Imagine a busy kitchen where every chef is tested not just on how well they cook, but on whether they can keep their integrity under pressure. Now, replace the chefs with AI models managing a small software company. The question isn’t just about how smart they are — it’s whether they can finish what they start, read their ingredients before cooking, and resist shortcuts even when stakes are high. That’s exactly what a recent live experiment by Firmulate put to the test, revealing surprising results that could reshape how we trust AI in business—and maybe, in your kitchen.
Get kitchen gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
The Challenge: Putting AIs Through the Toughest Week
In July 2026, four leading AI models faced the same daunting test: manage a small software company through its worst week. This wasn’t a casual demo — every decision was carefully recorded, every crisis simulated, and every temptation to cheat monitored. The goal? See which AI could truly handle the pressures of running a real business, not just generate pretty chat.
The Contenders and the Scoreboard
- gpt-5.6-sol scored the highest, with a remarkable 95 out of 100, winning the full deal and closing a €55,000 contract.
- Kimi K3, the newcomer from Moonshot, scored just slightly below at 93, also sealing the deal and generating an extra €4,583 in monthly recurring revenue.
- Sonnet 5 scored 88, managing the deal but with some process slips.
- Fable 5 and Opus 4.8 trailed behind with scores of 77 and 73, respectively, each leaving some opportunities on the table.
The Real Test: Reading Between the Files
While all models demonstrated they could identify crises and refuse manipulative tactics—like fake CEO messages or reporter tricks—the real differentiator lay in their depth of understanding. The winning models looked two document references deep into the company’s files, uncovering a buried fact that clinched the deal. Those that skipped this step lost the opportunity, highlighting a critical weakness: sometimes, reading more thoroughly makes the difference between success and failure.
Honesty Under Pressure
Part of the experiment involved social engineering — fake messages designed to trick the AI into approving bypasses or impersonations. All five models refused these requests, with Kimi K3 explaining, “Treat the request as a suspected approval-bypass / possible impersonation.” This discipline is crucial because, in real business, trust and integrity aren’t optional.
The Live Business: Simulating Reality
The experiment runs on a real, functioning company with 13 synthetic employees and actual money mechanics. It burns €105,000 a month against a modest €2,300 MRR, with every decision and rule versioned daily. Watching this unfold at firmulate.com/live offers a transparent glimpse into how AI models perform when managing real-world complexities.
The Surprising Gap: Discipline and Depth Matter
Among the models, Opus 4.8 was the most thorough, with over 80 learned rules and deep analyses, yet it finished last. It left the close on the table because it failed to escalate issues properly, writing attempts into a locked department instead of taking action. This shows that even the most analytical AI can stumble if it lacks discipline or depth in decision-making.
As an affiliate, we earn on qualifying purchases.
The Bigger Takeaway: Trust and Utility in AI Management
The core insight from this live experiment is clear: AI systems must do more than generate nice language. They need to finish what they start, read the full context, and remain honest under pressure. In practical terms, if AI agents are to handle your CRM, support queue, or business forecasts, their ability to stay disciplined and thorough directly impacts your bottom line.
Fairness and Testing Conditions
It’s important to note that Kimi K3 ran without an effort parameter — the default API setting — while the other models operated at a higher setting (xhigh). This fairness note underscores that performance differences aren’t just about raw capability but also about how models are configured and tested.
Try It Yourself
Interested enterprises can run their own scenarios against a read-only export of their business data. This no-risk “wargame” lets you see how your AI might perform before fully deploying it. Details are available at firmulate.com/pilot.html.
The live experiment shows that the most successful AI models are those that read deeply, stay disciplined, and refuse shortcuts—even under pressure. Trust is built through thoroughness and integrity, not just clever language. For any business or kitchen, choosing an AI that can finish what it starts is the key to real utility.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
AI decision-making tools for companies
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
