The headline on xAI's page is the price that didn't move. Grok 4.6 landed on 12 August at the same $2 in and $6 out per million tokens Grok 4.5 has charged since July, with five more points on the Artificial Analysis Intelligence Index and the same 500K context window. Fine. Then we opened the docs table, and two numbers had changed underneath the headline. Cached input went from $0.30 to $0.50 per million, a 67% rise on exactly the tokens an agent loop reuses hardest. And the long context band still starts at 200K, where every token in the request reprices at $4 in and $12 out, not just the tokens past the line. Your window is 500K. The cheap part of it is 200K. Here's what shipped, and where the invoice hides.
The short answer
Grok 4.6 shipped 12 August at Grok 4.5 prices, with five more points on the Artificial Analysis Intelligence Index and the same 500K context. The cache rate quietly went from $0.30 to $0.50 per million, and any prompt that reaches 200K reprices the entire request at $4 and $12. Where it genuinely earns the money is turn count: about 53 turns and 0.5B input tokens per long task, against roughly 103 and 2.0B for Opus 5 at max.
What xAI actually announced
Image: xAI
The pitch is long running agents and, in xAI’s words, “more ambitious interactive and visual work”. The company says 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index. The published numbers: CursorBench v3.2 at 69.9%, DeepSWE v1.1 at 65.9%, FrontierCode v1.1 at 61.3%, and a GDPval-AA v2 Elo of 1753. It’s live now in the API, Grok Build, Cursor, OpenRouter, Vercel and Cloudflare, with double the included usage inside Grok Build and Cursor for the first week.
Read the parity claim slowly. Matching GPT-5.6 Sol is parity, not a lead, and Artificial Analysis has Claude Opus 5 at max sitting two points above both on 63. Coming a month after Grok 4.5 launched at the same price, a five point gain is a real month of work. It just isn’t the frontier.
The number nobody put in the headline
Here’s the bit we went looking for. xAI’s own models table lists cached input for grok-4.6 at $0.50 per million below the threshold and $1.00 above it. The same table lists grok-4.5 at $0.30 and $0.60.
That’s a 67% rise, and it lands on the one line item agent workloads cannot avoid. Prompt caching is what makes a fat system prompt and a growing conversation affordable across fifty turns. If your loop reuses a 40K prefix on every call, you were paying $0.012 per turn for that prefix on 4.5. Now it’s $0.02. Multiply by turns, then by concurrent sessions, and it stops being rounding.
Honestly, I don’t think it’s a gotcha so much as a quiet correction: xAI priced 4.5’s cache aggressively and has walked it back. But it’s absent from the announcement, and the announcement is what most people will read.
The 200K cliff, and why it’s worse than it looks
xAI’s pricing note is one sentence: requests whose prompt reaches the listed token threshold are billed at the higher rate for all tokens in the request.
All tokens. Not the overflow. So a prompt at 199,000 tokens costs $0.398 to send. The same prompt at 201,000 costs $0.804. You crossed by 1% and the bill doubled, and the output tokens for that request doubled with it, from $6 to $12 per million.
Nothing about that is hidden, and banded pricing is common enough across the industry. What makes it sharp here is the shape of the product. It’s sold on 500K context and long running agents, and 60% of that window sits on the far side of the cliff. An agent that accumulates history is not going to sit politely at 190K forever. It will drift across, on its own, in the middle of a run, and nothing in the response will tell you it happened.
So if you’re building on this: measure prompt size before you send, and treat 200K as an architectural boundary rather than a billing detail. Compact or summarise the history when you approach it. A truncation step you control is cheaper than a doubling you didn’t plan.
Where it earns the money back
The per token rate is only half the invoice on agent work, and this is the part of the 4.6 story we’d actually pay attention to.
