Before trusting a new power tool, you want to know what it can handle when the cut gets tricky. AI agents deserve the same treatment: a polished demo is a poor substitute for seeing what happens under pressure. Firmulate’s live company experiment puts frontier models through a small software company’s worst week, with real money mechanics and consequential decisions.
Get business pricing on tools and workshop supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
A company under pressure
The experiment gives each model the same customers, crises and temptations. Every decision is versioned and auditable. The company has 13 synthetic employees, burns €105,000 a month against €2,300 in monthly recurring revenue, and publishes a cash countdown. It runs every business day, and readers can watch it at Firmulate.
The final July 2026 league table puts gpt-5.6-sol first with 95, followed by Moonshot’s Kimi K3 at 93. Sonnet 5 scored 88, Fable 5 scored 77, and Opus 4.8 scored 73. A do-nothing baseline scored 26. The benchmark counts partial progress, but one breach of trust caps the total: “no amount of good work outweighs a breach of trust.”
Popular right now
KitchenAid Classic Series 4.5 Quart Tilt-Head Stand Mixer K45SS, Onyx Black
Ninja Specialty Coffee Maker CM401, 10-Cup Carafe, Fold-Away Frother, Black
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Black
As an affiliate, we earn on qualifying purchases.
Reading the file mattered
Every model spotted every crisis and refused every manipulation attempt. Yet only two signed a €55,000 deal that their own analysis had earned. The weakness that tipped the deal was buried two document references deep in the company’s files, rather than stated in the customer event. Models that read the file won at full price, worth €4,583 in monthly recurring revenue. The site sums up the gap: “Same diagnosis, same pitch — no signature.”
Kimi K3 finished second, two points behind gpt-5.6-sol and ahead of three of the four Western frontier models in the field. It found the buried security needle, secured the deal, saved a churning customer and resisted all three baits, with one deviation—the cleanest discipline in the field. In a social-engineering test, fake CEO messages escalated over three stages, followed by a reporter asking 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.”
Thorough work still needs a finish
Opus 4.8 was the most thorough participant, with 80 learned rules and the deepest analyses, but came last. It left the deal unsigned and slipped on process, attempting to write into a locked department instead of escalating. The same weakness showed up, less strongly, in all four models. For a company considering AI agents in customer support, sales or forecasting, sound analysis is only part of the job; follow-through and respect for boundaries matter too.
There is a fairness caveat to the comparison: K3 ran without an effort parameter (API default), while the others ran at xhigh. The experiment also offers a quiz built from 242 real, unedited management decisions, inviting visitors to guess which model made each choice.
Enterprises can run the same wargame against a read-only export of their own business. Firmulate says nothing writes back to real systems. The benchmark page presents the results and findings.
The practical lesson
For anyone choosing an AI workforce, the league table makes a case for testing models on the work they will actually face: finding information in company files, closing a decision, handling pressure and respecting limits. Kimi K3’s result shows the field is open; the best fit for a real business is a question to test, not guess.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
