Grok / News

Grok 4.7, Grok Bot and Team Bots: what's new, what it costs, and how it did in our test

Grok 4.7 costs the same as Grok 4.6. We tested both on the same coding jobs. We also read xAI's pages about Grok Bot, Team Bots and where your data goes.

Four facts about Grok 4.7: $2 input and $6 output per million tokens, a 500,000 token context window, our test cost against Grok 4.6, and Grok Bot running in the US
Chart by Not an AI App from xAI's documentation and our own test of 5 October 2026.

xAI released Grok 4.7 on 21 September 2026. xAI now calls itself SpaceXAI. It says Grok 4.7 is its "most powerful model for coding and knowledge work". A week later, it added Team Bots: shared AI helpers for companies.

We read xAI's documentation. We also tested Grok 4.7 against the model before it, Grok 4.6. Both cost exactly the same per token. Tokens are the small pieces of text an AI reads and writes. Both models passed the same secret tests: tests we wrote ourselves and only ran after the AI had finished. But Grok 4.7 wrote about twice as many tokens to get there. So at the same price per token, it cost us about twice as much.

What is Grok 4.7?

Screenshot of xAI's release notes for 21 September: Grok 4.7 is available on the xAI API as grok-4.7, with a 500k context window and prices of $2, $0.50 and $6 per million tokens
xAI's release notes entry for Grok 4.7. Screenshot taken on 5 October 2026, cropped.

Grok 4.7 is xAI's newest AI model. xAI's release notes say it has been on the xAI API since 21 September 2026. The API is the paid connection for developers. There, the model is called grok-4.7.

The announcement calls it "our most capable model for coding and knowledge work". xAI says it "works longer on difficult tasks" and "checks its own work more carefully". It also says it uses "a new, larger base model compared to Grok 4.6".

xAI's model page lists these facts:

  • Context window: 500,000 tokens. That is how much text the model can look at at once.
  • Input and output: it reads text and images, and it writes text.
  • Knowledge cutoff: May 2026. Its knowledge stops there.
  • Thinking levels: four of them: low, medium, high (the default) and xhigh.
  • Output: "no text output limit", xAI says.

You can use it through the xAI API, in Cursor, and in xAI's own coding tool Grok Build. You can also use it through a gateway such as OpenRouter. A gateway is a service that gives access to many models through one account. That is how we tested it.

What does it cost?

Screenshot of xAI's Grok 4.7 page: 500,000 token context window, knowledge cutoff May 2026, no text output limit, $2.00 input and $6.00 output per million tokens
Grok 4.7 at a glance in xAI's documentation. Screenshot taken on 5 October 2026, cropped.

xAI's price list has two price levels. Which one you pay depends on how long your prompt is:

Grok 4.7, per million tokensInputSaved input (cached)Output
Prompts under 200,000 tokens$2.00$0.50$6.00
Prompts of 200,000 tokens or more$4.00$1.00$12.00

Saved input is text the model has already read once, such as long instructions. Reading it again is cheaper.

This is exactly the price of Grok 4.6. xAI's own announcement says so too: Grok 4.7 is "served at the same price and speed as Grok 4.6".

There is also a separate US regional endpoint. An endpoint is the web address an app sends its requests to. This one is for requests that must be processed in the United States. It costs 10% more.

How does that compare with other new models? GPT-6.1 Sol and Claude Sonnet 5.5 cost $2 for input and $10 for output. We wrote about that in our Gemini 4 Argon article. So Grok 4.7 is cheap on output. Per token, it is the cheapest of the group. But as our test shows, the number of tokens a model needs matters just as much.

xAI's test scores: strong in some places, not everywhere

Bar chart of xAI's own test scores: Grok 4.7 leads on EEBench and Harvey's legal test but is behind Fable 5.1 on CursorBench, Terminal-Bench and HealthBench
Chart by Not an AI App of the scores xAI published on 21 September 2026. These are xAI's claims, not our measurements.

