SpaceXAI

AI · Foundation Models · Developer Tools

Launch

SpaceXAI launches Grok 4.6, tuned for long-running AI agents

The model now pauses mid-task to check its own work, a shift toward agents that self-correct rather than just produce one fast answer.

• SpaceXAI (formerly xAI) released Grok 4.6 on August 12, 2026, a flagship model focused on long-running agents, coding, and knowledge work, less than a month after Grok 4.5.
• Grok 4.6 is built to stick with multi-step jobs like researching a topic, navigating a large codebase, or turning a rough idea into a working app, available immediately in Cursor and Grok Build.
• Training involved regenerating supervised fine-tuning trajectories with Grok 4.5 across reasoning, agent harnesses, and domains like STEM and software engineering, then reinforcement learning on tasks such as kernel optimization and CAD.
• Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Sol and trailing Claude Opus 5 (63) and Claude Fable 5 (62), up 5 points from Grok 4.5's 56.
• Pricing stays at $2/$6 per million input/output tokens, over 60% cheaper than Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30), while cache-hit pricing rose to $0.5 per million tokens from $0.3.
• On DeepSWE v1.1 coding benchmark Grok 4.6 jumped to 65.9% from Grok 4.5's 54%, but still trails GPT-5.6 Sol Max's 73%, showing gains outpace but don't yet close the gap with rivals.
• The release underscores a shift in the frontier AI race from raw one-shot intelligence toward agents that self-verify over long task chains, now the key battleground among top labs.