Grok 4.6 Review: Features, Benchmarks & Pricing

Grok 4.6 is xAI's flagship model update, released on August 12, 2026, about five weeks after Grok 4.5. xAI built it for long-running agents and more ambitious interactive and visual work — tasks where a model has to research, plan, build, and check its own output across many steps — while keeping Grok 4.5's price unchanged.

This Grok 4.6 review covers what changed, what the benchmark numbers say, what it costs, how it compares with Grok 4.7, which followed in September, and where its limits sit. It also covers access: Grok 4.6 is not yet integrated into Chat Smith, but Grok 4.5 is already live there, so you can use xAI's model line today without a separate subscription.

Grok 4.6 is a direct upgrade to Grok 4.5 rather than a new foundation model. According to xAI, the gains come from longer supplemental training on curated model-generated data, an improved optimizer and training recipe, and better self-checking on long, multi-step tasks.

The basics stay the same: a 500,000-token context window, text and image input, text output, and the API model ID grok-4.6. The one new control is an xhigh reasoning level on top of low, medium, and high, which remains the default. If you read our Grok 4.5 review, most of it still applies — Grok 4.6 simply goes further on long, demanding tasks.

xAI has moved fast since. Grok 4.7 arrived on September 21, 2026, and xAI now recommends it as the default for most work, including code. Grok 4.6 remains available; for an overview of where each version sits, see our full list of Grok models.

Four areas show where Grok 4.6 moves past Grok 4.5.

xAI positions Grok 4.6 for tasks that stay open across many steps: researching a topic, analyzing what it finds, working across a codebase, or turning an idea into a finished app or document. On GDPVal-AA, a test of economically valuable knowledge work, xAI reports an Elo of 1753. On AA-Briefcase, Artificial Analysis's private benchmark of long agentic projects, it debuted at 1577, behind Claude Opus 5. For planning-heavy jobs such as drafting with an AI business plan generator, the stronger self-verification is the most relevant change.

xAI reports 65.9% on DeepSWE v1.1, 69.9% on CursorBench v3.2, and 61.3% on FrontierCode v1.1. It also reports 26% on Terminal-Bench v3.0, a harder version than the v2.1 most leaderboards still use, so that figure should not be compared with v2.1 scores elsewhere. Grok 4.6 had day-one support in Cursor and reached GitHub Copilot's model picker two days after launch.

A key focus of the release is producing stronger first versions of interactive and visual projects — web pages, small apps, dashboards — with more self-testing before the result comes back. The model still outputs only text, so "visual" here means the code and layouts it writes, not generated images.

On Artificial Analysis's current Intelligence Index, Grok 4.6 scores 44.31 against Grok 4.5's 38.81, a large jump for a point release. Output speed also improved, from 59.6 to 72.9 tokens per second. The trade-off is that it thinks longer before answering: time to first token roughly doubled, from about 14 seconds to about 28.

The table below compares Grok 4.6 with the model it replaced and the one that replaced it. Benchmark, speed, and cost-per-task figures come from Artificial Analysis's current Intelligence Index; prices are xAI API rates per million tokens.

The pattern is clear. Grok 4.6 is noticeably smarter than Grok 4.5 at the same token price, but it starts answering more slowly and spends more reasoning tokens per task, so a finished job costs more. Grok 4.7 raises the ceiling again at the same list price.

Grok 4.6 costs $2 per million input tokens and $6 per million output tokens, with cached input at $0.50 — the same as Grok 4.5. Prompts of 200K tokens or more are billed at $4, $1, and $12 across the entire request, not just the tokens past the threshold. xAI's x_search tool, which pulls live posts and threads from X, costs $5 per 1,000 calls.

• Developers: available as grok-4.6 in the xAI API and through OpenRouter, Vercel, and Cloudflare.
• Cloud platforms: generally available on Amazon Bedrock, Google's Model Garden, and Microsoft Foundry.
• Coding tools: available in Cursor, GitHub Copilot, and xAI's Grok Build.
• Chat Smith: Grok 4.6 is not integrated yet. You can use Grok 4.5 today on Android, iOS, and the web, with higher usage limits on Chat Smith Pro at $9.99 per month or $39.99 per year.

Grok 4.6 has already been superseded. Grok 4.7 costs the same and scores higher, and xAI now points most users to it. If you build on the API, switching is a one-line change of model name.

It is close to the frontier rather than at it. On Artificial Analysis's index, it trails Claude Opus 5 by about 6.5 points, and on Arena's blind human-preference leaderboard it placed far lower than its benchmark scores suggest, with a preliminary rank of 48 in late August.

Practical limits matter too. The 500K context window is small next to rivals that offer 1M or more, the doubled rate starts at 200K, and a wait of roughly 28 seconds before the first token is noticeable in interactive use. The xhigh setting also adds cost and latency without a measurable gain on Artificial Analysis's index. Test Grok 4.6 against alternatives such as GPT-6 Sol on your real tasks before committing.

Grok 4.6 is not integrated into Chat Smith yet, but its predecessor is. Grok 4.5 is live in Chat Smith on Android, iOS, and the browser — pick it from the model list and start chatting, with no xAI account or Grok subscription needed.

Grok 4.5 shares Grok 4.6's 500K context window and price tier, and it actually starts answering faster, at about 14 seconds to first token against roughly 28. For everyday questions, writing, and research — such as working through long reports with an AI article summarizer — it remains a fast, capable choice. Grok 4.6's biggest gains show up mainly in long agentic and coding tasks.

Chat Smith also lets you compare Grok 4.5 with Gemini 3.5 Flash and other leading models in the same app, so you can put one question to several models and keep the answer that holds up. Chat Smith Pro adds higher usage limits and more premium models for $9.99 per month or $39.99 per year.

Grok 4.6 is a meaningful step up from Grok 4.5: smarter on long agentic and coding work, faster once it starts writing, and unchanged in price. Its weak spots are slower first responses, a smaller context window than its rivals, and the fact that Grok 4.7 has already replaced it as xAI's recommended model.

Until newer Grok versions reach more apps, Grok 4.5 remains an easy way to use xAI's models every day. You can try it now in Chat Smith, an AI chatbot that puts it next to the latest models from OpenAI, Anthropic, Google, and DeepSeek.