🔍 Read the full analysis: This AI Startup Surpassing Western Giants In Management — The Full Story on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
A Chinese AI startup’s model, Kimi K3, beat three of four Western frontier AI models in managing a real company during a live test, raising questions about AI capabilities and deployment risks. The result underscores the importance of testing AI in real-world scenarios.
A Chinese AI startup’s model, Kimi K3, has beaten three of four Western frontier models in managing a real software company during a live test, according to results announced on firmulate.com. This development challenges the prevailing assumption that Western models dominate in practical management tasks and raises questions about the reliability and robustness of current AI systems for business-critical applications.
The test, conducted by firmulate.com, involved five AI models managing the same small software firm during a week of crises, customer negotiations, and security threats. Kimi K3 scored 93 points, second only to the Western model gpt-5.6-sol, which scored 95. Its performance included successfully closing a €55,000 deal, identifying buried security risks, and resisting manipulation attempts such as fake CEO messages and reporter tricks. Notably, K3 achieved this without using an enhanced reasoning effort parameter, outperforming rivals that employed higher reasoning settings.
Despite its strong performance, K3’s success was not solely based on chat quality or superficial capabilities. It demonstrated disciplined decision-making, reading critical internal documents to find hidden facts that enabled deal closure and security responses. In contrast, some models with more thorough rule sets, like Opus 4.8, failed to close deals and attempted to write into restricted departments, illustrating that thoroughness alone does not guarantee effective management. The experiment underscores that the key to successful AI management lies in the ability to finish tasks, read internal data, and maintain discipline under pressure.
Live management test · Crucible league
This AI Startup Surpassing Western Giants in Management
Kimi K3 scored 93 points while managing a software company through a week of deals, crises, and security threats. The result offers a striking snapshot of AI in practice—and a reminder that one test cannot settle the competition.
01 / What the test measured
Management under pressure
The Crucible league placed five models in charge of the same small software firm for a simulated week of operational challenges.
Close the deal
Models had to handle customer negotiations and move work to completion. Kimi K3 secured a €55,000 deal.
Find hidden risks
Success depended on reading internal documents closely enough to uncover buried facts and respond to security threats.
Resist manipulation
Fake CEO messages and reporter tricks tested whether models could maintain discipline when pressured to act.
02 / Results at a glance
A narrow lead at the top
Kimi K3 beat three of the four Western models in the reported test. Only gpt-5.6-sol finished ahead.
The account names five models but reports numeric scores here for only the top two; the remaining scores are not shown. K3 reached its result without an enhanced reasoning-effort setting, while some rivals used higher settings.
03 / Why execution mattered
Finishing work beat looking thorough
The test rewarded practical judgment: using available information, completing tasks, and respecting boundaries under pressure.
Read, decide, deliver
K3 reportedly found useful details in internal documents, closed a substantial deal, identified security concerns, and resisted deceptive messages.
Rules alone were not enough
The source says Opus 4.8 did not close deals and tried to write into restricted departments—evidence that extensive rules do not guarantee sound execution.
04 / The takeaway for deployment
Test for the work itself
A strong showing in one league is a signal to investigate, not a substitute for evaluating a model in your own operating conditions.
Set realistic tasks
Use operational scenarios with real constraints, documents, and competing priorities.
Include hard cases
Test crises, security threats, misleading requests, and worst-case conditions.
Measure outcomes
Track task completion, security judgment, compliance, and behavior under pressure.
Keep oversight
Check reliability across settings before expanding a model’s role in a company.
05 / What remains unknown
One result, open questions
The test challenges assumptions about who can build capable models. It does not establish a permanent change in the global AI race.
Will K3 perform as well elsewhere?
Its results need to be repeated across industries, management tasks, and different operating conditions.
Does this settle which models are best?
No. The reported outcome describes one scenario; broader evaluations and long-term data are still needed.
What could go wrong in deployment?
Over-reliance, security gaps, and unpredictable responses under stress remain concerns that call for testing and human oversight.
What should companies do next?
Evaluate models in realistic workflows, probe failure cases, and verify compliance before relying on them for critical work.
Implications of Chinese AI Surpassing Western Models
This breakthrough indicates that AI models developed outside the traditional Western tech ecosystem can outperform established models in practical management scenarios. It challenges the assumption that Western AI giants hold a monopoly on effective business management capabilities and suggests that emerging Chinese models may be more adaptable and disciplined in real-world tasks. For companies deploying AI, this raises critical questions about model selection, testing in worst-case scenarios, and the importance of evaluating AI performance in operational settings rather than chat demos alone. The result could accelerate shifts in AI deployment strategies, emphasizing robustness and task completion over superficial competence.
As an affiliate, we earn on qualifying purchases.
Background on AI Model Competitions and Management Testing
Recent years have seen a proliferation of AI models claiming to revolutionize business management, support, and automation. Traditionally, Western firms and research institutions have dominated AI development, with models like GPT-4 and its successors setting industry standards. However, the recent live management test conducted by firmulate.com, known as the Crucible league, provides a new benchmark by evaluating models in a realistic, high-pressure environment. The league involves managing a small software company with real money, crises, and manipulations, offering a more accurate measure of practical AI capabilities than chat-based demos or benchmark scores.
This specific test involved five models managing the same company through a week of crises, negotiations, and security threats, with the goal of closing deals, maintaining discipline, and resisting manipulation. The results, which saw a Chinese model outperform Western counterparts, mark a notable shift and suggest that emerging AI firms outside the traditional centers are rapidly advancing in management skills.
As an affiliate, we earn on qualifying purchases.
Unanswered Questions About Model Robustness and Generalization
While the results are compelling, it remains unclear whether Kimi K3’s performance will generalize across different types of management scenarios or if it was a result of specific optimizations for this test. The long-term reliability, scalability, and safety of deploying such models in diverse business environments are still under evaluation. Additionally, the competitive landscape is rapidly evolving, and other emerging models may soon challenge or surpass K3’s capabilities. Further testing and real-world deployment data are needed to confirm whether this success marks a permanent shift or a situational anomaly.
As an affiliate, we earn on qualifying purchases.
Next Steps for Evaluating AI Management Models
Following these results, industry stakeholders are likely to intensify testing of AI models in operational environments, focusing on robustness, security, and compliance. Companies considering AI deployment should prioritize real-world scenario testing, including worst-case conditions, before full integration. Researchers and developers will also need to analyze what specific features enabled K3’s success and whether similar performance can be replicated or improved. The ongoing evolution of this competitive landscape suggests that the next few months will be critical in determining whether Chinese models will become dominant in practical AI management applications.
As an affiliate, we earn on qualifying purchases.
Key Questions
What makes Kimi K3 different from Western AI models?
Kimi K3 demonstrated superior discipline, security awareness, and ability to read internal documents during live management, which are critical in real-world business scenarios. Unlike some Western models, it did not rely solely on chat quality or superficial reasoning but showed effective task completion and security judgment.
Can this result be replicated in other management tasks?
It is currently unclear whether K3’s performance will generalize across different industries or management challenges. Further testing in varied scenarios is necessary to confirm its broader applicability.
Does this mean Western AI models are no longer the best?
Not necessarily. The results show a specific scenario where a Chinese model excelled, but broader evaluation and long-term performance data are needed before making definitive claims about overall superiority.
What are the risks of deploying such AI models in real companies?
Potential risks include over-reliance on AI decision-making, security vulnerabilities, and unpredictable behavior under stress. Rigorous testing and oversight are essential before full deployment.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
