Grok 4.6: Complete Guide to Pricing, Benchmarks, and the New xhigh Tier (2026)
xAI released Grok 4.6 on August 12, 2026, as grok-4.6 in the API — the successor to Grok 4.5, SpaceXAI’s Cursor-co-trained coding model. Everything below is verified against xAI’s own documentation (docs.x.ai/developers/grok-4-6) and the official launch announcement, not third-party reporting or aggregator listings.
What changed from Grok 4.5
Confirmed unchanged, per xAI’s own docs:
- Context window stays at 500,000 tokens
- Standard API pricing stays at $2.00 per million input tokens, $6.00 per million output tokens — identical to Grok 4.5
- Cached input pricing carries over from 4.5
Confirmed changed:
- A new
xhighreasoning effort level joins low/medium/high (high remains the default) - xAI’s own benchmark table shows Grok 4.6 improving over Grok 4.5 on every evaluation it reports, including a jump on DeepSWE v1.1 from 54% to 65.9%
- Knowledge cutoff moves to February 1, 2026
- Available via xAI API, Grok Build, Cursor (all plans), OpenRouter, Vercel, and Cloudflare
Not confirmed / could not verify:
- xAI does not publish a parameter count for Grok 4.6 anywhere in its documentation or launch materials. Treat any specific parameter figure you see elsewhere as not vendor-confirmed.
- We could not verify claims of a separate long-context pricing tier or threshold for Grok 4.6 in xAI’s own pricing documentation. If you see that claim elsewhere, treat it as unverified until xAI’s pricing page reflects it.
- xAI states its own benchmark table uses “the best of self-reported or publicly available results” for competitor figures — meaning the comparison numbers on xAI’s page are not independently audited, even where Grok 4.6’s own score is real.
The xhigh reasoning tier
Grok 4.5 already supported configurable reasoning across low, medium, and high. Grok 4.6 adds a fourth level, xhigh, above the existing high tier. As with Grok 4.5’s reasoning levels, higher effort trades more output tokens (and higher cost) for better results on harder problems — this is the same tradeoff pattern used by Claude’s effort levels and OpenAI’s reasoning tiers, just with an added tier at the top end.
Practically: xhigh is the tier to reach for on the hardest architecture decisions or debugging sessions, not for routine completions. Given that pricing is unchanged from 4.5, the cost of using xhigh comes entirely from the extra tokens it generates, not from a higher per-token rate.
Coding relevance
Grok 4.6 inherits Grok 4.5’s core positioning: a coding model co-trained with Cursor, meaning it understands IDE-specific patterns — tab completion, multi-file edits, project-aware context — as part of its base training rather than through post-training fine-tuning. Nothing in xAI’s Grok 4.6 materials indicates a change to that training relationship. If you were choosing Grok 4.5 specifically for its Cursor integration, that reason carries over unchanged to 4.6.
The DeepSWE v1.1 jump from 54% to 65.9% is the most concrete coding-relevant number xAI has published for this release. DeepSWE is an agentic software engineering benchmark; an improvement of that size, if it holds up under independent testing, would be a meaningful jump rather than a marginal one — but as with all vendor-reported benchmarks, this is xAI’s own number until third parties reproduce it.
Pricing and access
| Spec | Grok 4.6 |
|---|---|
| API model string | grok-4.6 |
| Release date | August 12, 2026 |
| Context window | 500,000 tokens |
| Input pricing | $2.00 per million tokens |
| Output pricing | $6.00 per million tokens |
| Reasoning levels | low, medium, high (default), xhigh |
| Knowledge cutoff | February 1, 2026 |
| Availability | xAI API, Grok Build, Cursor (all plans), OpenRouter, Vercel, Cloudflare |
Because pricing is unchanged from Grok 4.5, Grok 4.6 sits in the same place in the broader AI API pricing landscape as its predecessor did — see that guide for how $2/$6 compares against Claude Sonnet 5, GPT-5.6, and other current models. One note worth flagging directly from xAI’s own batch pricing documentation: Grok 4.6 has no batch API discount at all, unlike Sonnet 5, GPT models, and Gemini, which all offer a 50% batch discount. If you’re running bulk or offline workloads, that absence is worth factoring into a provider choice, even though Grok 4.6’s synchronous rate is competitive.
Grok 4.6 vs Grok 4.5
| Grok 4.5 | Grok 4.6 | |
|---|---|---|
| Release date | July 8, 2026 | August 12, 2026 |
| Pricing | $2 / $6 per million tokens | $2 / $6 per million tokens (unchanged) |
| Context window | 500,000 tokens | 500,000 tokens (unchanged) |
| Reasoning levels | low, medium, high | low, medium, high, xhigh |
| DeepSWE v1.1 | 54% | 65.9% (xAI self-reported) |
| Knowledge cutoff | Not specified in original release | February 1, 2026 |
There’s no pricing or context-window reason to stay on 4.5. If you’re already using Grok through any of the listed channels, upgrading to 4.6 costs nothing extra and xAI’s own data shows improvement across every benchmark it reports.
Comparisons
For how Grok 4.6 fits against other current models, see Claude Sonnet 5 and GPT-5.6 — both are benchmark-adjacent to Grok 4.6 per third-party tracking (Artificial Analysis reportedly ties Grok 4.6 with GPT-5.6 Sol at an Intelligence Index of 61, though we have not independently reproduced that score). For the Cursor-specific angle, see the Grok 4.5 vs Claude Sonnet 5 comparison, which remains largely applicable since pricing and context are unchanged between 4.5 and 4.6.
Who should use Grok 4.6
Same profile as Grok 4.5: Cursor users, teams optimizing for cost per task rather than cost per token, and anyone needing the 500K context window for large-codebase work. The xhigh tier adds a genuine option for the hardest problems without a pricing penalty beyond the extra tokens it consumes. If you were on the fence about Grok 4.5 specifically because of its hallucination rate or Cursor dependency, nothing in the 4.6 release addresses those tradeoffs directly — see the Grok 4.5 guide for the full limitations list, which still applies.
FAQ
Is Grok 4.6 more expensive than Grok 4.5?
No. Pricing is identical: $2 per million input tokens, $6 per million output tokens, confirmed on xAI’s own pricing documentation.
What is the xhigh reasoning tier for?
It’s a fourth reasoning effort level above high, for the hardest architecture and debugging problems. It costs more only in the sense that it generates more output tokens — the per-token price doesn’t change based on reasoning level.
Does Grok 4.6 have a bigger context window than Grok 4.5?
No, both are 500,000 tokens.
Is Grok 4.6 better at coding than Grok 4.5?
xAI’s own benchmark table shows improvement on every evaluation it reports, including a large jump on DeepSWE v1.1 (54% to 65.9%). This is vendor-reported; we have not independently verified these figures against third-party benchmark runs.
Does Grok 4.6 have a batch API discount?
No. Per xAI’s own pricing documentation, Grok 4.6 is excluded from xAI’s batch discount entirely — only four older/smaller models (grok-4.3 and the grok-4.20 family) qualify for a 20% batch discount. This is worth knowing if you’re comparing bulk-workload economics against providers like Anthropic or OpenAI that offer a 50% batch discount on their flagship models.