Grok 4.6 is xAI's newest flagship model, announced on August 12, 2026. It's a reasoning model (it thinks step by step before answering) with a 500K-token context window, tuned for coding and long multi-step agentic work. On Secret Chat it powers the whole Smart Grok family.

How is Grok 4.6 different from Grok 4.5?

It's an incremental upgrade announced about five weeks later. xAI describes the gains as post-training improvements rather than a new foundation: better performance on coding and agentic benchmarks, more self-testing on long tasks, and stronger first passes on visual and interactive projects. The context window and always-on reasoning are unchanged.

How do I use Grok 4.6 on Secret Chat?

Open the Secret Chat app and select the "Grok" model (the Smart one). Leave it on Smart Agentic for automatic routing, or pick a specific mode — Web Search, Deep Research or image generation. Every answer then comes from Grok 4.6.

Can I turn off Grok 4.6's reasoning?

No — Grok 4.6 always reasons before answering; reasoning can't be disabled. You can influence how hard it thinks (low/medium/high effort), and Secret Chat runs it at high effort by default. If you want a direct, non-reasoning answer instead, use the free DeepSeek USA model.

Does Grok 4.6 cost more credits than Grok 4.5?

Essentially no. The two rates that dominate a bill are identical ($2 per million input tokens and $6 per million output), so for ordinary use your credits go just as far. One rate did rise: reading a cached prompt prefix went from $0.30 to $0.50 per million tokens, which affects the earlier turns of a long conversation being re-read, not anything you type fresh. We bill Grok from the cost xAI reports on each response, so you pay the real figure rather than our estimate. The free DeepSeek USA model remains free for everyone.

Is Grok 4.6 private on Secret Chat?

Secret Chat forwards your prompts through its own proxy, so you use Grok 4.6 without an xAI account and without your requests being tied to you. No profile is built, no chat is associated with your identity, and your history stays in your browser. Provider-side retention may still apply, but your query reaches the model anonymized, not linked to your email or identity. Note that the content of your messages is sent verbatim: only your identity is removed.

Can Grok 4.6 search the web on Secret Chat?

Yes. Its Web Search and Deep Research modes attach xAI's live web and X search, so Grok 4.6 combines its reasoning with current information instead of being limited to training-cutoff knowledge.

Can I have Grok answer a question alongside the other models?

Yes; that's the AI Council. It sends one question to several models from different companies at once, reveals each answer as that model finishes, and then has a referee build a disagreement table showing which model asserted each claim, which contradicted it and which never mentioned it, plus a short synthesis of what at least two models agreed on. Grok is one of the four families available: run the Quartet for ChatGPT, Claude, Gemini and Grok together, or pick a two- or three-model panel that includes Grok. Panels use each family's mid tier rather than its heaviest, because the evidence comes from different companies disagreeing rather than from any single answer being the deepest obtainable.

Does Grok 4.6 also generate images?

Image generation in the Grok model runs on a separate image model, Grok Imagine 2, which this text upgrade doesn't change. You select it as the image generation mode next to the message box.

Grok 4.6 Is Now on Secret Chat AI: xAI's Newest Model

Published August 13, 2026 · Updated August 19, 2026

xAI's Grok 4.6 is now live on Secret Chat AI. The whole Smart Grok family (the automatic Smart Agentic mode and every manual mode behind it) has moved to xAI's newest model, one day after xAI announced it. Same place, same private workspace where you already chat with GPT, Claude, Gemini and Perplexity; a better model answering.

Grok 4.6 arrives about four weeks after Grok 4.5, which tells you what kind of release this is: not a new foundation, but a sharper version of the one that was already there. Below is what actually changed, what didn't, and (as always) how you get all of it without a profile ever being built around you.

What Is Grok 4.6?

Grok 4.6 is xAI's latest frontier model, announced on August 12, 2026. Like its predecessor it's a reasoning model: it works through a problem step by step before it replies. The headline points from xAI's own announcement:

Frontier-level on xAI's own scoring. xAI says Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, at a score of 61: a vendor's own benchmark claim, so read it as one, but it puts the model in the same conversation as the top of everyone else's lineup.
Tuned for long, multi-step work. xAI's pitch is that it "stays with complex tasks across many steps," with more self-testing and verification as a task runs long, and stronger first passes on visual and interactive projects. The gains are described as post-training work (better fine-tuning and reinforcement learning) rather than a bigger model underneath.
Reasoning is still always on. Grok 4.6 always reasons before answering; there's no way to switch that off. You can dial how hard it thinks (low, medium or high effort), and on Secret Chat it runs at high effort by default, so you get its deepest deliberation on every reply.
500K-token context window. Unchanged from Grok 4.5 — long enough to hold a large codebase, a contract set or a long research thread in a single conversation.

