Grok 4.6 Explained: Pricing, Benchmarks, and What Changed
xAI's Grok 4.6 keeps the same 500K context window as 4.5, but it posts a real jump on coding-agent benchmarks and ships with a two-tier pricing structure that punishes long prompts.
Contents6
xAI shipped Grok 4.6 on August 12, 2026. The pitch isn't a bigger context window or a flashier chat mode. It's long-horizon agent work: multi-step coding, tasks where the model needs to check its own output before moving to the next step, and interactive visual work instead of one-shot replies. xAI says the model now self-tests and verifies its work mid-task more often, rather than just producing an answer and stopping there.
What actually changed
The context window holds steady at 500,000 tokens, the same ceiling as Grok 4.5. Training data runs through February 1, 2026. So if you were hoping for a bigger window, this isn't that release. What did move is how the model handles multi-step work, and the benchmark numbers back that up.
Benchmark scores
On the Artificial Analysis Intelligence Index, Grok 4.6 scores 61, tying OpenAI's GPT-5.6 Sol and sitting one point behind Anthropic's [Claude](https://questloops.com/tools/claude) Fable 5 at 62. The more interesting movement is in the agent-specific tests:
| Benchmark | Grok 4.5 | Grok 4.6 |
|---|---|---|
| DeepSWE | 54 | 65.9 |
| APEX-Agents | 47.1 | 57.5 |
Those aren't rounding-error gains. A 12-point jump on APEX-Agents and an 11.9-point jump on DeepSWE both point at the same thing: this release was built for coding agents, not for winning the general chat leaderboard.
Pricing
Standard usage costs $2 per million input tokens and $6 per million output tokens. Push a prompt past 200,000 tokens and the rate jumps for the *entire* request, not just the overflow: $4 in, $12 out.
| Plan | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Standard (up to 200K tokens) | $2 | $6 |
| Long context (over 200K tokens) | $4 | $12 |
That's a real cost to plan for, not a minor surcharge. Anyone routing large codebases or long documents through the API will want to check where their typical prompt size lands before committing to a workflow built on Grok 4.6.
Where you can use it
Grok 4.6 is live on the xAI API, inside [Cursor](https://questloops.com/tools/cursor), on OpenRouter, and through Vercel and Cloudflare's model routing. If you already pay for one of those, there's a good chance you can try it today without a separate xAI subscription.
Is it worth switching to?
Depends what you're doing with it. If your work is long coding sessions where the model has to verify its own steps, the DeepSWE and APEX-Agents jumps are the numbers that matter, not the flat Intelligence Index score. For general chat and writing, the tie with GPT-5.6 Sol and the one-point gap behind Claude Fable 5 mean quality alone won't pull you off whatever you're already using. The context window sitting still at 500K is the one spot where Grok 4.6 is treading water while Google keeps pushing Gemini toward much larger windows.
Quick answers
**Did the context window get bigger?** No, it's still 500,000 tokens, unchanged from Grok 4.5.
**What's the actual price difference for long prompts?** Anything over 200,000 tokens bills the whole request at $4/$12 per million instead of $2/$6, roughly double.
**Where is it available?** xAI's own API, Cursor, OpenRouter, Vercel, and Cloudflare.
