Grok 4.6 is here
xAI just launched Grok 4.6, a little over a month after Grok 4.5. I loved Grok 4.5 because it was fast, reliable, and reasonably priced.
This time, as explained by xAI, the focus is on a few things:
- Staying with complex tasks across many steps
- Better self-testing and verification during longer tasks
- Stronger first passes on visual and interactive projects
On xAI's benchmarks, Grok 4.6 High scores 61 on the Artificial Analysis Intelligence Index, exactly matching GPT-5.6 Sol Max and coming just behind Fable 5 Max at 62. Benchmarks are benchmarks, though, and I am more interested in seeing how it feels during real coding work.
I have used it a little inside Grok Build, and my first impression is that it feels a bit slower than Grok 4.5. It is still fast enough, and the overall quality is good, but the difference is noticeable. I have not used it enough yet to say whether the stronger reasoning and long-running agent capabilities make up for the slower responses.
The normal API pricing is unchanged, but cached input is now more expensive.
| Price per 1M tokens | Grok 4.5 | Grok 4.6 |
|---|---|---|
| Input | $2 | $2 |
| Cached input, short context | $0.30 | $0.50 |
| Cached input, long context (≥200k) | $0.60 | $1.00 |
| Output | $6 | $6 |
Grok 4.6 has a 500k context window, reasoning levels up to xhigh, and a fast variant that costs twice as much. It is available through the xAI API, Grok Build, and Cursor. I am still on the 3-month SuperGrok offer, so I tried it in Grok Build.
Apart from this, xAI is offering 2x the usual included usage in Grok Build and Cursor for the first week.
I also discovered this write-up from Eric at Cursor, and it has some really good suggestions about using Grok 4.6. He has been using it as his daily driver for a few weeks. Here are the tips that stood out.
- skip phrases like "work very hard", they barely change the outcome
- a short prompt plus a clear preference is usually enough, because the model's taste is good
- write a long spec only when you already know exactly what you want
- say what done means, instead of telling it to keep going
- ask it to open the app, click through real user paths, and fix what it finds
- for visual work, tell it to capture a screenshot, list what is wrong, then fix only those things
- do not trust a summary that says it is finished, run it and look
The verification one is the highest leverage. He tried a two-page spreadsheet spec and a three-sentence prompt, and both apps came back almost the same. Adding that one "use it and check" sentence is what actually changed the result.
Cheers!