The honest summary: this is an incremental release, and a fast one. If Grok 4.5 was already doing what you needed, 4.6 will feel like the same model on a better day — most noticeably on long coding and agentic tasks, which is exactly where xAI aimed it.

Grok 4.6 Powers the Whole Smart Grok Family

On Secret Chat, "Grok" isn't a single toggle — it's the Smart Grok model with an intelligent Smart Agentic router plus a set of manual modes you can choose yourself. As of this release, every one of those runs on Grok 4.6:

Smart Agentic, the default. A router reads each message and decides how to handle it: a quick reply for casual chat, a web search when you need fresh facts, or deep research for an involved question — now all answered by Grok 4.6.
Web Search — Grok 4.6 with xAI's live web and X (Twitter) search attached, so answers combine its reasoning with current information rather than training-cutoff knowledge.
Deep Research — the same search tools at maximum reasoning effort, for questions worth waiting on.
Image generation — unchanged by this swap. Grok's image mode runs on Grok Imagine 2, a separate image model, so it isn't affected by the text upgrade.

You can also attach a PDF to any of them and ask about it — Grok reads the document and answers from it, and the copy we upload for that purpose is deleted again as soon as your answer comes back.

What Did Not Change

Three things are worth naming, because "new model" often quietly means "new everything":

The price you pay is almost unchanged, and we would rather be precise than round it off. xAI kept the two rates that dominate a bill identical ($2 per million input tokens and $6 per million output), so for ordinary use your credits go just as far. One rate did move: reading a cached prompt prefix went from $0.30 to $0.50 per million tokens. That applies to the earlier turns of a long conversation being re-read rather than to anything you type fresh, so it's a small share of a typical chat, and it's larger on very long threads. We bill Grok from the cost xAI itself reports on each response, so you pay the real figure either way rather than an estimate of ours.
The free quick tier is still not Grok. Grok 4.6 always reasons before replying and can't be told not to, so routing every throwaway question through a full reasoning pass would be slower and pricier for no benefit. That role still belongs to DeepSeek USA — free for everyone and served only from US providers, never from DeepSeek's own Chinese API. See why open weights make that possible.
The privacy model. Covered below — it's a property of the gateway, not of the model behind it.

How to Use Grok 4.6 on Secret Chat AI

There's nothing to install or switch on — it's already live:

Open the Secret Chat app and select the Grok model in the model picker.
Leave it on Smart Agentic to let the router handle everything, or pick a specific mode next to the message box — Web Search, Deep Research, or image generation — when you already know what you want.

From there, every answer comes from Grok 4.6. If you just want a fast, no-frills reply for a simple question, switch to the free DeepSeek USA model instead. Since 4.6 was tuned for tasks that run long, it rewards a prompt that gives it something substantial to chew on:

I want your best reasoning on a hard problem. Take your time, work through it step by step, weigh the approaches you would reject and why, check your own answer before you give it to me, and then give me your strongest recommendation. Here's the problem:

Private by Design, Whatever Grok You Pick

The upgrade changes the model, not the privacy. Secret Chat is a private gateway between you and the AI providers: your prompts are forwarded through our proxy, so you don't need an xAI account and your requests are never tied to you on the provider's side. We don't build a profile of you, and no chat is ever associated with your identity. Your conversation history lives locally in your own browser, and a prompt exists on our side only for as long as it takes to fetch your answer; there's no stored chat archive on our servers. You can pay with crypto if you'd rather not link a card to your AI use.

That matters especially with Grok. As we covered in Does Grok Save Your Chats?, xAI trains on Grok conversations by default and keeps your history. Reaching Grok 4.6 through Secret Chat lets you use the exact same model as a stranger: whatever xAI holds under its own terms and configuration, the record on that side carries our gateway's credentials and server address, not your name, your account or your IP.

One honest caveat, as always. "Anonymously" describes the link, not the words: no account identifier travels with your prompt, but that isn't a claim that the text stops being identifying. Write your own name or your case number into a message and it's all still sitting there in the message. Secret Chat removes you from your queries (it doesn't remove the data from your messages), so redacting identifying details before you send is still your call. Anonymity is also not privilege, not a legal exemption, and not a way to put anything beyond the reach of a court that's entitled to it. For the full picture of how anonymized Grok access works, see our dedicated Private Grok page.

