Scale · founder · 7 min read

Grok 4.6: a real jump for long agent runs, and a pricing cliff at 200K

SpaceXAI's Grok 4.6 is live in Cursor and Grok Build. The headline is $2/$6 per million tokens. The number that actually hits your bill is different.

Published August 21, 2026 ·

spacexai grok grok-4.6 cursor models pricing

SpaceXAI — the division of SpaceX that used to be xAI — released Grok 4.6 on August 12, 2026. Nine days later it’s still sitting at the top of Product Hunt, which tells you something about how much attention this category gets and not much about whether the model is good.

Here’s the version that matters if you’re building with these tools rather than benchmarking them.

Grok 4.6 is a genuine, large improvement over Grok 4.5, which shipped only five weeks earlier. It’s live in Cursor, in Grok Build, and through the API on OpenRouter, Vercel, and Cloudflare. And its advertised price is misleading in a way that will specifically bite people who use it the way vibe coders use models.

The pricing cliff nobody is putting in the headline

The number in every writeup is $2 per million input tokens and $6 per million output. Cached input is $0.50. That’s mid-priced for a frontier model — roughly a third of what Claude Opus 5 costs and well under GPT-5.6 Sol.

That rate holds only while your prompt stays under 200,000 tokens.

Cross 200K and the rate becomes $4 input, $1 cached, $12 output — and per xAI’s own API docs, the higher rate applies to every token in that request, not just the ones past the line. There’s no gradual ramp. One request at 199K tokens and one at 201K tokens cost roughly double per token, for the same work.

The model supports a 500,000-token context window, so more than half of its advertised context sits on the expensive side of that line.

Why this matters more for you than for an enterprise buying inference by the truckload: long agentic coding sessions are exactly the workload that accumulates context. A conversation where the agent reads twelve files, runs tests, reads the failures, and tries again is a conversation that quietly climbs toward 200K. You won’t see the crossing happen. You’ll see it in the bill.

If you’re using Grok 4.6 through Cursor or Grok Build on a subscription, this is abstracted away from you and you can stop worrying about it. If you’re calling the API directly — or building something on top of it — budget for the $4/$12 band, not the $2/$6 one.

Is it actually better?

Yes, clearly, over its predecessor. Less clearly over its competitors.

The generational gains are large and consistent. On DeepSWE v1.1, Grok 4.6 goes from 54% to 65.9%. On Terminal-Bench v3.0, from 15.7% to 26%. On APEX-Agents, from 47.1% to 57.5%. On CursorBench v3.2, 66.7% to 69.9%. These are not rounding errors. Five weeks between releases and double-digit jumps on several evals is a fast cadence.

But the frontier is crowded. GPT-5.6 Sol Max leads DeepSWE at 73%. Claude Fable 5 Max leads CursorBench at 70.5%, FrontierCode at 63.6%, and APEX-Agents at 59.2%. Terminal-Bench is the widest gap: Grok’s 26% against roughly 34% for both Sol Max and Fable 5 Max. On the independent Artificial Analysis Intelligence Index, Grok 4.6 scores 61 — good enough for third or fourth place depending on how you break its tie with GPT-5.6 Sol Max, behind Claude Opus 5 and Fable 5.

One caveat worth carrying: SpaceXAI’s comparison table uses the best self-reported or publicly available scores for competing models. That’s standard practice and it’s disclosed, but it is not a controlled four-model bake-off, and you shouldn’t read it as one.

The efficiency claim is the interesting part

Artificial Analysis measured Grok 4.6 completing its AA-Briefcase workloads in about 53 turns and 0.5 billion input tokens, versus roughly 103 turns and 2 billion tokens for Claude Opus 5 Max. Half the turns, a quarter of the tokens.

If that holds in production, it’s a bigger deal than any benchmark row, because the thing you actually pay for is finishing a task — not generating a million tokens. A more expensive model that needs half as many attempts is the cheaper model.

The honest counterpoint: Artificial Analysis puts Grok 4.6 at $0.84 per task on its intelligence-versus-cost frontier, which is worse than Grok 4.5 and worse than GPT-5.6 Luna, GLM-5.2, and Meta’s Muse Spark 1.2. So the efficiency story and the cost story disagree with each other right now. Agent costs depend enormously on harness design, caching, retries, and what you’re actually asking for. Treat both numbers as directional.

Where you can use it today

• Cursor — in the model picker now, on paid plans.
• Grok Build — SpaceXAI’s terminal coding agent. No standalone subscription; access comes through SuperGrok at $30/mo, X Premium+ at $40/mo, or SuperGrok Heavy at $300/mo.
• API — direct from xAI, or via OpenRouter, Vercel, and Cloudflare.

SpaceXAI ran double included usage in Cursor and Grok Build for the first week after launch. That window closed on August 19; if you tested it then and liked the economics, re-check them.

The part that isn’t about the model

Grok 4.6 arrived the day after Grok Bot, the first joint product from SpaceX/xAI and Cursor’s parent company Anysphere — persistent agents with their own cloud machines that keep working after you close your laptop. Access starts at $120/seat/month on Cursor Teams Premium, $200/mo on Cursor Ultra, or SuperGrok Heavy. That’s a different product for a different buyer, but it’s the clearest signal yet of what the Cursor acquisition was for.

And there’s a governance question you should price in if you sell to anyone regulated. The Grok brand carries a documented history — antisemitic outputs in 2025, the “white genocide” prompt-injection incident, and, more seriously, open investigations by U.K. regulator Ofcom, the U.K. Information Commissioner’s Office, and the European Commission under the Digital Services Act, largely concerning non-consensual sexualized imagery generated through Grok on X.

None of those findings concern Grok 4.6’s coding output, and it would be unfair to imply they do. But procurement teams don’t evaluate models in isolation from their vendors. If your customers include banks, hospitals, or government agencies, “which model is in your stack” is a question you may have to answer in a security review. Factor that in before you make Grok 4.6 your default rather than one option in a picker.

The bottom line

If you’re already in Cursor, try it — it’s free to switch in the model picker and it’s noticeably better than Grok 4.5 at staying on task through a long run. If you’re calling the API, assume the $4/$12 band. If you’re building something you’ll sell into a regulated industry, keep a second model configured.

And if you’re a non-technical founder wondering whether any of this changes what you should be using: it doesn’t, yet. The tools you build in matter far more than the model inside them. Which model powers your vibe coding tool has the fuller version of that argument.