Grok 4.7 Lands in Cursor at the Same $2 Rate

SpaceXAI released Grok 4.7 on September 21, 2026, at the same $2 and $6 token rates as Grok 4.6. The lab calls it its strongest model for coding and knowledge work, twice as fast at half the price of comparable systems, and already live in Cursor and Grok Build.

That is a 40-day follow-up to Grok 4.6, not a cut in what SpaceXAI charges for Grok itself. The cheaper bill shows up when a team compares Grok 4.7 with GPT-5.6 Sol or Fable 5.1, then runs the job in an editor SpaceX already owns.

Grok 4.7 Holds the $2 Rate Grok 4.6 Already Set

The launch post is blunt about the invoice. Standard Grok 4.7 is billed at $2 per million input tokens and $6 per million output tokens, the same list price Grok 4.6 has carried since August. SpaceXAI says the new model is served at that same price and the same speed as 4.6, and still sits in a cheaper class than the models it put on the same chart.

On that chart, GPT-5.6 Sol is $4 input and $20 output. Fable 5.1 is $10 input and $50 output. Grok’s $2 input rate is half of Sol and one-fifth of Fable. Its $6 output rate is 30% of Sol’s $20 and 12% of Fable’s $50.

TOKEN PRICE PER MILLION

OpenAI’s GPT-6 Astra, which the company priced on September 3 at $10 per million input tokens and $50 per million output tokens, matches Fable’s list rates. Grok 4.7 is not on that Astra post’s scorecard. It is on SpaceXAI’s, and the pitch there is price-performance on long coding jobs, not a clean sweep of every test.

The model id is grok-4.7. Docs list a 500,000-token context window, text and image in, text out, and tools for function calling, web search, X search, and code execution. Those specs matter once a run lasts hours instead of a single chat turn.

Day One Inside Cursor and Grok Build

Grok 4.7 did not wait for a slow enterprise rollout. SpaceXAI said it is available in Cursor and in Grok Build, the company’s own coding agent, plus the Grok API, third-party coding harnesses, and model routers and cloud platforms.

Cursor is not a random launch partner. SpaceX said in June it would buy Cursor’s parent, Anysphere, for $60 billion in stock. That deal closed on August 14 and folded Cursor into the SpaceXAI unit. Two days before the close, Cursor and SpaceXAI had already shipped Grok 4.6 on August 12 as a joint coding model, with double included usage for the first week.

Grok 4.7 now sits on every Cursor plan, and it is the default model in Grok Build. Gateways named in the docs are OpenRouter, Vercel, and Cloudflare. Vercel’s developer account listed Grok 4.7 on Vercel AI Gateway, fx, and Eve on September 21, with 40% off through September 27.

The first replies under SpaceXAI’s launch thread were not about CursorBench. People asked when Grok 4.7 would reach Grok Bot, the public Grok site, and a usage reset. The coding tools SpaceX already controls got the weights on day one. The chat surfaces did not.

Grok 4.7 is here.

It's a notable improvement over Grok 4.6 at the same price and speed. pic.twitter.com/H3OTBbXyvO

— SpaceX Super Intelligence (@SpaceXSI) September 21, 2026

That split is the product. A team that already pays for Cursor or Grok Build can point the new model at a repo the same afternoon. A user waiting on the main Grok app is still on the old stack until SpaceXAI flips those clients.

Terminal Work Nearly Doubles While Fable Still Leads

SpaceXAI’s own table is the honest version of the coding story. Grok 4.7 moves the needle hardest on work that runs for hours. It does not take every column.

On CursorBench 4.0, which the company says stresses longer-running coding tasks, Grok 4.7 scores 46.3% against 40.4% for Grok 4.6, 41.7% for GPT-5.6 Sol, and 51.8% for Fable 5.1. That is a 5.9-point gain over 4.6, and it still sits 5.5 points behind Fable.

