Grok 4.6 for Professionals: What It's Good At, What It Isn't, and the Data Question (September 2026)

The straight answer: Grok 4.6 is a serious work tool for one job, which is coding and cheap agent work, and a poor choice for writing, analysis, or anything a client will read. Before you put a single work document into it, read the data section below, because the consumer app and the developer API have opposite defaults. If you only remember one thing, remember that.

We have only covered Grok in passing, in the 2026 model launch scorecard. The questions readers keep asking are more basic: who makes it now, which version is current, what it costs, and whether it is safe for work. Here is all of that in one place.

Who makes Grok now

Grok is built by xAI (now SpaceXAI). SpaceX acquired xAI on February 2, 2026, and press reports say the unit was rebranded SpaceXAI in early July. For professionals this mostly changes who you are contracting with: the entity on any API agreement or enterprise contract is now a SpaceX subsidiary.

The bigger structural change is Cursor. SpaceX closed its $60 billion acquisition of Cursor, the AI code editor, on August 14, 2026. Cursor is now a first-party Grok surface, which is why the Grok 4.6 launch material talks about Cursor quotas and CursorBench scores rather than the chat app. If your developers use Cursor, they are using a SpaceXAI product whether or not they ever open grok.com.

One more piece of context, attributed to press coverage rather than the company: TechCrunch and Gizmodo have reported that Musk told staff Grok is "losing the AI race to Anthropic," and that all of xAI's original co-founders have departed. Read the product on its merits, but know that the organization behind it is in flux.

Which Grok is current

There is no Grok 5. The newest released model is Grok 4.6, which shipped on August 12, 2026, with the API id grok-4.6. It replaced Grok 4.5, which came out on July 8 and is still listed. A cheaper mid-tier model, Grok 4.3, is also still available.

Musk said on September 2 that Grok 4.7 "comes out in 10 days." Treat that as an announcement only. Nothing has been published about its capabilities or pricing, and xAI's release dates have slipped before. If you are evaluating Grok for a team this month, evaluate 4.6.

In the chat app, the picture is muddier. Paid tiers run Grok 4.5 or 4.6, and the Free tier runs an older, lighter model. xAI has not published a clear table of which tier gets which model as the default, so if the version matters to you, check the model picker after you sign in.

What it's actually good at

Coding and agent work, at a price that undercuts most of the competition.

xAI's own numbers put Grok 4.6 at 65.9% on DeepSWE v1.1, the long-horizon software engineering benchmark, and 69.9% on CursorBench. Artificial Analysis, the independent benchmark house, scores it at 61 on its Intelligence Index, which ties GPT-5.6 Sol. Its headline was that 4.6 "returns SpaceXAI to the intelligence frontier and leads on cost efficiency." The Hacker News thread on the launch ran to 632 points and 616 comments, and the consensus there matches the benchmark story: this is a capable model for code.

The cost side is where it gets interesting for teams. The API price is $2 per million input tokens and $6 per million output, which is in the same band as Claude Sonnet 5 rather than the frontier tier. Inside Cursor, the practical limit is generous. One user, @Fei2411 on X, described Grok 4.6 as close to unlimited usage inside Cursor on a paid plan. For a team that burns through coding-assistant quotas every month, that is the real pitch.

The model supports four effort levels, from low to xhigh, and a 500,000-token context window on xAI's own API. Microsoft Foundry lists it in preview at 200,000 tokens. There is no Batch API, so if your workflow relies on cheap overnight batch processing, Grok is not a drop-in replacement for the providers that offer it.

What it isn't good at

Writing, and any task where judgment and tone carry the value.

The early professional reaction on X has been unusually consistent. Tom Glass, posting as @TomGlassAI, wrote: "Grok 4.6 just isn't that good of a model for tasks outside coding. Its communication style + intuition are terrible." Another user, @aye_pete, was blunter: "Grok 4.6 is HORRIBLE at writing." And @A_A_Toulkeridis, comparing it to Anthropic's flagship, concluded: "Grok 4.6 is definitely not Opus 5 level."

Those are individual opinions, not benchmarks, but they line up with what the benchmarks do not measure. An Intelligence Index score of 61 says the model can solve hard, well-defined problems. It says nothing about whether it will draft a client email that sounds like you, or notice an unusual contract clause without being told to look. Those are the tasks most non-developer professionals actually need, and on those the community verdict is a clear no.

For which model suits which kind of work, see our comparison of ChatGPT, Claude, and Gemini for professionals. Grok is not in that comparison yet, and on current evidence it would sit with the coding-first models rather than the general assistants.

What it costs

API pricing, per million tokens, from xAI's developer documentation:

Model	Input	Cached input	Output	Context
Grok 4.6	$2.00	$0.50	$6.00	500K
Grok 4.5	$2.00	$0.30	$6.00	500K
Grok 4.3	$1.25	$0.20	$2.50	1M

Two footnotes on the 4.6 row. Prices double for requests above 200,000 tokens of context, and a "fast" variant is offered at twice the standard price. The knowledge cutoff is February 1, 2026. The model is available through the API, Cursor, Grok Build, OpenRouter, Vercel, Cloudflare, and Microsoft Foundry.

