SpaceXAI, the AI company Elon Musk previously ran as xAI, has released its new flagship model, Grok 4.6. Arriving roughly a month after Grok 4.5, it targets long-running AI agents, coding, knowledge work and visual application development.

The real story: the base model did not grow

The most notable thing about this release is not what it does but how it was made. Grok 4.6 is not a larger base model. SpaceXAI held the foundation constant and spent the entire improvement on post-training: a longer supplemental run than Grok 4.5 received, regenerated supervised fine-tuning trajectories, and reinforcement learning in agentic environments.

The supplemental run used curated, model-generated data for reasoning and advanced technical concepts. High-quality engineering data went into the mix, and both the optimizer and the training recipe were updated.

Then comes an interesting step: the company used Grok 4.5 itself to regenerate supervised fine-tuning trajectories across reasoning-effort levels, agent harnesses and domains spanning STEM, software engineering and knowledge work. Problematic traces were filtered out by model-based checks. Reinforcement learning followed, using agentic tasks in areas such as GPU kernel optimization, web development and computer-aided design.

Measurement and access

The model scores 61 on the Artificial Analysis Intelligence Index — five points above Grok 4.5 and tied with GPT-5.6 Sol Max.

  • Context window: 500,000 tokens
  • API name: grok-4.6, with pricing starting at $2
  • Default model in Grok Build; available in Cursor on all plans
  • Routable via OpenRouter, Vercel and Cloudflare
  • A new reasoning-effort tier, xhigh, sits above the ladder Grok 4.5 shipped with

There is no open-weights release and no self-hosting path. That means air-gapped deployments cannot use this model at all.

What changed on long tasks

According to the company, Grok 4.6 is better than its predecessor at turning a general product idea into a working first version. It can research a domain it does not know, lay out an application's structure, build the core interactions and refine the output over several rounds of user feedback. On visual and interactive projects the first outputs are stronger, and it can establish an application's structure and visual language in a single pass.

On long-running tasks, the model is meant to hold its context and to test and verify its own work before moving on. That targets the most common failure in multi-step agent work directly: drifting off task halfway through.

All of these claims rest on the company's own account. Independent verification is so far limited to the intelligence index score.