On Terminal-Bench 4.0, which covers multi-hour terminal work, Grok 4.7 scores 38.0% against 20.3% for Grok 4.6, 37.3% for Sol, and 57.9% for Fable 5.1. Cursor’s own launch note said the jump from 20.3% to 38.0% nearly doubles the prior model. Double 20.3 is 40.6, so the new score is close, not over. Fable remains well ahead.

SPACEXAI CODING AND WORK SCORES

The DeepSWE v1.1 mark of 71.0% is a high-effort score. Sol still leads that test at 72.7%, with Fable at 70.0% and Grok 4.6 at 65.2%. Electrical engineering is the clean Grok win on this chart: 64.0% on EEBench, ahead of Fable at 56.4%, Grok 4.6 at 53.0%, and Sol at 39.4%.

Office and professional scores are mixed in the same way. AA Briefcase v1.1, a multi-hour office test, has Grok 4.7 at 1,657, Grok 4.6 at 1,546, Sol at 1,487, and Fable 5.1 at 1,678. On the Harvey Legal Agent Benchmark, Grok 4.7 scores 19.6% against 15.8% for 4.6, 2.5% for Sol, and 6.7% for Fable. HealthBench Professional goes the other way: 56.7% for Grok 4.7, behind Sol at 60.5% and Fable at 62.1%.

Early Cursor users described the change as less babysitting of the harness, especially in Fast mode. That is a workflow claim, not a new leaderboard. Fable still owns the two longest coding tests SpaceXAI chose to publish. What 4.7 changes is the cost of sitting closer to those scores for hours at a time.

Why SpaceXAI Weighted Training Toward Multi-Hour Jobs

The company says Grok 4.7 uses a new, larger base model than Grok 4.6. It then ran a longer reinforcement learning pass on a harder mix of tasks, weighted toward problems that take many hours to finish. It also trained the model to natively understand the Grok Bot harness, which is the shared-computer agent stack, so the same weights are meant to handle chat and general knowledge work, not only a repo.

Grok 4.7 is our most capable model for coding and knowledge work. It works longer on difficult tasks, checks its own work more carefully, and comes with our best-calibrated safeguards to date.

SpaceXAI, Introducing Grok 4.7

SpaceXAI says the model is better at verifying its own work and managing longer context, and stronger at creating documents and presentations. In GDPval and AA Briefcase, the tasks are the kind lawyers, nurses, and financial analysts already do. Grok 4.7 beats 4.6 on both and, the lab says, performs in the same band as other frontier models.

The safeguard stack is new as well. SpaceXAI calls 4.7 the strongest Grok it has tested on refusals and jailbreak resistance. On LatchBio’s biosafety benchmark it reports 62.4%. On HackerBench v0.3, a test of risky and malicious cyber tasks, it says the model lets only 3.3% of risky dual-use prompts through while rarely blocking legitimate security work. Select cybersecurity partners get invite-only access to red-team features for defense research.

Those safety numbers sit beside a coding model that is trained to keep going. The same release that boasts a 3.3% leak rate on malicious cyber prompts is also the one that almost doubled Terminal-Bench by staying on a job longer. Teams that want the persistence will also inherit the new refusal stack, for better or worse, on the same API call.

Fast Mode Costs Double and Stays Off the Public API

SpaceXAI also serves a Fast variant of Grok 4.7 with twice the output speed at twice the price, so $4 input and $12 output per million tokens. Docs are explicit about the walls around that SKU. It is the same model on faster machines. It runs only in Cursor and Grok Build. It is not in Grok Build’s free tier. It is not on the public xAI API.

WHERE GROK 4.7 RUNS

• Cursor: Available on all plans, including the Fast variant.
• Grok Build: Default model of the coding agent; Fast is paid-only.
• Grok API: Standard grok-4.7 on the public xAI API and a US regional endpoint.
• Gateways: OpenRouter, Vercel, and Cloudflare, plus other harnesses and cloud routers.