xAI's announcement compares Grok 4.7 with three other models on seven tests. The others are Grok 4.6, OpenAI's GPT-5.6 Sol and Anthropic's Fable 5.1. These are xAI's own figures. We could not repeat the tests ourselves. They are worth reading closely, because Grok 4.7 does not win them all:

  • Ahead: on EEBench, a test of electrical engineering, Grok 4.7 scores 64.0%. Fable 5.1 scores 56.4%. On a legal test by the company Harvey, it is 19.6% against 6.7%.
  • Behind: CursorBench 4.0 is the main coding test xAI shows. There Grok 4.7 scores 46.3% and Fable 5.1 scores 51.8%. On Terminal-Bench 4.0 the gap is bigger: 37.6% against 57.9%. On HealthBench Professional, Grok 4.7 is behind both rivals.

One more thing does not add up. The subtitle of the announcement says "Twice as fast, at half the price of comparable models". But the text itself says Grok 4.7 has the same speed as Grok 4.6. The speed claim seems to be about the separate Fast version. We explain that one below.

Our test: Grok 4.7 against Grok 4.6

Table of our test: Grok 4.7 and Grok 4.6 on two coding jobs with secret test scores, average time and average cost per try
Our own test, 5 October 2026, through OpenRouter with default settings.

Both models cost the same per token. So the question is simple: does the newer model do better work, and how many tokens does it use for it?

We reused two jobs from our Codex test of GPT-6.1 Sol and GPT-6 Astra:

  • Fix a bug in a piece of code that reads times such as "1h30m".
  • Build a Dutch invoice calculation from a one-page description.

Each model did each job three times. We sent each job as one single request through OpenRouter, with the default settings. The model had to send back the complete file. Then we checked every answer with the same secret tests as before.

Grok 4.7Grok 4.6
Bug fix: secret tests passed23/24, 23/24, 23/2423/24, 23/24, 23/24
Invoice: secret tests passed15/15, 15/15, 15/1515/15, 15/15, 15/15
Tokens written, six tries39,96619,376
Cost, six tries$0.246$0.119
Time, six tries336 s250 s

The scores were the same. Every try of both models passed the same secret tests.

The difference was in how much each model wrote. Grok 4.7 wrote about twice as many tokens as Grok 4.6. Almost all of them were reasoning tokens. That is thinking the AI does before it answers. You pay for it, but you do not see it. Over six tries, Grok 4.7 used 35,858 reasoning tokens. Grok 4.6 used 16,567.

So at the same price per token, Grok 4.7 cost about twice as much for the same result. It also took about a third longer. And it was less steady. Per try, it wrote between about 3,400 and 9,400 tokens. Grok 4.6 wrote between about 2,900 and 3,900.

To be fair to Grok 4.7: xAI says it "works longer on difficult tasks". Our jobs were not hard enough to need that. So the extra thinking did not help here. On harder work, it might. And the amounts are small: all six Grok 4.7 tries together cost less than 25 cents.

One bug-fix test failed for both models, every time. Just like in the Codex test, that test checks a rule our job note never spelled out. So we do not count it against either model.

Two jobs and six tries per model is a small test. It shows how the two models behave on clear, everyday coding jobs. It does not show how they do on long projects.

Want to see the test? One secret test checks that "1h banana" is refused, because there is junk after the time. Another checks that an invoice line of 1.005 kilos is allowed. Our test page shows all three jobs, the secret tests and the exact words we sent. You can also download them and run them on any AI yourself.

Grok 4.7 Fast: only in Cursor and Grok Build

Screenshot of xAI's documentation: Grok 4.7 Fast is the same model on faster infrastructure at twice the price, only in Cursor and Grok Build, not on the public xAI API
The fast version in xAI's Grok 4.7 documentation. Screenshot taken on 5 October 2026, cropped.

xAI also offers Grok 4.7 Fast. Its documentation calls it "the same model served on faster infrastructure". In other words: the same model, on faster computers. It costs twice the normal price per token. For long prompts, it costs 1.5 times as much.

