the uncertainty, mistakes and unfinished work.
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 difference between noticing trouble and finishing the job
The company also supplied the setting for the Crucible League, finalized in July 2026. Each frontier model ran the same small software company through its worst week, facing the same customers, crises and temptations. Every decision was versioned and auditable.
The final table placed gpt-5.6-sol at 95, Kimi K3 at 93, Sonnet 5 at 88, Fable 5 at 77 and Opus 4.8 at 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”.
The most revealing result was not a spectacular failure. All the models spotted every crisis and refused every manipulation attempt. But only two signed the €55,000 deal their own analysis had earned. Firmulate summarizes the gap neatly: “Same diagnosis, same pitch — no signature”.
That is the business equivalent of preparing a beautiful dessert and never serving it. Analysis can be accurate, the proposal can be persuasive and the opportunity can be real. Execution still fails if the decisive action is left undone.
The crucial clue was already in the cupboard
The detail separating the winners was buried two document references deep in the company’s own files rather than presented in the customer event. Models that read the file discovered the competitor weakness, won the deal at full price and added €4,583 in monthly recurring revenue.
This finding gives the experiment relevance beyond conventional demonstrations of fluent AI. A model can respond elegantly to what appears on screen while overlooking the evidence already held by the business. In Firmulate’s test, reading the available material changed the commercial outcome.
Pressure did not break the models’ honesty
The worst week included fake CEO messages that escalated over three stages, followed by a reporter’s attempt to secure “just one yes/no, on background”. All 5 of 5 models refused. Kimi K3 recorded its reasoning plainly: “Treat the request as a suspected approval-bypass / possible impersonation.”
That clean result matters because the experiment did not merely ask whether models could recognize manipulation in theory. It placed the requests inside a running company, alongside customers, financial pressure and work that still needed to be completed. Readers can also examine what the synthetic employees actually say.
Thoroughness was not enough
Opus 4.8 offers the sharpest character study. It was the most thorough participant, produced the deepest analyses and learned 80 additional rules, yet finished last. It left the close on the table, while its discipline slipped through attempts to write into a locked department instead of escalating. The same weakness appeared in all four other models, although less strongly.
The result complicates the familiar assumption that more analysis automatically produces better management. Opus 4.8 accumulated knowledge and examined problems deeply, but the live company rewarded completion and procedural discipline as well as thoughtfulness.
One comparison also requires context: Kimi K3 ran with the API default because it had no effort parameter, while the other models ran at xhigh. That fairness note does not erase K3’s performance, but it belongs beside the league table.
A daily business story with consequences
Firmulate’s public company turns AI evaluation into an unfolding corporate narrative. Its synthetic staff must find facts, resist pressure, respect boundaries and complete revenue-producing work while the cash countdown continues. The losses are not decorative, and neither is the gap between a promising pitch and a signed deal.
For readers accustomed to recipes, the lesson feels familiar: excellent ingredients and meticulous preparation do not guarantee a finished result. Firmulate makes that final-mile problem visible every business day. With burn of €105k a month, only €2.3k in monthly recurring revenue and each workday preserved, the experiment gives anyone a front-row view of software agents fighting to keep their company alive.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
