Recently, SpaceX AI officially launched its upgraded flagship artificial intelligence model, Grok4.6. This model focuses on handling complex, long-term agent tasks and has demonstrated strong performance that matches or even exceeds the leading OpenAI model GPT-5.6Sol in multiple core benchmark tests.

Looking back at its technological evolution, SpaceX AI had previously collaborated with Cursor to launch Grok4.5 in July, which was trained using the computing power of tens of thousands of NVIDIA GB300 GPUs. Building on this foundation, Grok4.6 achieved a further breakthrough. Its training process focused on comprehensive optimization for long-term complex tasks, extensively utilizing model-generated data that includes reasoning capabilities, advanced technical concepts, and high-quality engineering code. Additionally, the model underwent extensive agent-based reinforcement learning training, covering knowledge work, general programming, and multiple specialized domain environments, enabling it to produce higher-quality initial results when building visual and interactive applications.

image.png

In the comprehensive index of nine popular AI benchmark tests by the authoritative evaluation organization Artificial Analysis, Grok4.6's overall score has caught up with GPT-5.6Sol, currently trailing only Claude Opus5 and Fable5Max. Specifically, in the GDPval-AA v2 test evaluating logic and knowledge level, Grok4.6 achieved an Elo score of 1753; and in the private benchmark test AA-Briefcase assessing long-term knowledge work agent tasks, its Elo score reached 1577 points. Both key performances are second only to Claude Opus5.

While delivering significant performance improvements, Grok4.6 continues to maintain a highly competitive operating speed and pricing structure. The model can run at 80 TPS, with the base version's input and output prices being $2 and $6 per million tokens, respectively. The high-speed variant version is $4 and $12 per million tokens. Currently, Cursor and Grok Build users can start using it directly, and developers can also conveniently access it through the API.