There is a catch. Fast is "available only in Cursor and Grok Build". It is not part of Grok Build's free plan. And it "is not available on the public xAI API". So if you use the API directly, or a gateway such as OpenRouter, you get the normal speed.

What is Grok Bot?

Screenshot of xAI's Grok Bot overview: Bots you can keep around, AI teammates with names and jobs that work on a persistent cloud computer with a browser, filesystem and terminal
The opening of xAI's Grok Bot documentation. Screenshot taken on 5 October 2026, cropped.

Grok Bot is xAI's helper app. It launched on 11 August 2026. The documentation describes "Bots you can keep around: AI teammates with names, jobs, and context that compounds over time". Put simply: Bots that build up knowledge of your work over time.

Each Bot works on its own computer in the cloud, which keeps running. It has a web browser, a place to store files and a terminal for typing commands. So it can carry out tasks, not just write about them. Grok Bot is a different product from the Grok chat app on grok.com.

You cannot buy Grok Bot on its own. xAI says it "is included with every paid individual Cursor plan and with the Cursor Teams plan". You can also link a personal SuperGrok, SuperGrok Plus or SuperGrok Heavy subscription. Cursor's help page adds X Premium+. SuperGrok Lite does not include it. How much you can use it resets every week.

Update, 13 October 2026: since 12 October, new customers get Grok Bot through The X Subscription, from the Premium plan at $30 a month. Existing SuperGrok and Cursor subscribers keep their current plans. A Grok Bot update of 11 October also lets Bots get their own email address and pay with X Money or Stripe Link, after your approval. See our explainer.

Two details stand out in xAI's own security documentation. Grok Bot runs on Cursor's computers. And "Cursor manages model selection": you cannot pick the model, and xAI does not promise which model does the work. We did not test Grok Bot, because it needs one of those paid plans.

Update, 17 October: we have now tested Grok Bot on a family's admin, against Claude Cowork. Both scored 24 out of 24; see Grok Bot vs Claude Cowork.

Team Bots: one Bot for a whole team

Screenshot of xAI's Team Bots documentation: a Team Bot is one Bot that your whole team talks to, in the Grok Bot app or in Slack
xAI's description of Team Bots. Screenshot taken on 5 October 2026, cropped.

xAI announced Team Bots on 28 September 2026. According to xAI, they are in public beta on Teams and Enterprise plans. That means people on those plans can use them, but they are still being tested. The documentation says: "A Team Bot is one Bot that your whole team talks to."

The owner sets it up once. They give it the plugins, secrets, skills and files it needs. Then every teammate can chat with it in the Grok Bot app or in Slack.

xAI says each person's chats with a Team Bot stay private. Use counts against each teammate's own allowance, not the owner's. A Team Bot can hold up to 25 secrets, such as passwords or API keys. Audit logs, a record of who did what, are only available on Enterprise plans.

So a shared helper holds a team's secrets on a computer in the cloud. Whoever looks after security in a company should check that carefully.

For readers in Europe: where does your data go?

Screenshot of xAI's Grok Bot security page: Grok Bot computers run in the United States today, and Cursor's US-only data residency programme does not apply by default
Where data is processed, on xAI's Grok Bot security page. Screenshot taken on 5 October 2026, cropped.

As European users, this is the part we would check first. It is about data residency: where your data is stored and processed.

  • Grok Bot: xAI's security page says "Grok Bot computers run in the United States today". Cursor has a programme that keeps data in the US only. xAI says that programme does not apply to Grok Bot by default.
  • The API: the regional endpoints page says xAI "may route requests between regions for capacity or reliability, so the processing location is not guaranteed". In other words: you cannot be sure where your request is processed. There is a US endpoint. We found no EU endpoint.
  • Training: for the API, xAI's security FAQ says it "never trains on your API inputs or outputs without your explicit permission". It keeps requests for 30 days, for checks (auditing). Teams can switch on zero data retention, which means nothing is kept.

xAI's release notes say Grok 4.5 became available in the EU in July 2026. We found no such note for Grok 4.6 or 4.7. And the plans page of the normal Grok app still mentioned Grok 4.6 when we looked on 5 October.

