Grok 4.6 Is Out: Agents, Pricing, And What To Do With It
SpaceXAI released Grok 4.6 on 12 August 2026. The company is positioning it as the default for code, chat, and agent work: a longer training run than Grok 4.5, more self-testing on long trajectories, and a stronger first pass on visual and interactive projects. Knowledge cutoff is 1 February 2026. Context window is 500K tokens, which is half of DeepSeek V4's 1M.
It is in Cursor and Grok Build today, and in the API (plus OpenRouter, Vercel, and Cloudflare). SpaceXAI is offering 2x included usage in Grok Build and Cursor for the first week. After that you are on the meter.
Official Prices
Per one million tokens, from the xAI model docs:
| Prompt size | Input | Cached input | Output |
| --- | --- | --- | --- |
| Under 200K tokens | $2.00 | $0.50 | $6.00 |
| 200K tokens or more | $4.00 | $1.00 | $12.00 |
Two details that will show up on an invoice.
Long-context pricing applies to the whole request. Once the prompt crosses 200K tokens, every token in that request (input and output) is billed at the higher rate. A 199K prompt and a 201K prompt are not "almost the same bill." They are two different products.
The fast variant is twice the price. SpaceXAI's launch note says so. Do not put it on a high-volume path because the latency graph looked nicer in a demo.
Cached input is $0.50 per million under 200K, which is a 75% discount, not the 90% some other vendors advertise. Still worth designing for. Rate limits on the API are 150 requests per second and 50 million tokens per minute.
What The Benchmarks Are Claiming
SpaceXAI says Grok 4.6 High matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 61, and leads or sits near the top of several agent and coding boards they published (GDPVal-AA, CursorBench 3.2, FrontierCode 1.1, APEX-Agents). Those figures are from the company's own announcement, mixing self-reported and public competitor scores. Use them as a screening filter, not as an acceptance test.
The part that matters for a build decision is the job they trained for: long-running agents that stay on a task across many steps, plus a better first pass on interactive apps. If your workload is classification and extraction at volume, this is the wrong model. DeepSeek V4 Flash is an order of magnitude cheaper on that mix.
How It Sits Next To DeepSeek V4
On a 6,000-input / 800-output request, Grok 4.6 costs about $0.0168. Flash costs about $0.00106. Pro costs about $0.00331. Grok is the frontier-priced option in this week's set, Flash is the volume option, Pro is the cheap-but-heavier reasoning option.
That is the routing table, not a loyalty test. Pin `grok-4.6` only on the steps where the extra capability shows up in your evaluation set: multi-step coding agents, long research traces, visual app first drafts. Send the plumbing to Flash. Escalate the hard 10% to Pro or to Grok, and measure which one actually wins on your cases.
Grok 4.6 also has a product wrapped around it. Grok Bot, launched a day earlier at x.ai/bot, is SpaceXAI's bet that the model should arrive as a teammate with its own computer, not as another chat completion.
What To Do This Week
If you already run Cursor or Grok Build, spend the 2x-usage week on a real internal task, behind an evaluation set, and keep the traces. A launch-week demo is not a migration.
If you are on the API, pin the dated model name, log tokens and cost per step from day one, and do not let a 200K-token prompt happen by accident. The doubled rate is the kind of bug that looks like a successful agent run until finance asks why.
If you do not have an evaluation set, that is the project. Grok 4.6, DeepSeek V4, and whatever ships in September will keep arriving. The teams that absorb them are the ones who already built the thin swap layer and the acceptance tests. A scoped DX audit is how we put that layer in place.
---
Sources: Introducing Grok 4.6 · xAI Grok 4.6 docs · DeepSeek pricing