AI Catchup

Grok 4.6 Launched in Cursor With a One-Week 2x Usage Window, Now Ended

By 10 min read

SpaceXAI released Grok 4.6 on August 12, 2026, focused on long-running agents and visual work. Its first-week 2x included-usage offer in Cursor and Grok Build has ended, and Grok 4.7 succeeded it on September 21 at the same price. Grok 4.6 API pricing starts at $2 per million input tokens and $6 per million output tokens.

Status (October 9, 2026): The first-week 2x offer ended in August. SpaceXAI released Grok 4.7 on September 21, 2026, at the same price and speed as Grok 4.6, and Cursor still offers both.

SpaceXAI released Grok 4.6 on August 12, 2026, and put it in the Cursor model picker the same day. The company, publishing at x.ai, positions the model for long-running agents and more ambitious interactive and visual work, and it offered 2x included usage in Cursor and Grok Build for the first week, an offer that has since ended. API pricing starts at $2 per million input tokens and $6 per million output tokens. (SpaceXAI; Cursor)

This is a post-training step on the Grok 4.5 line, not a claim that one model now wins every coding benchmark. If you are choosing a Grok model today, start with Grok 4.7, which SpaceXAI serves at the same price and speed. Keep a bake-off against GPT-5.6 Sol Max and Fable 5 on your own terminal-heavy tasks before you change a default.

Key Takeaways

  • Grok 4.7 has succeeded it. SpaceXAI released Grok 4.7 on September 21, 2026, "served at the same price and speed as Grok 4.6." (SpaceXAI)
  • Grok 4.6 launched August 12 in Cursor, Grok Build, the API at x.ai, and partner gateways including OpenRouter, Vercel, and Cloudflare. (SpaceXAI; Grok 4.6 docs)
  • Cursor access is on all plans, per SpaceXAI's model overview. (Grok 4.6 docs)
  • First-week promo, now ended: 2x included usage inside Cursor and Grok Build. Cursor's docs dated it one week from August 12, 2026 and dropped the line by August 21. (SpaceXAI; Cursor)
  • API model name is grok-4.6. Context window is 500,000 tokens. SpaceXAI's own docs disagree on the knowledge cutoff: the Grok 4.6 overview says January 2026, while the API models list says February 1, 2026. Inputs are text and image; output is text, with no text output limit. (Grok 4.6 docs; API models list; xAI release notes)
  • Standard API rates are $2 input / $0.50 cached input / $6 output per million tokens below 200,000 prompt tokens, and $4 / $1 / $12 at or above that threshold, billed for all tokens in the request. (Grok 4.6 model page)
  • Reasoning effort supports low, medium, high (default), and xhigh. SpaceXAI says reasoning cannot be disabled. (Reasoning docs)
  • The headline benchmark claim is narrow: SpaceXAI's launch post says Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, a composite of nine benchmarks. The same page's table labels that comparison column GPT-5.6 Sol Max, at 61 vs 61. Fable 5 Max still leads the index at 62. (SpaceXAI)

Where You Can Use It Today

SurfaceWhat SpaceXAI / Cursor say
CursorAvailable from launch day; the first-week 2x included-usage offer has ended; docs say all plans
Grok BuildAvailable from launch day, with the same first-week 2x offer (ended); described at launch as the coding agent's default model
API at x.aiModel name grok-4.6 via the Responses API and Chat Completions
GatewaysOpenRouter, Vercel, and Cloudflare

(SpaceXAI; Cursor; Grok 4.6 docs)

Grok Build installs from SpaceXAI's CLI:

curl -fsSL https://x.ai/cli/install.sh | bash

(SpaceXAI)

Do not conflate this first-week 2x window with Cursor's permanent doubled Cursor Models pool from July. That pool doubling is a separate plan change. The Grok 4.6 posts describe a launch offer inside Cursor and Grok Build and say nothing about how the two relate -- but Cursor's documentation has since answered it. See the update below.

