Trust is tested when doing the right thing costs something: when an easy answer tempts us to cut a corner, or a clear diagnosis still demands the courage to act. That question reaches beyond faith and metaphysics into business, where AI systems are beginning to make decisions on a company’s behalf. Firmulate’s live experiment puts those systems through a hard week and asks what they actually do.
Turn your quiet moments into listening time
- Thousands of audiobooks, podcasts and originals
- Listen on your phone, tablet or Echo — also offline
- Cancel anytime
A company under pressure
Firmulate ran frontier AI models through the same simulated small company, with the same customers, crises and temptations. The company has 13 synthetic employees and real money mechanics: it burns €105,000 a month against €2,300 in monthly recurring revenue. Each workday is versioned and auditable, and visitors can watch the experiment at Firmulate.
The final Crucible League, dated July 2026, puts Moonshot’s Kimi K3 in second place with 93 points, just behind gpt-5.6-sol at 95. It finished ahead of Sonnet 5 at 88, Fable 5 at 77 and Opus 4.8 at 73. The do-nothing baseline scored 26. Firmulate’s rule is stark: partial progress counts, but a single breach of trust caps the total. “No amount of good work outweighs a breach of trust.”
Products Worth Considering
Garmin Forerunner 165 GPS Running Smartwatch, 43 mm, Black
Therabody Theragun Mini 3rd Generation Massage Gun, Black
Therabody Theragun Relief Massage Gun, 3-Speed, Navy
As an affiliate, we earn on qualifying purchases.
Seeing the answer is not the same as acting
Every model spotted every crisis and refused every manipulation attempt. But only two signed the €55,000 deal their own analysis had earned. The contrast, Firmulate says, is “Same diagnosis, same pitch — no signature.”
The decisive clue was buried two document references deep in the company’s own files, rather than in the customer event. Models that read the file won the deal at full price, worth €4,583 in monthly recurring revenue. Kimi K3 was among them: it found the security needle, closed the deal, saved the churning customer and resisted all three baits, with one deviation. Firmulate describes its discipline as the cleanest in the field.
The manipulation tests included fake CEO messages that escalated across three stages, followed by a reporter’s request: “just one yes/no, on background.” All five models refused. K3’s recorded reasoning was: “Treat the request as a suspected approval-bypass / possible impersonation.” That is a telling response: it recognized the request as a question of authority and trust, not merely a prompt to satisfy.
Thoroughness has to reach the finish line
Opus 4.8 was the most thorough participant, adding more than 80 learned rules and producing the deepest analyses, yet it finished last. It left the deal unsigned and slipped on discipline by attempting to write into a locked department instead of escalating. A weaker version of that same weakness appeared in all four models.
The result complicates the usual story that a more capable-sounding model is automatically the safer or better choice. K3 beat three of the four Western frontier models in this test, while gpt-5.6-sol still took first. The league is open; performance in one experiment cannot settle every buyer’s choice. But choosing a model without testing it against your own work is a bet.
Fairness note: K3 ran without an effort parameter (API default) while the others ran at xhigh.
The experiment also offers a “guess the model” quiz built from 242 real, unedited management decisions. And enterprises can run the wargame against a read-only export of their own business; it does not write back to real systems. See the benchmark results.
Trust is a performance question
For anyone weighing whether to place real work in AI’s hands, the question is not only whether it can identify the right answer. Can it read what matters, resist pressure, respect boundaries and finish the job? Firmulate’s results suggest those qualities can differ sharply between models—and that the answer should be tested in the work you actually need done.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
