Grok 4.7: What Actually Changed

If you searched Grok 4.7, Grok 4.7 pricing, or Grok 4.7 vs Grok 4.6, here is the short version.

On 21 September 2026, xAI (which now also calls itself SpaceXAI) released Grok 4.7, its new model for coding and knowledge work. The pitch is simple:

Its price hasn't changed from Grok 4.6, but there is a bigger brain underneath and it is much better at long, multi-step coding jobs.

This guide explains what actually changed, in plain English.

Quick verdict

What Grok 4.7 is

xAI calls Grok 4.7 its most capable model for coding and knowledge work. Think of it as xAI's new default workhorse for:

1. AI coding agents (Cursor, Grok Build, GitHub Copilot)
2. Office work: documents, presentations and spreadsheets
3. Long, multi-step tasks that take hours rather than seconds

API model id: grok-4.7

It takes text and images as input and gives text as output. It supports function calling, structured outputs (JSON), and a reasoning "effort" dial: low, medium, high (default) and xhigh.

1) Under the hood: a bigger model, trained for longer

This is the real change, and it is not just a tune-up.

According to xAI:

1. Grok 4.7 uses a new, larger base model than Grok 4.6
2. It had a longer reinforcement learning run on harder tasks, weighted toward problems that take many hours to finish
3. It is better at checking its own work and managing long context
4. It got extra training on anonymized Cursor workflow data to improve coding

The model card lists a pretraining data cutoff of June 2026, with some training data from as late as August 2026.

Simple takeaway: Grok 4.6 was good at quick coding answers. Grok 4.7 is built to stay on task for long jobs and verify before it says "done".

2) Price: same rate card, one new rule to know

xAI's published API prices (per 1M tokens):

Three things to know:

1. Same headline price as Grok 4.6, served at the same speed
2. Long prompts cost double. If a request's prompt reaches 200K tokens, every token in that request is billed at the higher rate
3. No Batch API discount is listed for Grok 4.7

There is also a fast variant with about twice the output speed at twice the price. Reports say it is only available inside Cursor and Grok Build, not on the public API.

Simple takeaway: On the rate card, Grok 4.7 is one of the cheapest frontier-class models. For comparison, Claude Opus 5.5 lists at $4 / $20.

3) Coding and agents: the biggest jump

This is where Grok 4.7 improved most. Here are xAI's own numbers (Grok 4.7 at xhigh effort vs Grok 4.6 at high effort):

The Terminal-Bench score almost doubled. That fits xAI's story of a model trained on hours-long tasks.

Independent testing points the same way. Artificial Analysis measured Grok 4.7 (xhigh, inside Grok Build) at 56 on its Coding Agent Index, up from 47 for Grok 4.6.

Honest context: Grok 4.7 is not the top scorer. In xAI's own table, Claude Fable 5.1 still leads on CursorBench 4.0 (51.8%) and Terminal-Bench 4.0 (57.9%). xAI's claim is price-performance: near-frontier results at a much lower per-token price.

For the wider September model race, see our comparison: GPT-6 Astra vs Claude Fable vs Gemini 3.8 Flash.

4) Knowledge work: better documents and slides

xAI says Grok 4.7 is better at creating documents and presentations. On the benchmarks it published:

1. AA Briefcase v1.1 (multi-hour office work): 1,657 vs 1,546 for Grok 4.6
2. Harvey Legal Agent Benchmark: 19.6% vs 15.8%
3. HealthBench Professional (clinical reasoning): 56.7% vs 48.5%

It is also the default model in the Grok add-ins for Microsoft Word, PowerPoint and Excel.

Simple takeaway: It isn't only for coders. If your team writes reports, decks or analysis, it is worth a test. Still, xAI's own model card says it should not make high-stakes medical, legal or financial decisions without human experts checking.

5) The catch: more thinking = more tokens

Here is the part most launch posts skip.

Artificial Analysis found that Grok 4.7 uses far more output tokens per task than Grok 4.6:

1. About 81,000 output tokens per task at xhigh, vs about 38,000 for Grok 4.6 at xhigh
2. Measured cost per task: about $3.74 at xhigh and $2.73 at high

Because output tokens are the expensive part, a low per-token price does not always mean a low bill. In the same testing, Claude Opus 5.5 at its default effort cost about $1.34 per task, even with its higher rate card.

Simple takeaway: Start at high (the default), not xhigh. Only turn it up for the hardest tasks, and set spend caps on long agent runs.

Cost tip: If your agents pass lots of JSON around, a more compact format can cut input tokens. Our free JSON TOON Converter runs in your browser, so you can test this without uploading your data anywhere.