Pricing, Context, and Reasoning

MeterBelow 200k prompt tokensAt or above 200k prompt tokens
Input$2.00 / 1M$4.00 / 1M
Cached input$0.50 / 1M$1.00 / 1M
Output$6.00 / 1M$12.00 / 1M

The 200,000-token threshold is a cliff: SpaceXAI says requests whose prompt reaches that size are billed at the higher rate for all tokens in the request, not only the overflow. Context window is 500,000 tokens. (Grok 4.6 model page; xAI release notes)

The launch posts also mention a fast variant at twice the price. The published API model page currently documents grok-4.6 with the long-context rates above, and does not list a separate fast-variant model ID. Treat "twice the price" as a launch-post claim until you see a named SKU in the picker or the docs. (SpaceXAI; Cursor; Grok 4.6 model page)

For API users, SpaceXAI recommends setting a prompt_cache_key (Responses API) or the x-grok-conv-id header (Chat Completions) so multi-turn agent loops hit the same server and actually get cache discounts. (Grok 4.6 docs)

How Grok 4.6 Compares on Published Benchmarks

These figures are from SpaceXAI's launch table. The company says competitor scores are the best of self-reported or publicly available results. They are not an independent bake-off on your codebase.

EvalGrok 4.6 HighGrok 4.5 HighGPT-5.6 Sol MaxFable 5 Max
AA Intelligence Index61566162
CursorBench v3.269.9%66.7%67.2%70.5%
DeepSWE v1.165.9%54%73%70%
FrontierCode v1.1 (Extended)61.3%56.6%60.6%63.6%
APEX-Agents57.5%47.1%56.7%59.2%
Terminal-Bench v3.026%15.7%34.6%34.1%
APEX-SWE56.4%53.6%--58.8%
GDPVal-AA v21753152617281741
AA-Briefcase1577131315021574
Harvey LAB (Vals)15.8%12.9%2.5%11.3%

(SpaceXAI)

The useful read is the shape, not a crown. Grok 4.6 closes most of the gap on Grok 4.5, ties the launch table's GPT-5.6 Sol Max column on the composite intelligence index, and still trails Sol Max and Fable 5 Max on DeepSWE and Terminal-Bench. If your work looks like long-horizon software engineering in a terminal, those two rows are the ones to test before you switch a default. If your work looks like turning a product idea into a first visual pass, that is the job SpaceXAI is selling.

What SpaceXAI Says Changed

SpaceXAI says Grok 4.6 got a longer supplemental training run than Grok 4.5, then used Grok 4.5 to regenerate supervised-fine-tuning trajectories across reasoning efforts, agent harnesses, and domains including STEM, software engineering, and knowledge work. Reinforcement learning covered knowledge work, general coding, and domain-specific environments such as kernel optimization, web development, and computer-aided design. (SpaceXAI; Cursor)

The product claim that follows is about persistence, not peak IQ. SpaceXAI says the model stays with complex tasks across many steps, is especially strong at turning a broad product idea into a working first version, and showed more self-testing and verification on longer trajectories. It also says first passes on visual and interactive projects are stronger than with Grok 4.5. Those are vendor observations from internal tests, not a public eval you can rerun. (SpaceXAI)

How a Reader Uses This

If you already live in Cursor, test Grok 4.7 before you settle on 4.6. The first-week window is over, and both models bill at the same rates. Rerun one real task you already trust: a multi-file feature, a visual UI pass, or a long agent loop that usually needs babysitting. Compare it against your current default on the same prompt.

If your bottleneck is terminal-heavy agent work, do not take the composite index as the decision. SpaceXAI's own table has GPT-5.6 Sol Max at 34.6% on Terminal-Bench v3.0 versus 26% for Grok 4.6 High. Run that class of task yourself. Cursor's router write-up already argues that no single model wins every turn; this release does not change that.

If you call the API, pin grok-4.6 and watch the 200k cliff. A long agent transcript that crosses 200,000 prompt tokens doubles input, cache, and output rates for the whole request. Set a cache key. Keep a cheaper model in the router for short tool-calling turns.

