
In the rapidly evolving landscape of artificial intelligence, the question is no longer just about how well an AI can generate text or answer questions—it’s about whether these models can act reliably in real-world business scenarios. A recent live experiment reveals that a newcomer to the field, Kimi K3 from Moonshot, not only competed with, but outperformed seasoned giants, raising critical questions for enterprises about choosing AI partners.
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
The Live AI Company Wargame: Putting Models to the Test
Imagine putting AI models through a week of the toughest business crises—same customers, same crises, same temptations to cheat or manipulate. This is precisely what the recent experiment by Firmulate did. Each model was tasked with managing a small software company’s operations, decision-making, and crisis resolution, with every action recorded and auditable.
The models involved—gpt-5.6-sol, Kimi K3 by Moonshot, Sonnet 5, Fable 5, and Opus 4.8—were tested under identical conditions, facing the same challenges. The goal: see which models could spot every crisis, resist manipulation attempts, and ultimately close a critical €55,000 deal based on their own analysis.
Results: A Newcomer Takes the Lead
- The top scorer was gpt-5.6-sol with 95 points, closely followed by Kimi K3 with 93 points.
- Sonnet 5 and Fable 5 scored 88 and 77 respectively, while Opus 4.8 scored 73.
- All models identified every crisis and refused manipulation attempts, demonstrating a high level of honesty and discipline.
What set K3 apart was its ability to uncover the buried, critical information hidden two document references deep within the company’s files—an insight that led directly to closing the deal at full price, generating an additional €4,583 MRR (monthly recurring revenue).
Integrity Under Pressure
Throughout the experiment, all models refused social engineering attempts, including staged CEO messages and a reporter trick. K3’s reasoning was clear: “Treat the request as a suspected approval-bypass / possible impersonation.” This disciplined response is vital for enterprise applications where trust and security matter immensely.
The Real-World Company and Its Challenges
The live company operated with 13 synthetic employees, handling real money mechanics with a burn rate of €105k/month against €2.3k MRR. Every day, the system was versioned and monitored, and its outcomes are publicly visible at firmulate.com/live. The company’s operational complexity provided a rigorous test bed for the models’ decision-making abilities under real-world pressures.
Lessons from the Deep Analysis
The detailed analysis showed that the most thorough participant, Opus 4.8, with over 80 learned rules, was the last place in the league. Its discipline slipped during the close, and some decisions were left unresolved instead of escalation. This highlights that even deep analyses and comprehensive rules don’t guarantee success if discipline falters.
Importantly, the fairness note clarifies that K3 was run without an effort parameter (the API default), whereas other models operated at xhigh. This means K3’s performance was achieved without extra optimization or tuning, emphasizing its innate capability to perform reliably.
Products Worth Considering

Arlo Pro Wireless Security Camera, 2K HDR, 6th Gen, 4-Camera Kit

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity

Philips Hue Play HDMI Sync Box 8K
As an affiliate, we earn on qualifying purchases.
Why This Matters for Enterprises
In practical terms, this experiment underscores that the critical question for businesses adopting AI is not just about chat quality or superficial performance metrics. Instead, they should focus on whether the AI can finish what it starts, read relevant documents before acting, and resist pressure or manipulation under stress.
Choosing an AI model is now a strategic decision—one that impacts trustworthiness, security, and ultimately, bottom-line results. As the league table shows, emerging models like Kimi K3 are closing the gap quickly, challenging industry veterans and redefining standards of performance and integrity.

The recent live AI experiment demonstrates that a newcomer like Kimi K3 can outperform established models in real-world business scenarios, emphasizing the importance of discipline, thoroughness, and trustworthiness when selecting AI. The league is open, and enterprise leaders should base decisions on actual performance under pressure, not just superficial metrics.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