6) Safety: a new safeguard stack

xAI says Grok 4.7 ships with an entirely new safeguard stack:

1. On xAI's HackerBench v0.3 (risky cyber tasks), it let only 3.3% of risky dual-use prompts through, while rarely blocking legitimate security work
2. It topped LatchBio's biosafety benchmark at 62.4%
3. Select cybersecurity partners get invite-only access to its red-team abilities for defense research

One detail worth noticing: the Grok 4.7 model card says xAI never silently downgrades or falls back to other models. Compare that with Claude Opus 5.5, where some cyber and biology requests can be quietly routed to older models (we covered this in Claude Opus 5.5: What Actually Changed).

And one rule that never changes: do not paste API keys, passwords or customer data into any AI agent, whichever model it runs on. See Stop Pasting Secrets Into AI.

7) Where you can use it today

1. Grok API: grok-4.7 (console.x.ai)
2. Cursor: available on every plan
3. Grok Build: xAI's terminal coding agent, where Grok 4.7 is the default model
4. GitHub Copilot: rolling out gradually to Pro, Pro+, Max, Business and Enterprise (VS Code, Visual Studio, JetBrains, Xcode, Eclipse, Copilot CLI and more). Business and Enterprise admins control access through model policy
5. Model gateways: OpenRouter, Vercel, Cloudflare, Snowflake, Databricks Mosaic and others
6. Office add-ins: Word, PowerPoint, Excel

Not yet: xAI says Grok 4.7 will come to its consumer apps (web, mobile, and Grok on X) at a later date.

Note for Indian teams: xAI's docs list API regions in the US only (us-east-1, us-west-2, us-central-1). If you have data-residency rules, check them before sending client data.

Grok 4.7 vs Claude Opus 5.5 at a glance

The two launched a day apart (21 and 22 September), so many teams are comparing them.

In short: Grok 4.7 wins on price per token and response speed. Opus 5.5 wins on benchmark scores and context size, and can end up cheaper per finished task.

Should you switch this week?

Switch or test now if:

1. You already use Grok 4.6: it's the same price, with a clearly better model
2. You work in Cursor or GitHub Copilot and want a cheap, fast option for agent coding
3. Your product needs answers to start fast (chat-style assistants)

Wait / test carefully if:

1. You need the very top coding scores (Fable 5.1 and Opus 5.5 still lead)
2. Your prompts regularly go above 200K tokens (double pricing kicks in)
3. You need data to stay outside the US, or you need Batch API discounts

30-minute bake-off

1. The same multi-file bug on Grok 4.7 (high) vs your current default
2. The same long refactor with a spend cap, comparing the real cost per task, not the rate card
3. The same report or slide draft

Keep the model that needs the least babysitting for the lowest total bill.

FAQ (search questions)

When was Grok 4.7 released?

On 21 September 2026 by xAI (SpaceXAI).

How much does Grok 4.7 cost?

$2 per million input tokens and $6 per million output tokens, with cached input at $0.50. If a prompt reaches 200K tokens, the whole request is billed at $4 / $12 ($1 cached). Check xAI's docs for live pricing before budgeting.

What is Grok 4.7's context window?

500,000 tokens, with text and image input.

Is Grok 4.7 better than Grok 4.6?

Yes, on every benchmark xAI published: for example CursorBench 4.0 went from 40.4% to 46.3% and Terminal-Bench 4.0 went from 20.3% to 37.6%. It also uses more output tokens per task.

Is Grok 4.7 in GitHub Copilot?

Yes. It is rolling out gradually to Copilot Pro, Pro+, Max, Business and Enterprise plans.

Can I use Grok 4.7 in the Grok app?

Not yet. xAI says it will add Grok 4.7 to its web, mobile and X apps later. For now, it's available through the API, Cursor, Grok Build, Copilot and gateways.

Bottom line

Grok 4.7 is not a small version bump. The real change is a bigger, longer-trained model at the same low price, with big jumps on long coding and terminal tasks. The trade-off is that it spends more tokens per task, so measure your real bill.

If your team codes in Cursor or Copilot, this is a cheap test worth running this week.

Working with AI tools every day? Keep your sensitive data in your browser. SR Infobiz offers 28+ free, privacy-first tools that run entirely in your browser with no sign-up, including a JSON Formatter, JWT Decoder, Hash Generator and One-Time Secret Sharing for API keys: explore the free tools.

Need help choosing an AI model, building an AI workflow, or shipping software for your business? Our team can help. Email inquiry@srinfobiz.com.