Consumer plans are harder to pin down, because xAI's help pages block automated fetching and the tiers have changed several times this year. Third-party price trackers broadly agree on the following, so treat these as approximate and check grok.com before you buy:

• Free: a lighter model, not the flagship
• X Premium: roughly $8 per month
• SuperGrok Lite: roughly $10 per month
• SuperGrok: roughly $30 per month
• X Premium+: roughly $40 per month
• SuperGrok Plus: reported at about $100 per month, but this figure is the least certain
• SuperGrok Heavy: $300 per month

The flagship models are paid-only. If you are on the free tier, you are not testing Grok 4.6.

Can you put work data in it?

This is the section to read twice. Grok has two front doors, and their data defaults are opposites.

The API and tools built on it. xAI's security documentation says it never trains on API inputs or outputs without permission. Data is retained for 30 days, encrypted, and there is a zero-data-retention toggle in the developer console. xAI reports SOC 2 Type 2 certification. Those are the same commitments the major providers make for their business tiers, and they cover Cursor and the other integrations that call the API.

The consumer grok.com app. According to press coverage and the app's own settings, Grok trains on your chats by default unless you turn off "Improve the model" under Settings > Data, or use Private Chat for a given conversation. For users outside the EU, public posts on X are also used for training. Ireland's Data Protection Commission has an open investigation into xAI's use of European users' data. Our general guide to whether AI trains on your data covers the same setting for every major provider; Grok is the one where the default is most clearly against you.

Then there are the incidents. These are the reasons compliance teams hesitate, and a work audience should know them:

• July 12 to 14: the Grok Build command-line tool uploaded users' entire Git repositories, including committed secrets, to a Google Cloud storage bucket, even when users had told it not to and had the opt-out enabled. xAI disabled uploads. Axios and The Next Web reported it; no public postmortem has been found.
• August 19 to 24: Grok returned gibberish output to users for several days, as reported by TechCrunch.
• Late August: a class action was filed alleging that Grok was trained on child sexual abuse material, reported by Gizmodo. This is an allegation, not a finding.
• First half of 2026: a Resemble AI report tied Grok to 87% of the deepfake files it could trace to a source model.

None of that changes the API's contractual terms. It does change how much benefit of the doubt a security reviewer is likely to extend, and it is fair to expect a longer vendor review for Grok than for its competitors.

Grok Bot and the "AI teammates" pitch

On August 11, xAI launched Grok Bot in beta, positioned as "AI teammates" that run on their own cloud computer, sign into your apps, and carry out tasks while you do something else. It is available on SuperGrok, SuperGrok Plus, SuperGrok Heavy, and Cursor Pro+. On September 3, xAI followed with Grok Bot for Enterprise, which adds access controls, network restrictions, and audit logging. No pricing has been published for the enterprise version.

The pitch is the same one every provider is now making: an agent with its own browser and credentials that acts on your behalf. Given the July incident above, where a Grok tool ignored an explicit opt-out, the sensible posture is to try Grok Bot on accounts and data you could afford to lose before connecting it to anything else. Grok Build, the app builder, and Grok Imagine, the image and video generator, round out the line, and Grok voice is reported to be in Tesla vehicles.

Full guide: What Is Grok Bot, and Should It Sign Into Your Work Apps?.

Bottom line for professionals

• If your team writes software and is quota-constrained, Grok 4.6 through Cursor or the API is worth a real trial. It is cheap, capable, and the API data terms are standard.
• If you write, advise, analyze, or produce anything a client reads, stay with Claude, ChatGPT, or Gemini. The community verdict on Grok's writing and judgment is not close.
• If you use the grok.com app at all, turn off "Improve the model" in Settings > Data today, or use Private Chat. If your work is regulated or confidential, do not put it in the consumer app under any setting.

Sources

1. Grok 4.6 — xAI news, August 12, 2026
2. Models and pricing — xAI developer documentation
3. Security FAQ — xAI developer documentation
4. Introducing Grok Bot — xAI news, August 11, 2026
5. Grok Bot for Enterprise — xAI news, September 3, 2026
6. Grok 4.6: benchmarks and analysis — Artificial Analysis
7. Use Grok models in Microsoft Foundry — Microsoft Learn
8. SpaceXAI's Grok uploaded customer code to the cloud despite opt-outs — Axios, July 14, 2026
9. Grok keeps sending gibberish responses to users — TechCrunch, August 20, 2026
10. SpaceXAI releases Grok 4.5, which Elon describes as an Opus-class model — TechCrunch, July 8, 2026
11. @TomGlassAI on X: "Grok 4.6 just isn't that good of a model for tasks outside coding"
12. @aye_pete on X: "Grok 4.6 is HORRIBLE at writing."
13. @A_A_Toulkeridis on X: "Grok 4.6 is definitely not Opus 5 level"
14. @Fei2411 on X, on Grok 4.6 usage limits inside Cursor
15. @JayaNayak21 on X, relaying Musk: "Grok 4.7 will be released in 10 days"