If you like having several models on tap for the same question, that's the whole idea. See why using GPT, Claude, Gemini, Grok and Perplexity together beats relying on any single one. The AI Council automates it, and Grok is one of the four families that sit at the table. One question goes to several models from different companies at once, each answer is revealed the moment that model finishes, and a cheap referee model then extracts the factual claims into a disagreement table — which model asserted each claim, which contradicted it, which never mentioned it — followed by a short synthesis of what at least two models agreed on and a list of things to verify before acting. You choose the size: a Duet of two models, a Trio of three, or the Quartet, which runs ChatGPT, Claude, Gemini and Grok together. On the Duet and Trio you also choose which families take the seats, so you can put Grok next to Claude and nothing else if that's the comparison you want.

Two details about how Grok answers there. Panels seat each family's mid tier rather than its heaviest: a council is evidence because independent companies disagree, not because each individual answer is the deepest one obtainable, and depth from one model is what Deep Research is for. And when a question turns on current or checkable facts, the whole panel answers with live web search attached, each provider using its own, for Grok that means the same xAI web and X search described above. Grounding is decided once for the entire panel, so the members stay comparable. Two limits we would rather state than imply: agreement is evidence, not proof, because models share training data and can be wrong together, which makes disagreement the reliable signal; and a council widens where your question travels, since it reaches every model on the panel plus the referee. Each request is still anonymous, but anonymization protects who you are, not what you type. The finished check is stored only on your device, and the Session Privacy Report lists every model that took part.

Conclusion

Grok 4.6 is xAI's newest model, and it now powers the entire Smart Grok family on Secret Chat — Smart Agentic, Web Search and Deep Research alike, with image generation continuing on Grok Imagine 2. It's an incremental step rather than a reinvention: the same 500K context and always-on reasoning, sharpened for long, multi-step work, at no change to what your credits buy.

Try it today at Secret Chat AI, with the full multi-model lineup, your history in your own browser, and every query reaching the model anonymously.

Frequently Asked Questions

What is Grok 4.6?
Grok 4.6 is xAI's newest flagship model, announced on August 12, 2026. It's a reasoning model (it thinks step by step before answering) with a 500K-token context window, tuned for coding and long multi-step agentic work. On Secret Chat it powers the whole Smart Grok family.
How is Grok 4.6 different from Grok 4.5?
It's an incremental upgrade announced about five weeks later. xAI describes the gains as post-training improvements rather than a new foundation: better performance on coding and agentic benchmarks, more self-testing on long tasks, and stronger first passes on visual and interactive projects. The context window and always-on reasoning are unchanged.
How do I use Grok 4.6 on Secret Chat?
Open the Secret Chat app and select the "Grok" model (the Smart one). Leave it on Smart Agentic for automatic routing, or pick a specific mode — Web Search, Deep Research or image generation. Every answer then comes from Grok 4.6.
Can I turn off Grok 4.6's reasoning?
No — Grok 4.6 always reasons before answering; reasoning can't be disabled. You can influence how hard it thinks (low/medium/high effort), and Secret Chat runs it at high effort by default. If you want a direct, non-reasoning answer instead, use the free DeepSeek USA model.
Does Grok 4.6 cost more credits than Grok 4.5?
Essentially no. The two rates that dominate a bill are identical ($2 per million input tokens and $6 per million output), so for ordinary use your credits go just as far. One rate did rise: reading a cached prompt prefix went from $0.30 to $0.50 per million tokens, which affects the earlier turns of a long conversation being re-read, not anything you type fresh. We bill Grok from the cost xAI reports on each response, so you pay the real figure rather than our estimate. The free DeepSeek USA model remains free for everyone.
Is Grok 4.6 private on Secret Chat?
Secret Chat forwards your prompts through its own proxy, so you use Grok 4.6 without an xAI account and without your requests being tied to you. No profile is built, no chat is associated with your identity, and your history stays in your browser. Provider-side retention may still apply, but your query reaches the model anonymized, not linked to your email or identity. Note that the content of your messages is sent verbatim: only your identity is removed.
Can Grok 4.6 search the web on Secret Chat?
Yes. Its Web Search and Deep Research modes attach xAI's live web and X search, so Grok 4.6 combines its reasoning with current information instead of being limited to training-cutoff knowledge.
Can I have Grok answer a question alongside the other models?
Yes; that's the AI Council. It sends one question to several models from different companies at once, reveals each answer as that model finishes, and then has a referee build a disagreement table showing which model asserted each claim, which contradicted it and which never mentioned it, plus a short synthesis of what at least two models agreed on. Grok is one of the four families available: run the Quartet for ChatGPT, Claude, Gemini and Grok together, or pick a two- or three-model panel that includes Grok. Panels use each family's mid tier rather than its heaviest, because the evidence comes from different companies disagreeing rather than from any single answer being the deepest obtainable.
Does Grok 4.6 also generate images?
Image generation in the Grok model runs on a separate image model, Grok Imagine 2, which this text upgrade doesn't change. You select it as the image generation mode next to the message box.