Today marks the launch of Grok 4.6, an upgraded version that builds on its predecessor, Grok 4.5. This new iteration focuses intensively on long-term tasks and advanced interactive and visual capabilities. It adeptly manages complex, multi-step processes, whether that's in-depth research, information analysis, coding tasks, or transforming concepts into fully functional applications or artifacts.
Grok 4.6 is achieving remarkable intelligence across multiple benchmarks in coding and knowledge work, specifically matching GPT-5.6 Sol on the Artificial Analysis Intelligence Index, which aggregates results across nine distinct metrics.
The benchmarking results reveal Grok 4.6 performing exceptionally well alongside competitors. Notably, it achieves strong scores in various evaluations, with competing models listed per their developers’ system cards or specified leaderboards.
Today, Grok 4.6 is accessible via Cursor and Grok Build. To celebrate the launch, users will receive double the included usage during the first week, enabling a chance to explore the features and capabilities of 4.6 right away.
The development of Grok 4.6 involved an extended supplemental training phase compared to Grok 4.5. This included using carefully selected model-generated data for reasoning, sophisticated technical concepts, and high-quality engineering information, paired with an enhanced optimization process. This effort has culminated in a more robust foundation for subsequent supervised fine-tuning (SFT) and reinforcement learning (RL) phases.
Subsequently, Grok 4.5 was utilized to refine SFT trajectories across various reasoning tasks and domains such as STEM, software engineering, and knowledge work. Problematic traces were filtered out using model-based assessments, resulting in an SFT checkpoint that indicates markedly improved performance and behavior.
The model's training regime included a diverse array of agentic RL tasks spanning knowledge work, general coding, and particular fields like kernel optimization and web development. Testing has shown Grok 4.6 excels at converting broad product concepts into functional first versions. Its capabilities extend to researching unfamiliar subjects, structuring applications, executing core interactions, and refining outputs through iterative feedback, demonstrating an impressive sustainability in longer projects.
We also observed an increase in self-assessment and verification during extended tasks, with Grok 4.6 proving capable of validating its own work prior to proceeding.
When it comes to visual and interactive projects, Grok 4.6 outperforms its predecessor, Grok 4.5, by establishing application structure and visual aesthetics in a single pass, making it particularly effective in scenarios where quick iterations based on substantial starting points yield the best outcomes.
Improvements have been made to Grok 4.6's safeguards to better align with its advanced capabilities. The safety framework is crafted to balance utility and security across valid applications, making Grok 4.6 valuable for use cases like vulnerability patching, accelerating engineering design cycles, and enhancing AI research.
Our rigorous evaluation process for safeguards has expanded, incorporating the broadest suite of pre-deployment testing ever seen and extensive post-deployment and third-party assessments.
Grok 4.6 is now available for use in Cursor and Grok Build, as well as through the API and partners like OpenRouter, Vercel, and Cloudflare. Pricing starts at $2 per million input tokens and $6 for output tokens, with an expedited option available at double the price.
To encourage immediate exploration, we're offering double the included usage for the first week inside Grok Build and Cursor, inviting users to engage with Grok 4.6 right away.