The US regional endpoint at us.api.x.ai/v1 keeps inference in the United States and bills token use at a 10% premium. Docs also push teams to set a prompt cache key so multi-turn jobs actually hit cache, and they point long agent loops at context compaction. Encrypted reasoning is always returned on the Responses API for grok-4.7, even when a client does not ask for it, so a multi-hour run can pass reasoning items back unchanged.

Reasoning effort is a lever, not a single mode. The named settings are low, medium, high, and xhigh, with high as the default. SpaceXAI’s published DeepSWE number uses high effort. CursorBench and Terminal-Bench in the launch table are shown under Grok 4.7 xHigh against Grok 4.6 High, GPT-5.6 Sol Max, and Fable 5.1 Max. A user who leaves the default on high is not running the exact setup on that chart.

The Release Calendar Did Not Pause

Grok 4.7 arrives in a month when the large labs talked about slowing down and then kept shipping. Anthropic chief executive Dario Amodei published “We Must Pace the Frontier” on September 12, 2026, and wrote that the industry must slow the pace at which it improves the capabilities of AI models. He said pacing is not a halt, and that progress will still seem fast. SpaceXAI’s 40-day gap from Grok 4.6 to Grok 4.7 fits that last clause better than the first.

THE SIX WEEKS AROUND GROK 4.7

1. August 12, 2026: Cursor and SpaceXAI release Grok 4.6 at $2 and $6 per million tokens.
2. August 14, 2026: SpaceX closes the $60 billion all-stock Cursor deal and folds the editor into SpaceXAI.
3. September 3, 2026: OpenAI launches GPT-6 Astra at $10 and $50 per million tokens.
4. September 12, 2026: Amodei publishes the pacing essay and commits Anthropic to embedded third-party evaluators.
5. September 21, 2026: SpaceXAI ships Grok 4.7 at the same $2 and $6 rates, first in Cursor and Grok Build.

OpenAI’s Sam Altman said the same week as Amodei’s essay that he agreed the frontier should be paced and that OpenAI would adopt independent evaluators. GPT-6 Astra had already been in paid ChatGPT tiers and the API for nine days by then. SpaceXAI’s launch post nine days after that, on September 21, called Grok 4.7 a notable improvement over Grok 4.6 at the same price and speed.

xAI was absorbed into SpaceX in February 2026 and later took the SpaceXAI name. Grok is no longer a private-lab side bet. It is a public-company coding product with an editor attached. The second-order effect of Grok 4.7 is that a 40-day model bump, priced like the last one, now lands inside that editor before it lands in the chatbot a lot of people still think of as Grok.

Frequently Asked Questions

Is the Grok 4.7 Fast Variant on the Public API?

No. Grok 4.7 Fast is the same model served on faster machines, billed at twice the standard token rates, and SpaceXAI’s docs say it is available only in Cursor and Grok Build. It is excluded from Grok Build’s free tier and is not offered on the public xAI API, so an API customer calling grok-4.7 gets the standard speed SKU.

What Is Grok 4.7’s Knowledge Cutoff?

SpaceXAI’s developer docs list a May 2026 knowledge cutoff for grok-4.7, alongside text and image input and text output with no text output limit. Anything after that cutoff has to come from tools such as web search, X search, or whatever files a harness attaches at run time.

How Is the US Regional Endpoint Billed?

The US regional endpoint at us.api.x.ai/v1 keeps inference inside the United States and prices token usage at a 10% premium to the standard $2 and $6 rates. Teams that need that residency pay the premium on both input and output; everyone else can use the main xAI API without it.

What Reasoning Effort Settings Does Grok 4.7 Accept?

The docs name four reasoning settings: low, medium, high, and xhigh, with high as the default. SpaceXAI’s launch table compares Grok 4.7 xHigh with Grok 4.6 High, and the 71.0% DeepSWE v1.1 score is marked as high effort, so a default high call is not the same run as the xHigh column on CursorBench 4.0.