Grok 4.7: what changed, what it costs, and how it compares to Claude and ChatGPT

The short answer: if you already call Grok 4.6 over the API, change the model id to grok-4.7. The rate card did not move. But budget for bigger bills per task, because Grok 4.7 writes far more reasoning tokens than 4.6 did. If your team works in ChatGPT or Claude every day, stay there. On the independent index we trust most, Grok 4.7 is still well behind Claude Fable 5.1 and GPT-6 Astra.

Facts below were checked against xAI's Grok 4.7 announcement, xAI's release notes and Artificial Analysis's Grok 4.7 benchmark write-up as of 22 September 2026. Scores that xAI published are labelled as xAI's own. We do not publish a LiveBench score for Grok 4.7 because LiveBench has not published one.

What launched

On 21 September 2026, xAI released Grok 4.7 as the successor to Grok 4.6, which came out on 12 August. According to xAI's release notes, this is what you get:

• Model id: grok-4.7 on the xAI API.
• Context window: 500K tokens, the same as Grok 4.6. It takes text and images in and returns text only.
• Price for prompts under 200K tokens: $2 input / $0.50 cached input / $6 output per 1 million tokens.
• Price for prompts above 200K tokens: $4 / $1 / $12.
• Reasoning effort: you choose low, medium, high (the default) or xhigh.
• A fast variant, grok-4-7-fast: it doubles the token rates for about twice the output speed. It is only available inside Cursor and Grok Build, not on the public xAI API.

xAI says Grok 4.7 is available in Cursor, Grok Build, the Grok API, third-party coding tools, model routers and cloud platforms. It is listed on OpenRouter as x-ai/grok-4.7. xAI's release notes do not mention retiring Grok 4.6, so it is still served.

xAI's own description of the upgrade is that Grok 4.7 "works longer on difficult tasks" and "checks its own work more carefully." Treat that as the vendor's pitch. The numbers are more useful.

How it scores: xAI's numbers

xAI published a comparison table against its own previous model, OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5.1. All of these figures are xAI's own.

Two things stand out. First, the jump from 4.6 is real on coding in a terminal: Terminal-Bench 4.0 nearly doubles. Second, even on xAI's own chart, Claude Fable 5.1 wins three of these four rows, and it wins Terminal-Bench by about 20 points. Anthropic's own figure for Fable 5.1 on Terminal-Bench 4.0 is 55.8%, so the gap is large whichever number you use.

How it scores: the independent check

Artificial Analysis tested Grok 4.7 at xhigh effort and published the results on launch day:

• Intelligence Index: 46, up 2 points from Grok 4.6's 44. The leaders on the same scale, Claude Fable 5.1 and GPT-6 Astra, sit at 53.
• Coding Agent Index: 56 when run inside Grok Build, xAI's own coding agent. That is up 9 points from Grok 4.6 and ranks 4th, behind Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.
• Agentic knowledge work (AA-Briefcase): 1,657 Elo, up 111 from Grok 4.6, placing it just behind Claude Opus 5 and Claude Fable 5.1.
• Hallucination rate: 29%, down from 34% for Grok 4.6. Accuracy was broadly flat at 47% against 48%.
• Speed: about 188 output tokens per second on long prompts, and about 7.1 minutes per index task.

One caution on comparing scores. Artificial Analysis has re-based its Intelligence Index since Grok 4.6 launched. In August, xAI quoted Grok 4.6 at 61 on the older scale. On today's scale, Grok 4.6 is 44. Do not compare the 61 from our Grok 4.6 post with the 46 here. Compare numbers only within the same article.

The price did not change. The bill might.

This is the part most launch coverage skips. Artificial Analysis measured Grok 4.7 using about 81,000 output tokens per Intelligence Index task. That is more than double Grok 4.6 at the same xhigh setting (38,000), and three times GPT-6 Astra (27,000).

You pay for output tokens. So the same $6 per 1M output rate buys fewer finished tasks than it did on Grok 4.6. If you move a workflow from 4.6 to 4.7 and leave the effort level at xhigh, expect your output-token spend on that workflow to rise, even though the rate card is identical.

For context, here is a rough comparison using only the sourced numbers above. This is our arithmetic, not a vendor figure, and it counts output tokens only. At 81,000 tokens and $6 per 1M, one Grok 4.7 index task costs about $0.49 in output. At 27,000 tokens and $50 per 1M, one GPT-6 Astra task costs about $1.35. Grok 4.7 is still cheaper per task, but by a much smaller margin than the $6 against $50 rate cards suggest.

The practical fix is simple. Start new Grok 4.7 workflows at medium or high effort (the default) and only move to xhigh on the tasks where you can see it helps.

Grok 4.7 vs Claude and ChatGPT for a small business

Everyday chat (drafts, email, "help me think"). Keep ChatGPT or Claude if that is where your team already works. What you pay for is the app around the model: memory, files, projects, and a UI people actually open. This launch was an API and coding-tool release, and we could not verify any change to the consumer Grok app, so we do not treat it as a reason to move chat.

