Grok 4.6, xAI's newest coding model, is now selectable in CRHQ through the Cursor provider. It joins Grok 4.5 and the rest of the Cursor lineup, and it is live on every instance.
What it is
Grok 4.6 is a focused upgrade to Grok 4.5, built for long-running agentic work: tasks that run across many steps without losing track of the goal. It is not a fresh model from scratch. It is a new round of post-training on Grok 4.5, which is why the gains show up most in agentic coding and reliability rather than raw knowledge.
It came out of xAI's acquisition of Cursor, which gave xAI a large body of real coding data and the tooling to turn it into a more dependable coding model. That is the same Cursor provider CRHQ already uses, so Grok 4.6 arrives through a path you may already have connected.
How it benchmarks
The numbers below are as reported by xAI and by independent trackers, not our own measurements. On the whole, Grok 4.6 lands in the same tier as the leading coding models, at a lower price point.
| Benchmark | What it measures | Grok 4.5 | Grok 4.6 |
|---|---|---|---|
| DeepSuite | Real-world coding feel | ~54% | ~66% |
| Terminal-Bench | End-to-end CLI and tool tasks | ~15% | ~26% |
| Frontier Code | Hard coding problems | ~57% | ~61% |
On Artificial Analysis's overall intelligence index, Grok 4.6 is reported as roughly neck and neck with GPT-5.6 Sol and just behind Fable 5 and Opus 5, which places it around third overall. On OpenAI's GDPval knowledge-work benchmark, xAI reports Grok 4.6 High taking the top score.
One honest caveat worth repeating: some of Cursor's own benchmark data was reportedly trained into the model, so its CursorBench score should be read with a little skepticism. The independent benchmarks above are the better guide.
What people are saying
"It's fast, it's cheap, it's reliable, and it's good at doing long running tasks with lots of sub agents and not losing track of what it's supposed to be getting done." — Theo, t3.gg
"This is a phenomenal model, competitive with the top models out of OpenAI and Anthropic." — Matthew Berman
The recurring theme from reviewers is speed, price, and staying coherent across long agentic runs. It is the first non-Anthropic, non-OpenAI model several of them say they would actually reach for by default.
How to use it
- Start a new session, or edit an agent's default model.
- Choose provider Cursor.
- Pick Grok 4.6 from the dropdown.
- Optionally set an effort level. Grok 4.6 adds a native xhigh tier that Grok 4.5 did not have, so at the higher effort settings you get more thinking than 4.5 could offer.
- Send your message as normal. Delegation, artifacts, skills, and tools all work identically.
As with the other Cursor models, your satellite needs a Cursor account connected, under Settings, then AI Models, then Cursor. Usage is billed through your own Cursor plan.
Cost tracking
Per-turn tokens and cost appear in each session's Costs tab. Grok 4.6 is a Cursor subscription-pool model, but we track its API-equivalent value the same way we do for Claude and GPT, using xAI's published rates.
| Model | Input, per 1M tokens | Output, per 1M tokens | Cached input |
|---|---|---|---|
| Grok 4.6 | $2.00 | $6.00 | $0.50 |
| Grok 4.5 | $2.00 | $6.00 | $0.30 |
Grok 4.6 raised only the cached-input rate over 4.5. These are equivalent token values based on xAI's rates. Actual billing happens on your Cursor plan. Turns that ran before this pricing landed keep their original recorded value.
Availability
| Fact | Value |
|---|---|
| Availability | All CRHQ satellites |
| Live since | 2026-08-15 |
| Provider | Cursor |
| Client action required | None, beyond connecting a Cursor account if you have not already |
| Billing | Through your own Cursor plan |
A note on pace: reviewers point out that xAI has been shipping quickly since the Cursor acquisition, and a Grok 4.7 has already been teased. Because these models come through Cursor, when the provider ships a new one we can surface it the same way, without you changing anything.