Grok 4.6 and Grok Bot: a guide to AI that works across apps
Grok 4.6 updates the AI model; Grok Bot adds a way to work in applications. Learn the difference and how to test a useful task before giving it more access.
By Monolith
xAI released Grok 4.6 on August 12, 2026. It focuses on tasks where AI takes several steps, including coding and visual work. A separate September 3 update to Grok Bot adds business controls around an AI worker using a cloud computer. Grok 4.6 is the model; Grok Bot is a product that uses an agent to work in applications.13
Understand the model and the worker around it
A model can interpret instructions and generate an answer. A product such as Grok Bot adds a computer and connections so that the model can attempt steps in applications. Knowing which part is being offered helps you ask a provider the right questions.
Suppose a weekly task involves opening a dashboard, collecting figures and putting them into a report. A model on its own does not have your dashboard account. A connected agent could be given approved access and asked to prepare the draft. The value would come from completing those steps correctly with less manual copying.
What changed from Grok 4.5
xAI reports improvement in coding, research and working inside applications. It also says the model checks its progress more often during a long task. In a useful business process, that might mean returning to a source file when an answer looks incomplete, rather than continuing from a mistaken assumption.1
The practical test is whether those checks improve the result. Give it a task with a known finish line and inspect the work it leaves behind. Time spent “thinking” is only useful when it helps the system complete the right job.
| Evaluation | Grok 4.5 High | Grok 4.6 High | GPT-5.6 Sol Max | Fable 5 Max |
|---|---|---|---|---|
| DeepSWE v1.1 | 54% | 65.9% | 73% | 70% |
| APEX-Agents | 47.1% | 57.5% | 56.7% | 59.2% |
| CursorBench v3.2 | 66.7% | 69.9% | 67.2% | 70.5% |
| AA Intelligence Index | 56 | 61 | 61 | 62 |
Each row in the table measures a different skill. Read across a row to compare that task, rather than treating a high number on one row as better than a lower number on another. The table shows xAI’s reported improvements over Grok 4.5.
Competitor scores also come from different sources and effort settings. Effort controls how much processing a model spends on a response. Grok and Sol having the same displayed Intelligence Index score does not establish that they are equally good at your client work.1
API pricing and availability
Developers pay usage fees when connecting Grok to software. Input is the information sent in; output is what the model generates. A faster option can cost more per unit, so compare both the turnaround time and the total cost of the task.
| Detail | What it means |
|---|---|
| Starting API rates | $2 input and $6 output per million tokens. |
| Fast variant | Twice the starting rate. |
| Bedrock announcement | $0.50 cached input per million tokens; 500K context window. |
| Effort options | Low, medium, high and xhigh. Confirm the provider and request conditions. |
Grok 4.6 launched through Cursor, Grok Build, the API and listed partners, with Bedrock support following on August 19. These are different ways to access the model. Temporary launch promotions and developer rates do not establish the price or allowance of a consumer subscription.12
What Grok Bot adds
Grok Bot works through applications on a cloud computer, a computer hosted remotely rather than on your desk. xAI describes teaching routines through demonstrations and reusing templates for later tasks. The product supplies a working environment around the AI model.3
The September enterprise update adds ways to control who can use the Bot, which networks it can reach and what activity is recorded. These features matter when several people use the same business systems. The introductory two-week free-use offer was temporary.3
Working through an app's interface can help when that app has no suitable API, a direct connection for other software. It also creates room for mistakes, such as selecting the wrong customer record. Decide what the Bot can read and draft, and which actions require approval, before connecting accounts.
For example, give the Bot permission to read the records needed for a report and save a draft. It does not need permission to change the customer database for that assignment. Keeping access tied to the actual job also makes errors easier to investigate: your team knows which systems the task was supposed to touch.
| System | Initial access | Human approval |
|---|---|---|
| CRM | Read permitted records; prepare notes | Changing stages or sending follow-ups |
| Support desk | Summarize tickets; draft replies | Sending, refunds and policy exceptions |
| Vendor portal | Read approved information | Placing or changing orders |
| Accounting | Read-only summaries | Payments and financial-record changes |
Practical uses to test
Try a website check with a clear assignment: open specified pages, identify a visible problem and save a report with screenshots. Or ask for a draft report from a known set of files. Both tasks produce something a person can inspect before any customer-facing change.
Compare the result with the way your team currently does the work. Include the time spent finding missing files, correcting errors and confirming completion. A convincing final message does not show that the right page was checked or the correct figures were used.
A useful trial ends with a reviewable result: a report, a draft reply or a proposed change to a customer record. Decide whether that result saves enough work to continue. Expand access only when there is a specific next task and your team understands how it will be checked.
Terms used in this article
| Term | Plain-language meaning |
|---|---|
| Token | A small unit of information counted by the model. Usage fees often depend on how many are read and generated. |
| Input / output | The information you send / the response the model generates. |
| Cached input | Eligible information stored for reuse, sometimes charged at a lower rate. |
| Benchmark | A defined test. Its score depends on the tasks, settings and supporting tools. |
| Context window | How much information fits into one request; it is not a guarantee of perfect recall. |
When did Grok 4.6 launch?
Are all the benchmark figures independently measured?
What does Grok 4.6 cost?
Read for this feature. The numbers match the markers in the text.
Want to apply this to a project? Here is the related service.
25 AI instructions to try on everyday work
Prompts are instructions you give an AI tool. This PDF includes 25 examples for inquiries, marketing and admin. Adapt them to your task and check the results.