Grok 4.7 vs Claude. Claude Fable 5.1 costs $10 input / $50 output per 1M tokens on Anthropic's API. That is five times Grok 4.7 on input and a little over eight times on output. Fable 5.1 also scores higher on every independent number above and wins most of xAI's own table. If quality on hard coding or long documents is what you are paying for, Fable 5.1 is still the stronger model. If you want near-frontier work on a tighter budget and can accept a step down, Grok 4.7 is a credible middle option.

Grok 4.7 vs ChatGPT (GPT-6 Astra). GPT-6 Astra has the same $10 / $50 rate card as Fable 5.1 and ties it at the top of the Intelligence Index. OpenAI built it for long chains of computer use: clicking through forms, driving a browser. Grok 4.7 is not sold on that. Pick Astra if the job is operating software. Pick Grok 4.7 if the job is reasoning or writing on a budget.

API work already running on Grok 4.6. Migrate, but watch the bill. It is a one-line model swap at the same price, and it does better on coding and knowledge work. Set the effort level on purpose instead of leaving it at xhigh.

Volume automation. Grok 4.7 is not the cheap tier. DeepSeek V4.1 Flash costs $0.15 / $0.60 per 1M tokens off-peak. Use Flash for bulk sorting, tagging and extraction, and save Grok 4.7 for the slice of work where Flash's mistakes would cost you more than the extra tokens.

What we will (and will not) put on the leaderboard

Our AI model tracker lists Grok 4.7 with its price, the xAI-reported Terminal-Bench figure and the independent Artificial Analysis index, and each number is linked to its source. Our benchmarks page shows LiveBench's published leaderboard unaltered. LiveBench now lists "Grok 4.7 xHigh" at 77.4 overall, and that row appears on our page exactly as LiveBench publishes it. We do not add benchmark rows of our own.

We have not published a parameter count, training-data claims, token speed beyond Artificial Analysis's measurement, or consumer-app availability for Grok 4.7. We saw those claims circulating but could not trace them to xAI.

The move this week

1. If Grok 4.6 is in an API workflow, switch to grok-4.7 and set reasoning effort explicitly. Check your output-token spend after a week.
2. If your team lives in ChatGPT or Claude, do not migrate chat for this release.
3. If you need a lot of cheap tokens, stay on DeepSeek V4.1 Flash. Use Grok 4.7, Fable 5.1 or Astra only on the hard slice.

For the rest of September's launches, see our September 2026 model roundup and the running AI model release watch.

Splitting work between everyday chat, a volume API and a frontier API is exactly the call a Business AI Plan is built to make. Start with the free 2-minute AI plan finder if you want that mix named for your own workflow instead of a leaderboard.

Frequently asked questions

When was Grok 4.7 released? xAI released Grok 4.7 on 21 September 2026. It is on the xAI API as grok-4.7, and it is also available in Cursor, Grok Build and OpenRouter.

How much does Grok 4.7 cost? For prompts under 200K tokens, $2 per 1M input tokens, $0.50 per 1M cached input tokens and $6 per 1M output tokens. Above 200K tokens, the rates are $4, $1 and $12. That is the same as Grok 4.6. The faster grok-4-7-fast variant costs double and is only offered inside Cursor and Grok Build.

Is Grok 4.7 better than Claude? Not on the independent numbers. Artificial Analysis scores Grok 4.7 at 46 on its Intelligence Index against 53 for Claude Fable 5.1, and xAI's own table shows Fable 5.1 ahead on Terminal-Bench 4.0, CursorBench 4.0 and GDPval. Grok 4.7 is much cheaper per token, so it can be the better value for work that does not need the top model.

Is Grok 4.7 better than ChatGPT? GPT-6 Astra, OpenAI's current flagship, ties Claude Fable 5.1 at 53 on Artificial Analysis's index, ahead of Grok 4.7's 46. Astra costs $10 / $50 per 1M tokens, against Grok 4.7's $2 / $6. However, Grok 4.7 uses about three times as many output tokens per task, so the real cost gap is smaller than the rate cards suggest.

Should I switch from Grok 4.6 to Grok 4.7? If you use Grok 4.6 through the API, yes. The price is the same and the coding and knowledge-work scores are higher. Set the reasoning effort deliberately, because at xhigh Grok 4.7 writes more than twice as many output tokens as 4.6 did.

Is Grok 4.6 being retired? xAI has not announced a retirement date for Grok 4.6. Its release notes list 4.7 as the new frontier model but do not say 4.6 is going away.

Sources: xAI — Grok 4.7, xAI release notes, Artificial Analysis — Benchmarking Grok 4.7, OpenRouter — Grok 4.7, Anthropic — Claude Fable 5.1, OpenAI — GPT-6 Astra model docs, DeepSeek pricing