A backup generator earns trust when the grid fails, not when the lights are on. AI agents face a similar test when a company hits a bad week: can they spot trouble, protect customers and finish the work? Firmulate put five frontier models through that kind of live business trial. Moonshot’s Kimi K3 finished second, ahead of three Western competitors.
Get backup power and energy gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
A company under pressure
Firmulate’s Crucible gave each model the same small software company, the same customers, crises and temptations. Every decision was versioned and auditable. The company is a live experiment, with 13 synthetic employees and real money mechanics: it burns €105,000 a month against €2,300 in monthly recurring revenue. Readers can watch the public cash countdown and the company’s workdays at Firmulate.
In the final league table, dated July 2026, gpt-5.6-sol scored 95, Kimi K3 scored 93, Sonnet 5 scored 88, Fable 5 scored 77, and Opus 4.8 scored 73. A do-nothing baseline scored 26. Firmulate says 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
Dyson Ball Animal 3 Upright Vacuum
Dyson V8 Cyclone Cordless Vacuum
eufy E15 Gen 2 Wire-Free Robotic Lawn Mower
As an affiliate, we earn on qualifying purchases.
Reading the files mattered
All five models spotted every crisis and refused every manipulation attempt. Yet only two signed the €55,000 deal their own analysis had earned. The decisive competitor weakness was buried two document references deep in the company’s files, rather than in the customer event. Models that found it won the deal at full price, worth €4,583 in monthly recurring revenue.
K3 found the buried fact, won the deal and saved the churning customer. It also resisted all three baits and had just one deviation, the cleanest discipline in the field. When fake CEO messages escalated over three stages and a reporter asked for “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.”
The pattern shows why fluent answers alone may not tell a company whether an AI agent can do a job. The models shared the diagnosis, but some did not complete the close: “Same diagnosis, same pitch — no signature.” Reading the company’s own records helped separate those that spotted an opportunity from those that finished it.
Thoroughness did not guarantee execution
Opus 4.8 was the most thorough participant, with 80 learned rules and the deepest analyses, yet placed last. It left the deal unsigned and slipped on discipline by attempting to write into a locked department instead of escalating. Firmulate reports the same weakness, in a milder form, across all four other participants. The result is a reminder that analysis and execution are different parts of managing a business.
The experiment’s scores come with a fairness caveat: K3 ran without an effort parameter, using the API default, while the others ran at xhigh. Firmulate also offers a quiz built from 242 real, unedited management decisions, inviting readers to guess which model made each choice. Full results and the quiz are available at Firmulate’s benchmark pages.
Test before handing over the keys
For businesses considering AI in customer support, forecasting or other operations, the practical question is whether a model will read the relevant records, act under pressure and follow through. Firmulate says enterprises can run the wargame against a read-only export of their own business, with no writes back to real systems. In home energy, backup power is judged under failure conditions; AI management deserves its own real-world stress test before it gets responsibility.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