Artificial Analysis measured long horizon tasks and found Grok 4.6 resolving them in roughly 53 turns and about 0.5 billion input tokens on average, against roughly 103 turns and 2.0 billion for Claude Opus 5 at max effort. Same job, four times fewer input tokens, half the round trips. They put the cost at about $0.84 per task.
Which reframes the cache rise, a bit. Paying 67% more per cached million while consuming a quarter of the tokens is still a much smaller bill. It also reframes the Opus 5 comparison we ran last month: Opus keeps the capability lead, and it takes a lot more turns to spend it.
Fewer turns also means less wall clock, which matters more than most cost spreadsheets admit when a human is waiting on the agent.
One thing we could not confirm
The announcement says nothing about the European Union. That’s a live question here, because Grok 4.5 was not available in the EU when it launched on 8 July, and access reportedly opened on 17 July after xAI worked through EU AI Act obligations for models flagged as carrying systemic risk. Whether 4.6 shipped to EU accounts on day one, we don’t know, and we’re not going to guess in a table. Check your own console before you plan a migration around it.
Should you move
If you’re already on grok-4.5 in Cursor or the API, moving is a model string and you get five index points plus xhigh for the hard calls. Re run whatever evals you trust first, then look at your cache spend, because that line is the one that changed.
If you’re on a frontier model for agent work and the invoice is the thing keeping you up, this is worth a bake off on your own tasks. Not because 61 beats 63, it doesn’t, but because finishing in half the turns is a real number that shows up on a real statement.
And if your prompts run long, do the 200K arithmetic before anything else. On a big context agent that single threshold will move your bill more than the choice of model will.
Sources
Model details, benchmark figures, availability and the first week usage offer come from xAI’s own Grok 4.6 announcement of 12 August 2026, whose announcement card is the image above. Pricing for both grok-4.6 and grok-4.5, the cached input rates, the 200K threshold and the rule that a request is billed at the higher rate for all its tokens are from the xAI models documentation, with the effort levels and the note that reasoning cannot be disabled from the xAI reasoning guide. The Intelligence Index scores, the GDPval Elo, and the turn and token counts per long horizon task are from Artificial Analysis. The context window and release date were cross checked against the OpenRouter listing. The Grok 4.5 EU timeline is reported rather than confirmed by us, and we have found no xAI statement covering 4.6.
Frequently asked questions
How much does Grok 4.6 cost?
Two dollars per million input tokens, fifty cents per million cached input, and six dollars per million output, for any request whose prompt stays under 200K tokens. At or above 200K the whole request is billed at four dollars input, one dollar cached and twelve output. Those are xAI's published documentation rates as of 13 August 2026.
Is Grok 4.6 more expensive than Grok 4.5?
On the headline rate, no: input and output are identical at $2 and $6. On cached input, yes. Grok 4.5 charges $0.30 per million cached tokens below 200K and $0.60 above it. Grok 4.6 charges $0.50 and $1.00. That is a 67% rise on the part of the bill that agent loops and long system prompts lean on hardest.
What is the 200K threshold on Grok 4.6?
A pricing band, not a hard limit. The context window is 500K either way. But xAI's pricing note says requests whose prompt reaches the listed token threshold are billed at the higher rate for all tokens in the request. So a 205K prompt does not pay the cheap rate on the first 200K and the dear rate on the last 5K. It pays double on the lot.
Is Grok 4.6 better than Claude Opus 5?
Not on raw capability. Artificial Analysis puts Grok 4.6 at 61 on its Intelligence Index against 63 for Claude Opus 5 at maximum effort, and Opus 5 stays ahead on the GDPval agent Elo. Grok 4.6 wins on cost per finished task, largely because it burns about a quarter of the input tokens over roughly half the turns on long horizon work.
Can I turn reasoning off on Grok 4.6?
No. xAI's documentation says reasoning cannot be disabled on this family. The reasoning_effort parameter picks low, medium, high or xhigh, and high is the default. The xhigh level is new with 4.6; if you send xhigh to Grok 4.5 it is silently treated as high.