What This Announcement Does Not Establish

The launch posts do not publish a calendar end date for the 2x Cursor / Grok Build offer, a named fast-variant model ID, or an independent benchmark suite. They also do not say how the first-week 2x offer interacts with Cursor's existing first-party usage pool. SpaceXAI's safety section says safeguards were improved and calibrated, with a wider pre-deployment test suite, but it does not publish those test results here.

Update (August 19, 2026): Cursor's documentation has since closed two of those gaps, and the launch posts still have not. Checked today:

  • The offer had a documented duration, and Cursor's docs called it a discount rather than a doubling. Cursor's models-and-pricing page stated: "A 50% launch discount for Grok 4.6 applies for one week starting August 12, 2026." Cursor's blog described the same offer as "2x included usage ... for the first week." Cursor staff established in July that these are one thing and not two: "50% off and doubled usage for Grok are the same thing, not two separate multipliers." No clock time was ever published for the end of the week. Update, August 21, 2026: the offer is over. That sentence is no longer on Cursor's models-and-pricing page, checked today in both the rendered page and its markdown source. There was no end announcement; the line's removal is the only end signal Cursor gave, and it matches the documented week expiring on or about August 19.
  • The pool interaction is documented. Cursor's Grok 4.6 help page says "Grok 4.6 uses the same first-party / Composer credit treatment as Grok 4.5. Credits that apply to Grok 4.5 also apply to Grok 4.6. Included plan usage from the first-party pool covers 4.6 the same way," and the Cursor Models pool is now listed as Cursor Grok 4.6, Grok 4.5, and Composer 2.5. Grok 4.6 draws on the existing doubled pool rather than a new allocation of its own.
  • Still unpublished: a named fast-variant model ID and an independent benchmark suite. Cursor's own pricing table does list a "Grok 4.6 (Fast)" row at $4 input / $12 output per million tokens.

Corrections and Updates

October 9, 2026: This page presented the first-week 2x usage offer as live; it ended in August, and SpaceXAI released Grok 4.7 on September 21, 2026 at the same price as Grok 4.6.

Sources

Read next

Related coverage

Frequently Asked Questions

What is Grok 4.6?

Grok 4.6 is SpaceXAI's August 12, 2026 frontier model for coding, agentic tasks, and knowledge work. The company says it builds on Grok 4.5 with a focus on long-running agents and more ambitious interactive and visual work.

Where can I use Grok 4.6?

SpaceXAI and Cursor launched it on August 12, 2026 in Cursor and Grok Build, with the API at x.ai plus OpenRouter, Vercel, and Cloudflare. Cursor's models documentation still lists Grok 4.6, and SpaceXAI's API models page still lists grok-4.6 alongside the newer grok-4.7.

How much does Grok 4.6 cost?

API pricing starts at $2 per million input tokens and $6 per million output tokens. Cached input is $0.50 per million below 200,000 prompt tokens. Prompts at or above 200,000 tokens are billed at $4 / $1 / $12 for input, cached input, and output.

Is there a Grok 4.6 launch promo?

Not anymore. SpaceXAI and Cursor offered 2x included usage inside Cursor and Grok Build for the first week. Cursor's documentation called it a 50% launch discount for one week starting August 12, 2026, and that line was gone from the docs by August 21, 2026, so the offer has ended.

Is there a newer Grok model than Grok 4.6?

Yes. SpaceXAI released Grok 4.7 on September 21, 2026 and calls it its most capable model for coding and knowledge work. It is served at the same price and speed as Grok 4.6, starting at $2 per million input tokens and $6 per million output tokens, in Cursor, Grok Build, and the API.

Does Grok 4.6 beat GPT-5.6 Sol?

Not across the board. SpaceXAI's launch post says Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index. The same table labels that column GPT-5.6 Sol Max at 61 vs 61. Fable 5 Max still leads the index, and Sol Max leads DeepSWE v1.1 and Terminal-Bench v3.0.

Get an email when Cursor changes its plans, prices or limits

We check every change against Cursor's own pages before we send. You also get the weekly AI Catchup; one click unsubscribes.