Update, 8 October: Grok Bot will also use Claude, Midjourney and Suno

Screenshot of Elon Musk's post on X about Grok Bot, shown in X's automatic Dutch translation: SpaceX will use the best back-end model for each task, including Claude Opus 5.5, Midjourney and Suno
Elon Musk's post on X, seen in the site owner's X app with X's automatic Dutch translation. Screenshot from 8 October 2026, cropped.

On 7 October 2026, Elon Musk wrote on X that Grok Bot will from now on use "the best back end model for any given task". He named Claude Opus 5.5, Midjourney and Suno, and "other leading APIs". Claude Opus 5.5 is made by Anthropic. Midjourney makes images and Suno makes music.

So when you ask Grok Bot something, the answer may not come from Grok at all, but from another company's AI. Musk did not say which tasks go to which model, or when this starts.

That also matters for your data. Your question may then be sent to Anthropic, Midjourney or Suno, and their rules may apply. We found no official page from xAI or SpaceX yet that explains this. We will update this article when there is one.

Questions people ask

List of questions people ask Google about this topic, with the 4 questions we answer marked
"People also ask" questions from Google (US), 9 October 2026, via DataForSEO. Chart by Not an AI App.

What is the current version of Grok?

Grok 4.7 is xAI's newest model. It has been on the xAI API, the paid connection for developers, since 21 September 2026. You can also use it in Cursor, in xAI's coding tool Grok Build and through services like OpenRouter. On 5 October, the plans page of the normal Grok app still mentioned Grok 4.6.

Why is Grok 4.6 so good?

We cannot say why, but in our test it did as well as the newer Grok 4.7. Both passed the same secret tests, tests we wrote ourselves, on two coding jobs. Grok 4.6 wrote about half as many tokens, the small pieces of text an AI writes. So at the same price per token, it cost $0.12 against $0.25.

Is Grok 4 available now?

Grok 4.7, the newest version, is. Developers can use it since 21 September 2026 through the xAI API, Cursor and Grok Build. For Europe, xAI's release notes only say Grok 4.5 became available in the EU in July 2026. We found no such note for Grok 4.6 or 4.7.

How much does Grok 4 cost?

Grok 4.7 costs $2 per million tokens in and $6 per million tokens out, for prompts under 200,000 tokens. Longer prompts cost twice as much. Grok 4.6 has the same price, xAI's price list shows. Our article does not list prices for older Grok 4 versions.

How we checked

Diagram of our method in three steps: read xAI's documentation, test Grok 4.7 against 4.6 on two coding jobs, check and compare
Diagram by Not an AI App: how we checked.

On 5 October 2026 we read xAI's announcement, release notes, price list and Grok 4.7 page. We read the Grok Bot overview, Team Bots and security pages. We also read the regional endpoints and API security pages, and Cursor's Grok Bot plans page.

The screenshots are our own pictures of xAI's documentation. We only cropped them. xAI's news pages blocked our automatic browser, so we could not take screenshots there. We read the Grok 4.7 announcement as plain text and made a chart of its test scores from that. The Team Bots announcement we could only read through a tool that summarises pages. So for Team Bots, we relied on the documentation wherever we could.

For our test, we sent each job as one request through OpenRouter. We did that three times per model, with the default settings. We checked each returned file with secret tests. Before that, we had checked those tests against our own correct answers. The time, token counts and cost come from our script and from OpenRouter's reply. The owner of this site paid for the requests.

Our first run gave usable scores, but our script measured the time wrongly. So we fixed the script and ran the whole test again. This article uses only the second run.

We did not test Grok Bot, Team Bots or the Fast version. None of the images was made by AI. Prices, plans and availability can change at any time.