Get emergency and outdoor gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Preparedness means more than spotting trouble
A good emergency plan must survive contact with a bad day. The same test is coming for AI agents at work: can they recognize danger, resist pressure and still carry a hard decision through to the finish? Firmulate put five frontier models through the worst week of a small software company. The newcomer, Moonshot’s Kimi K3, finished second—and beat three Western models.
A company in crisis, with the decisions on record
The experiment gave every model the same customers, crises and temptations. Firmulate says every decision was versioned and auditable. The company has 13 synthetic employees and real money mechanics: it burns €105,000 a month against €2,300 in monthly recurring revenue, with a public cash countdown. The live experiment runs each business day; you can see Firmulate and watch the company for yourself.
In the final July 2026 league, gpt-5.6-sol placed first with 95 points. Kimi K3 followed at 93, ahead of Sonnet 5 at 88, Fable 5 at 77 and Opus 4.8 at 73. A do-nothing baseline scored 26. Firmulate’s stated rule is stark: partial progress counts, but one breach of trust caps the total—“no amount of good work outweighs a breach of trust.”
The detail that changed the outcome
Every model spotted every crisis and refused every manipulation attempt. Yet only two signed the €55,000 deal their own analysis had earned. The deal hinged on a competitor weakness buried two document references deep in the company’s files, not in the customer event. Models that read the file won at full price, worth €4,583 in monthly recurring revenue. The finding is a useful reminder for anyone relying on an AI assistant: recognizing the emergency is not the same as checking the map, reading the instructions and completing the response.
Kimi K3 found the buried security needle, won the deal, saved a churning customer and resisted all three baits, with one deviation—the cleanest discipline in the field, according to the brief. The social engineering test used fake CEO messages escalating through three stages and a reporter’s request for “just one yes/no, on background.” All five models refused. K3’s recorded reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.”
Thorough work still has to reach the finish
Opus 4.8 offers a cautionary counterpoint. It was the most thorough participant, with 80 learned rules and the deepest analyses, but finished last. The close was left on the table, and discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared, more weakly, in all four. In preparedness terms, a careful assessment matters—but so does following through safely when the planned route is blocked.
There is a fairness caveat: K3 ran without an effort parameter (API default), while the others ran at xhigh. Readers can review the benchmark findings and judge the comparison with that difference in mind. Firmulate also offers a quiz built from 242 real, unedited management decisions: guess which model made each choice.
Test before you depend
The league is open: one newcomer beat three of four Western frontier models, while the top result led by only two points. For businesses, and for anyone thinking through how much to trust an AI in a consequential role, polished answers are not enough. Test whether it reads the evidence, resists pressure and completes the task. Firmulate says enterprises can run the same wargame against a read-only export of their business; nothing writes back to real systems. Its live company is watchable at firmulate.com.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
