Grok 4.6

Grok 4.6, released August 12, 2026 by SpaceXAI (formerly xAI), is the company's latest frontier reasoning model. Built on the same 1.5 trillion-parameter V9 foundation as Grok 4.5, it delivers its gains through a heavily reworked post-training pipeline rather than raw scale, shifting focus from raw intelligence to long-running agents and interactive, visual work — multi-step research, codebase navigation, and turning product ideas into polished applications.

Model Specifications

What's New vs Grok 4.5

• Post-training overhaul: Same 1.5T V9 base, but significantly improved SFT and reinforcement learning, trained on agentic RL tasks spanning knowledge work, general coding, web development, CAD, and kernel optimization.
• Self-verification behavior: The model increasingly tests and verifies its own outputs before moving forward across multi-step tasks.
• Stronger first attempts at visual and interactive projects — it can establish an application's structure and visual language from a product idea before refining.
• Longer supplemental training run using model-generated reasoning data and engineering material.

Benchmark Highlights (self-reported)

• AA Intelligence Index: 61 (up from 56 on Grok 4.5)
• APEX-Agents: 57.5% (up from 47.1%)
• DeepSWE v1.1: 65.9% (up from 54%)
• CursorBench v3.2: 69.9% (up from 66.7%)
• Harvey LAB (Vals): 15.8%, the best professional/legal work score in its comparison set
• Known weakness: Terminal-Bench v3.0 at 26% trails rivals (34-35%), a genuine gap in real terminal work

Pricing

Consumer plans include a limited free tier, SuperGrok at $30/month, SuperGrok Plus at $100/month, and SuperGrok Heavy at $300/month. At $2/$6 per million tokens, Grok 4.6 is roughly half the price of comparable frontier models like Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30).

Notes

All benchmarks are self-reported by SpaceXAI with no independent replication yet. The 1.5T parameter count comes from Elon Musk's X posts rather than a formal model card. A successor, Grok 4.7 (2.1T parameters), is expected in late August or early September 2026.

Conclusion

Grok 4.6 shows that xAI's competitive edge now lies in agentic training and interactive work rather than model scale alone. For agent builders and heavy coding users, it offers frontier-level capabilities at roughly half the API cost of its main competitors — with real-time X data access remaining its signature differentiator.