Grok 4.6 and Grok Bot: xAI's play for the always-on coworker

xAI released Grok 4.6 on August 12, 2026, one day after Grok Bot entered public beta. Taken together they are the clearest statement yet of where xAI wants to sell: not a chat window, a workforce. The model is a genuine step up from 4.5. The bot product is the more interesting and the more dangerous of the two.

Grok 4.6: the numbers, with their sources

On the Artificial Analysis Intelligence Index, an independent aggregate, Grok 4.6 scores 61, tied with GPT-5.6 Sol and one point behind Claude Fable 5. The jumps over its own predecessor are large: DeepSWE from 54 to 65.9, APEX-Agents from 47.1 to 57.5. It also tops the GDPVal-AA row, an economic-value benchmark, at 1753. Context stays at 500,000 tokens, there are now four reasoning-effort levels with a new xhigh, and the knowledge cutoff is February 1, 2026.

The honest read from that table is the one GEO Toolbox gave: Fable wins more rows than anyone, most of Grok's wins are inside the noise, and Grok 4.6 is a credible mid-priced frontier model rather than the top of the index. The word doing the work there is mid-priced.

Update, September 3, 2026: Claude Fable 5.1 shipped on September 1 and is not in this table. The only shared row published so far is Terminal-Bench 4.0, where Anthropic lists 5.1 at 55.8 and GPT-5.6 Sol at 37.3, as reported by Tech-ish; xAI has not published a Grok 4.6 figure on it. Read the Fable column above as the June model.

Two dollars in and six out puts Grok 4.6 level with GPT-5.6 Luna on output price while scoring like Sol. That is the pitch, and for prompts under 200,000 tokens it is a real one. Cross that line and the entire request bills at double, which erases the advantage for anyone stuffing a whole knowledge base into every call. Availability at launch was API, Grok Build, Cursor, Microsoft Office add-ins, and the usual gateways; the consumer chatbot was not on the launch list.

Grok Bot: an agent with its own computer

Grok Bot launched in beta on August 11. The idea is simple to state and hard to do well: you create a persistent Bot with a job, give it access to the applications and websites it needs, and it works through multi-step tasks on its own computer in the cloud, with a browser, a filesystem, and a terminal. It signs into your tools with your credentials, keeps going when your laptop is closed, and comes back when it needs approval or has finished. Bots coordinate in shared threads, learn routines from demonstrations, and carry memory across conversations.

The part worth pausing on: Bots work through the user interface, which means they can operate legacy software and internal systems that have no API and never will. For a lot of local businesses, that is the entire back office. It also means a Bot with the wrong permissions can edit CRM records, answer support tickets, and change vendor orders exactly as confidently as it does the right things. VentureBeat flagged the operational risk plainly, and so will anyone who has watched an automation misfire against production data.

What this means for agencies and businesses

• Model choice is now a cost lever, not a loyalty. Grok 4.6 under 200K tokens is a legitimate default for research, drafting, and agentic runs that do not need the last point on the index.
• Always-on agents need an approval design before they need a subscription. Which actions run unattended, which pause for a human, and what a Bot is never allowed to touch is the whole project.
• The UI-driving approach is the unlock for shops running old software. It is also the reason to scope permissions per Bot, per system, and test against a copy first.