THE VERDICT, UP FRONT
● Ship on Grok 4.6. It exists, it scores 61 on the AA Intelligence Index, and it costs 2 dollars in and 6 dollars out per million below 200K input.
● Do not architect around 4.7. There is no model ID at docs.x.ai, no published price, no context window and no benchmark card.
● Watch the 200K line. Cross it and the entire request reprices to 4 dollars and 12 dollars. Not just the overage. The whole request.
● Realistic 4.7 window: early to mid September, per Musk on August 13. Nothing official from xAI.
The comparison, as far as one is possible
| Spec |
Grok 4.6 |
Grok 4.7 |
| Status |
Shipped August 12, 2026 |
Supplemental training |
| Parameters |
1.5T (V9 foundation) |
~2.1T, reported and unconfirmed |
| Price per 1M tokens |
2 in / 6 out under 200K, 4 in / 12 out above |
None published |
| Context window |
500,000 tokens |
None published |
| AA Intelligence Index |
61 (Grok 4.5 was 56) |
No independent score exists |
| Knowledge cutoff |
February 1, 2026 |
Unknown |
| Serving speed |
Current xAI ceiling |
Musk says slightly slower |
| Token efficiency |
Baseline |
Musk says better |
| Model ID at docs.x.ai |
grok-4.6 |
Does not exist |
Note how many rows on the right say nothing. That is the whole argument. You cannot make a build decision against a column of blanks, and every filled cell on that side traces back to a Musk post on X rather than an xAI model card, benchmark suite, or blog.
Grok 4.6, the model you can actually call
Verdict: the best price-to-index ratio xAI has shipped, with one billing trap you have to design around.
Launched August 12, 2026. It matches GPT-5.6 Sol Max on the Artificial Analysis Intelligence Index at 61 each, at the same 2 dollars and 6 dollars per million that Grok 4.5 carried. Context expanded from Grok 4.5 to 500K tokens. Knowledge cutoff is February 1, 2026.
Where it is live: xAI API, Grok Build, Cursor, and Grok Bot on X. Third-party routing is available through OpenRouter, Vercel and Cloudflare, so you can switch without changing providers. Existing Grok subscribers on X rotate in automatically. Cursor users have to pick Grok 4.6 from the model dropdown to trigger the doubled usage promo.
THE 200K PRICING TRAP
Any request with more than 200,000 input tokens reprices to 4 dollars in and 12 dollars out — and the entire request reprices, not the portion above the threshold. A 210K-token request costs double a 199K-token request end to end. If your agent accumulates context across turns, this is the single most expensive line in your bill and the easiest to miss.
Honest limitation: GPT-5.6 Sol still leads on DeepSWE repo-scale coding. Grok 4.6 leads APEX-Agents and posts 69.9 percent on CursorBench against Grok 4.5's 66.7 percent, so the strength is long-running agent work rather than whole-repository reasoning.
Best for: long-running agents, Cursor workflows, cost-sensitive inference under 200K context.
Grok 4.7, and the strangest delay reason in frontier AI
Verdict: real, further along than xAI usually admits before shipping, and still unplannable.
Musk described it as a new pre-train at 2.1 trillion parameters rather than a post-training refresh of the 4.6 base. His framing of the tradeoff was direct: better than 4.6 in every way, except slightly slower to serve, with better token efficiency.
That tradeoff is worth taking seriously even though the source is a founder post. The argument is that a larger model can be worth using at lower tokens per second if it finishes the same task in fewer total tokens. For Grok 4.5, xAI claimed roughly 80 tokens per second and token efficiency around twice comparable models on some tasks. If 4.7 improves on that again, the effective cost per completed task can fall even as the per-second speed drops.
The delay is the interesting part. On August 13 Musk confirmed initial pre-training was complete and the company had moved into a supplemental run feeding a large volume of SpaceX engineering data, excluding ITAR-restricted material. That pushed the window to roughly three to four weeks out.
Most frontier slips are compute contention or safety evaluation. This one is a deliberate bet on a proprietary corpus no other lab can access, and it signals where xAI thinks the moat is: engineering and physical-systems reasoning rather than general capability. It is also why the AA Intelligence Index may not capture what 4.7 is actually good at, if the claim holds.
Underneath it all: Colossus in Memphis, past 500,000 GPUs across two buildings, with a third facility in Southaven, Mississippi.
Decision framework
| If you are... |
Do this |
| Shipping a product this month |
Grok 4.6. Route through OpenRouter so a swap later costs you nothing |
| Running agents that accumulate long context |
Grok 4.6 with a hard 200K input cap and context trimming, or the pricing doubles |
| Doing repo-scale code generation |
GPT-5.6 Sol still leads DeepSWE. Grok is not the pick here |
| In Cursor already |
Select Grok 4.6 from the dropdown for the doubled usage promo |
| Doing engineering or physical-systems work |
Worth re-evaluating when 4.7 lands, given the SpaceX training corpus |
| Locked into a procurement cycle |
Do not write 4.7 into a contract. No price, no SLA, no model ID |
How to wait without waiting
The practical move is not choosing between them. It is building so the choice does not matter.
- Route through an aggregator. OpenRouter, Vercel and Cloudflare all carry Grok 4.6. Switching model IDs later becomes a config change instead of an integration.
- Cap input at 200K and enforce it in code. Not as a guideline, as a hard limit with context trimming. This survives whatever 4.7 prices at, and it is the difference between a 2 dollar and a 4 dollar effective rate today.
- Keep an eval suite of your own tasks. When 4.7 ships, an independent score will not exist for weeks. Your own 50-task suite gives you an answer on day one instead of day thirty.
xAI shipped Grok 4.5 on July 8 and Grok 4.6 on August 12. At that cadence the cost of switching is the thing to optimise for, not the choice of any single model.
FAQ
When exactly does Grok 4.7 launch?
No date has been published by xAI. As of August 13, Musk indicated roughly three to four weeks with supplemental training running, which points to early or mid September. There is no grok-4.7 model ID at docs.x.ai, no pricing page and no benchmark card.
Is Grok 4.7 confirmed at 2.1 trillion parameters?
Reported, not confirmed. The figure comes from Musk on X. xAI has published no parameter count, and it has not published one for 4.6 through an official model card either.
Why does the Grok 4.6 price double above 200K tokens?
xAI reprices the entire request rather than the overage once input passes 200,000 tokens: 4 dollars in and 12 dollars out per million instead of 2 and 6. A request at 201K costs roughly twice what one at 199K costs, end to end. Cap and trim input in code.
Is Grok 4.6 better than GPT-5.6 Sol?
They tie at 61 on the Artificial Analysis Intelligence Index. Grok 4.6 leads APEX-Agents and posts 69.9 percent on CursorBench. GPT-5.6 Sol leads on DeepSWE repo-scale coding. Pick by workload, not by index score.
What is SuperGrok Heavy and do I need it?
It is xAI's 300 dollar per month subscription tier. For API-based development it is irrelevant, since the API bills per token. It matters only if your usage is through the Grok consumer product.
Will Grok 4.7 be worth switching to immediately?
Unknown, and that is the point. Musk claims better on everything except serving speed. No independent benchmark will exist at launch, and xAI's pattern is to announce on X with the model card following days or weeks later. Run your own eval suite before moving production traffic.