Grok 4.6 launched on August 12, 2026, as xAI's frontier model built specifically for long-running agents and multi-step coding work. The model ties OpenAI's GPT-5.6 Sol Max at 61 on the Artificial Analysis Intelligence Index, a composite benchmark, but the row-by-row results tell a more complicated story. Fable 5 Max leads on most coding benchmarks, while Grok 4.6 wins on knowledge-work tasks like document reasoning and professional analysis.

What Actually Changed From Grok 4.5 to 4.6?

xAI did not build a larger model. Instead, the company ran a longer supplemental training phase on top of Grok 4.5's existing foundation, then regenerated fine-tuning data and applied reinforcement learning in agent environments. The underlying model remains at roughly 1.5 trillion parameters, the same scale as 4.5. What is new is the focus: Grok 4.6 is tuned to stay on task across many steps without drifting, and xAI reports that on longer runs it started checking its own work before moving forward.

The model also adds a new "xhigh" reasoning setting above the existing low, medium, and high levels, and it received supplemental training on anonymized Cursor workflow data through a collaboration with the coding IDE Cursor. The context window remains 500,000 tokens, unchanged from 4.5, so anyone claiming a context bump is mistaken.

How Does Grok 4.6 Actually Perform Against Rivals?

The benchmark picture is mixed. On xAI's published evaluation table, Grok 4.6 High wins three rows outright: GDPVal-AA v2, AA-Briefcase, and Harvey LAB, a document-understanding and data-reasoning benchmark. However, the margins on GDPVal and AA-Briefcase are narrow, only a handful of Elo points, and both benchmarks exclude Claude Opus 5, which leads the overall Intelligence Index. That leaves Harvey LAB as Grok's clearest win, though the gap versus GPT-5.6 Sol (15.8% versus 2.5%) is wide enough to warrant independent verification.

Fable 5 Max leads on the coding-focused benchmarks that matter most to developers:

• CursorBench v3.2: Fable scores 70.5% versus Grok's 69.9%, a narrow edge but a win nonetheless
• FrontierCode v1.1 Extended: Fable reaches 63.6% compared to Grok's 61.3%
• APEX-Agents and APEX-SWE: Fable leads both agentic coding benchmarks, scoring 59.2% and 58.8% respectively
• DeepSWE and Terminal-Bench: GPT-5.6 Sol Max takes these two, with Grok posting its weakest result on Terminal-Bench v3.0 at 26%

Versus Grok 4.5, the upgrade is unambiguous across every published benchmark. Grok 4.6 gains 5 points on the composite Intelligence Index, 227 points on GDPVal, and double-digit improvements on most coding evals. The price stays the same: $2 per million input tokens and $6 per million output tokens for the standard variant, with a 2x fast variant available at $4 and $12.

Where Can You Actually Use Grok 4.6?

Grok 4.6 is available across multiple platforms starting immediately. Cursor, the AI-powered code editor, got day-one access on all plans. Grok Build, xAI's in-house agent scaffolding tool, ships with the model as the default. The SpaceXAI API offers direct access, and the model is also available through partner gateways including OpenRouter, Vercel, and Cloudflare. xAI is offering 2x included usage in Grok Build and Cursor for the first week as a launch promotion.

One important caveat: Grok 4.6 is a closed, hosted model with no open-weights release and no self-hosting path. If you need to run the model on your own infrastructure, this is not an option.

How to Evaluate Grok 4.6 for Your Workflow

• For agent-heavy work: Test Grok 4.6 on multi-step tasks that require the model to check its own work and iterate. The training specifically targets this use case, so it may outperform rivals on your actual workflows even if benchmarks look mixed
• For pure coding tasks: Run side-by-side comparisons with Fable 5 Max on your codebase. Fable leads most published coding benchmarks, so if raw coding performance is your priority, benchmark it first before committing
• For knowledge work and document analysis: Grok 4.6 shows clear wins on professional reasoning tasks like document understanding and data analysis, so this is a strong fit if your work involves parsing and reasoning over complex documents
• For cost-conscious teams: Grok 4.6 pricing matches Grok 4.5, so there is no premium for the upgrade. The 2x included usage promo in the first week gives you a risk-free window to test before committing to production use

What Does This Mean for the Broader AI Agent Race?

Grok 4.6's design philosophy reveals a shift in how frontier labs are thinking about model releases. Rather than chasing the highest composite benchmark score, xAI optimized for a specific use case: long-running agents that need to stay focused and verify their own work. That is a different bet than building the "smartest" general-purpose model, and it reflects the reality that agent architecture and harness design often matter as much as raw model capability.

The benchmark table also exposes a limitation of vendor-published evaluations. xAI chose which benchmarks appear, and competitor figures are the best of self-reported or publicly available results, not a controlled head-to-head run in a single lab. That means the table shows where each model shines, not necessarily where it would rank if tested identically. For teams making real decisions, independent benchmarking or hands-on testing remains essential.

Grok 4.6 is available now, with the model card and full evaluation table published by xAI on August 12, 2026. The broader rollout tied to the new model is expected to expand access beyond the current early-access tier in the coming weeks.