OpenAI’s rumored GPT-6, with 10 trillion parameters, has intensified the AI race, pushing capabilities and risks to extremes while heralding an era of massive capital expenditure. The risk of autonomous malicious behavior has forced multiple labs to cage their most advanced models, making the core competition a game of computing power, capital, and safety control. Across top models, Alibaba’s Qwen 3.8 has 3.5 trillion parameters, Moonshot AI’s Kimi K3 has 2.8 trillion, and OpenAI’s Claude Mythos 5 is estimated at 8 trillion. DeepSeek-R1, which triggered the “DeepSeek moment” in 2025, had 671 billion parameters, showing frontier models have swelled by more than an order of magnitude in over a year. GPT-5.6 Sol can independently complete an image compression tool in under two hours, handling the full workflow, while Mythos 5, after targeted access to cybersecurity agencies, autonomously identified over 10,000 high-risk or critical software vulnerabilities and can execute full-chain operations from code auditing to constructing attack payloads. Parameter count determines memory and expression capacity, but intelligence also depends on data, training methods, and architecture. Alibaba’s DAMO Academy M6 reached 10 trillion parameters in 2021 but was an early, limited form relying on sparsification and CPU offloading, incomparable to today’s end-to-end