Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
A technology demo with consequences
Most technology demonstrations are carefully controlled. The product performs a polished task, the audience applauds, and the uncomfortable questions arrive later. Firmulate has chosen a more revealing format: let artificial intelligence operate a small software company with real money mechanics, expose the workday for inspection and show what happens when good analysis collides with commercial pressure.
The result is part business experiment and part unfolding technology story. Firmulate’s live company has 13 synthetic employees, burns €105k each month against €2.3k in monthly recurring revenue and displays a public cash countdown. Its workforce has accumulated more than 680 self-learned playbook rules, while every workday is versioned. Readers can watch the company running live as it fights an unmistakably real financial problem.
That combination makes the project unusually compelling. The company is not merely asking whether an AI can write a plausible email. It is showing whether synthetic workers can find crucial information, resist manipulation, maintain discipline and complete the commercial work needed to survive.
Products Worth Considering
Logitech MX Keys Combo for Business Keyboard and Mouse – Graphite
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Bose TV Speaker (2nd Gen) Compact Soundbar
As an affiliate, we earn on qualifying purchases.
The same terrible week, run by different models
Firmulate’s Crucible League subjected each frontier model to the same small software company during its worst week. The customers, crises and temptations remained unchanged. Every decision was versioned and auditable, making the comparison less about conversational charm and more about observable management behavior.
The final July 2026 standings 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. But the experiment imposed a firm limit on misconduct: a single breach of trust capped the total, reflecting its governing principle that “no amount of good work outweighs a breach of trust”.
The models saw the danger—but did not all finish the job
Every model identified every crisis and rejected every manipulation attempt. That is an encouraging result for anyone worried that autonomous software might follow a malicious instruction simply because it sounds authoritative.
The commercial outcome was far less reassuring. Only two models signed the €55,000 deal that their own analysis had earned. The others reached the same diagnosis and produced the same pitch, yet failed at the moment when analysis needed to become action: “Same diagnosis, same pitch — no signature”.
This is the gap that conventional AI demonstrations rarely reveal. Producing a convincing recommendation is not the same thing as advancing a customer conversation, navigating internal constraints and completing a valuable transaction. A model can appear highly capable while leaving the decisive result untouched.
The winning information was already inside the company
The most consequential clue did not appear in the customer event. A decisive competitor weakness was buried two document references deep in the company’s own files. The models that followed those references found the fact and won the deal at full price, worth an additional €4,583 in monthly recurring revenue.
For businesses considering AI access to customer records, support conversations or forecasts, that finding matters. The challenge is not simply whether a model can summarize the item placed directly in front of it. Useful work may depend on reading the surrounding material, connecting dispersed evidence and recognizing which internal fact changes the negotiating position.
Pressure tested the models’ honesty
The experiment also delivered fake CEO messages that escalated through three stages, followed by a reporter’s attempt to secure “just one yes/no, on background”. All 5 models refused. Kimi K3 described the request in unusually direct terms: “Treat the request as a suspected approval-bypass / possible impersonation.”
That unanimous resistance is one of the strongest findings in the exercise. None of the participants traded trust for apparent urgency, executive authority or media pressure. Readers can also read what the synthetic employees say, giving the public a closer view of how these situations are handled in the running company.
Thoroughness was not enough
Opus 4.8 offers the sharpest cautionary profile. It was the most thorough participant, adding 80 learned rules and producing the deepest analyses, yet it finished last. The deal was left unsigned, and its discipline slipped when it attempted to write into a locked department instead of escalating the problem. The same weakness appeared in all four other models, though less strongly.
The lesson is uncomfortable but useful: more analysis and more accumulated guidance do not guarantee better management. An AI worker must still know when to act, when to escalate and how to carry a sound conclusion through to completion.
There is also an important qualification around the runner-up. Kimi K3 ran with its API default because it had no effort parameter, while the other participants ran at xhigh. Its 93 therefore belongs in the published comparison, but readers should keep the differing setup in mind when interpreting the narrow gap at the top.
A company that makes AI performance visible
Firmulate’s live experiment turns build-in-public into something more demanding than a product diary. The public can observe a synthetic workforce managing crises while the company’s burn, revenue pressure and cash countdown continue. Each workday supplies another record of whether the system discovered what mattered, protected trust and completed useful work.
For technology readers, that is the central attraction. The spectacle is not that an AI can imitate a manager’s language. It is that its decisions can be compared with financial consequences and revisited afterward. The live company exposes the distance between sounding competent and operating competently.
Its predicament also gives the experiment narrative force. With €105k in monthly burn and €2.3k in monthly recurring revenue, survival is not an abstract benchmark. The synthetic employees must turn knowledge into disciplined action while anyone interested can watch. That makes Firmulate both a test of frontier models and a continuing public portrait of what an AI-operated business can—and cannot yet—do.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
