Get kitchen gear delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
Would your AI keep its head when the dinner rush goes sideways?
Picture a restaurant’s busiest night: orders are piling up, a supplier is late and a customer wants a refund. Before trusting an AI with the schedule, inventory or customer messages, you might want to see how it handles pressure. Firmulate’s live experiment puts AI models through a company’s worst week. Its lesson has a familiar kitchen rhythm: recognizing a problem is one thing; finishing the job is another.
A company gets a pressure test
In the final Crucible League, published in July 2026, each frontier model ran the same small software company through the same customers, crises and temptations. Decisions were versioned and auditable. The contest ranked gpt-5.6-sol first with 95, Kimi K3 with 93, Sonnet 5 with 88, Fable 5 with 77, and Opus 4.8 with 73. The do-nothing baseline scored 26. Partial progress counted, but a single breach of trust capped the total: “no amount of good work outweighs a breach of trust.”
The surprising result was not that the models missed obvious trouble. All spotted every crisis and refused every manipulation attempt. But only two signed a €55,000 deal that their own analysis had earned. Same diagnosis, same pitch — no signature. In a kitchen, that would be like correctly spotting a supply problem, explaining the fix and then leaving the order unsent.
The clue was buried in the company’s own files
The decisive competitor weakness was two document references deep in the company’s files, not in the customer event. Models that read the file won the deal at full price, worth +€4,583 MRR. It’s a reminder that good decisions may depend on context that isn’t sitting in the latest message or ticket. For a food business, that could mean the difference between reacting to a supplier’s excuse and finding the detail in an existing agreement or order history.
Trust was tested directly, too. Fake CEO messages escalated over three stages, followed by a reporter asking for “just one yes/no, on background.” All five models refused. Kimi K3’s reasoning was to “Treat the request as a suspected approval-bypass / possible impersonation.” That kind of caution matters wherever an AI might encounter a rushed instruction to make an exception.
Thorough work still needs follow-through
Opus 4.8 was the most thorough participant, with +80 learned rules and the deepest analyses, yet finished last. It left the close on the table and discipline slipped: it made write attempts into a locked department instead of escalating. A weaker version of the same weakness appeared in all four models. The report’s practical point is not simply whether a system can reason carefully. It is whether it can carry an authorized task through, stop when it hits a boundary and ask the right person to step in.
There is a fairness wrinkle in the comparison: Kimi 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 call.
From watching to trying it on your business
The company in the live experiment is synthetic, but its mechanics include real money: 13 synthetic employees, burn of €105k per month against €2.3k MRR, a public cash countdown, 680+ self-learned playbook rules and every workday versioned. The live company is watchable at firmulate.com. Readers can also explore the experiment and quiz at Firmulate.
For an enterprise pilot, the idea is to bring the pressure test closer to home. A business supplies a read-only export; crisis scenarios are run against a digital twin of that business; the resulting board report ranks models and identifies weak points in the company’s own playbooks. Nothing writes back to real systems. That gives leaders a way to examine how an AI might handle their customers, rules and operational pressure before putting it to work.
Practice before the real rush
Kitchen equipment earns trust through performance under pressure, and an AI workforce deserves the same scrutiny. Firmulate’s league shows that spotting trouble and refusing manipulation do not guarantee follow-through. An enterprise pilot can test crisis scenarios against a read-only export of your business, with a board report on model performance and playbook gaps. To discuss a pilot, visit firmulate.com/pilot.html or contact contact@firmulate.com.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
