Wes Roth compares Grok 4.6 High with leading Anthropic and OpenAI models and says the release appears competitive rather than merely benchmark-optimized. He attributes the improvement over Grok 4.5 to a longer supplemental training run, higher-quality engineering data and changes to the optimizer and post-training recipe.
His first practical test asked Grok Build to create a portal-based game environment. Over several hours and with some steering, the model produced connected portals, conserved momentum, a playable puzzle, generated assets and voice effects. A second prototype connected an emulator, synthetic speech and chat controls for an automated game streamer, though it remained incomplete.
Roth also examines xAI's broader agent strategy. Grokbot gives assistants cloud virtual machines, delegated sub-agents, recurring routines and cross-device access, while Grok 4.6 is priced below comparable frontier APIs. He treats future Grok claims as unverified, but sees the current release and always-on agent design as evidence that xAI has become a serious frontier competitor.
Watch on YouTube
