Grok 4.6 builds on Grok 4.5 with longer multi-step task execution, stronger first passes for interactive and visual applications, and more self-testing during extended work.
OUR VERDICT
A significant agent-focused update, especially for Grok Build and Cursor users
Grok 4.6 is most relevant to developers and teams building applications or artifacts across many steps. Its value should be measured on completed, reviewed work rather than self-reported benchmark scores.
What changed
SpaceXAI says Grok 4.6 received a longer supplemental training run, regenerated supervised trajectories, and agentic reinforcement learning across coding, knowledge work, web development, CAD, and other technical environments. The company reports improvements in sustained execution, self-testing, and visual or interactive first drafts.
Availability and price
Grok 4.6 is available in Grok Build, Cursor, the SpaceXAI API, and partners including OpenRouter, Vercel, and Cloudflare. API pricing starts at $2 per million input tokens and $6 per million output tokens; a fast variant is listed at twice the price. Verify exact model names, limits, regional access, and current price before deployment.
How to evaluate
Use long tasks with explicit acceptance tests. Measure completed requirements, failing tests, visual corrections, unsafe actions, intervention rate, recovery, latency, and total token cost. Compare with Grok 4.5 using the same harness.
Bottom line
Grok 4.6 is a genuine product update focused on sustained agent work. The strongest evidence will come from whether it finishes a representative project with fewer interventions and corrections.
Read the Grok review, ChatGPT vs. Grok, and best AI agents for work.
Our editorial read
A broad assistant option for current-event exploration, brainstorming and multimodal tasks.
Best fit
Users who value a conversational interface tied to a fast-moving information environment and can verify claims.
Where we would hesitate
Avoid treating fast answers as verified reporting, especially for legal, financial or safety-sensitive questions.
Decision rule
Choose Grok for exploratory breadth; choose a source-grounded research tool when citation quality is the deciding factor.
Evidence note: This judgment synthesises the linked official documentation, release notes and product context. It is not a claim that AI Tools Manual independently benchmarked every tool.
Official source check
Official sources used for product facts: x.ai, help.x.com. These links support named features, access, plan or policy statements; they are not independent benchmark results.
The links above are included because this page names or compares these products. They support the documented product facts only; the editorial judgment, workflow advice and limitations remain our own and are not presented as independent test results.
Recent official changes
SpaceXAI’s official news page adds two availability changes after the Grok 4.6 launch: Grok Bot is now included with SuperGrok Plus, Cursor Pro+, and all Cursor Teams plans (Aug 21), and Grok 4.6 is available through the Gemini Enterprise Agent Platform and Amazon Bedrock. This changes the access and integration decision more than the underlying model verdict. Check the partner plan, region, model name, rate limits, data handling, and billing path before deployment. The announcement does not establish independent quality or cost performance.
Primary sources
- https://x.ai/news?category=all
- https://x.ai/news/grok-bot-is-now-included-with-more-plans
- https://x.ai/news/grok-4-6-on-amazon-bedrock
High-risk review required: verify partner availability, plan entitlements, regional access, model identity, billing, data handling, and limits before publication. The announced features have not been independently tested by AI Tools Manual.
August 2026 update: Grok Bot connects to X
xAI announced that Grok Bot can now connect to an X account. After signing in with the X connector, a Bot can search posts, read a timeline, check mentions, and summarize what is happening on X. Paid Grok Bot users receive free X API credits to start, while xAI creates a developer account when needed.
Editorial judgment: This is useful for social listening and rapid context gathering, but it changes the data-permission boundary of the workflow. Teams should decide which account, lists, and mentions may be read, and should treat any generated summary as a starting point rather than a verified record. Do not let an agent publish, reply, or take account actions without an explicit approval step.
What to verify before connecting
- Which X data the connector can read and which actions it can request.
- Whether API credits, plan access, and regional availability apply to your account.
- How long prompts, retrieved posts, and outputs are retained.
- Whether sensitive or private information may be pulled into the Bot context.
The announcement describes the first version of the integration; it does not establish independent accuracy, complete coverage of X, or a guarantee that summaries reflect every relevant post. We have not independently tested the connector. Review our AI privacy checklist and compare the Grok tool profile before enabling it.
September 2026 update: Grok Bot for Enterprise
SpaceXAI announced that Grok Bot is available for enterprises. Grok and Cursor Enterprise customers receive free usage for the first two weeks and can invite their whole organization, including people without an existing seat. The product describes Bots as persistent workers that run in isolated cloud environments, use connected apps and websites, and continue tasks until they finish or need a decision.
The enterprise release adds controls for access, networks, and auditing. SpaceXAI says each user’s work runs in a separate environment, and a Bot has no access by default; it only reaches accounts the user signs into. These are useful guardrails for organizations experimenting with always-on agents, but they do not remove the need to scope connectors, review actions, and set human approval points.
Before inviting a team
- Define allowed apps, domains, credentials, and network paths.
- Require review for messages, purchases, code merges, and other external actions.
- Test cancellation, retries, logs, and hand-off behavior.
- Confirm Enterprise eligibility, billing, retention, and regional availability.
What the recent Grok direction means
Grok Build is expanding from a restricted beta toward broader web and mobile availability, with publishing and sharing of the apps it creates. The Memory in Grok Build entry signals that projects may retain more context over time. Grok Bot is a separate agent concept that can operate across tools and return for approval. Readers should keep these surfaces distinct: a memory feature, a build environment, and an always-on bot have different risks.
Who benefits and who should be cautious
Creators and developers who want to prototype a small website, game, or dashboard may value the shorter path from prompt to a shareable result. Teams handling private repositories, customer data, or regulated actions should wait until retention, connector scope, approval behavior, and administrative controls are documented for their plan. Memory can improve continuity, but it can also preserve incorrect assumptions or sensitive context.
Our assessment and checks
The most useful question is not whether Grok can build an app, but whether a reviewer can see what it changed and stop unsafe actions. Start with a disposable project, inspect generated code and network calls, disable unnecessary connectors, and require a human checkpoint before publishing. Verify pricing and availability directly in the account because xAI’s news page does not establish universal access.
Primary source and review note
Primary source: xAI News. Re-check memory settings, plan coverage, safety controls, and terms before publication.
September 2026 official update: Grok Voice Transcribe 2.0
xAI’s official news page lists the September 18 launch of Grok Voice Transcribe 2.0 and describes it as a speech-to-text model focused on accuracy and cost effectiveness. The announcement is a product notice, not an independent test: it does not provide enough public detail to claim a specific word-error rate, language ranking, or total cost for every workload.
Who should care
Developers building voice notes, meeting capture, or audio search may want to add Grok to a short technical evaluation, especially if they already use xAI APIs. Buyers should first confirm the API surface, supported audio formats, regional availability, retention behavior, and whether pricing is separate from ordinary Grok usage. For regulated recordings, transcription should remain a reviewable draft until a human checks names, numbers, and permissions.
Editorial decision
Our view is to treat Voice Transcribe 2.0 as a promising new option rather than a proven leader. Grok is a good fit for teams that value a fast-moving vendor and can run their own acceptance checks; it is a poor fit when a procurement decision requires a published independent benchmark or a mature compliance package. See the Grok review, then use the guide library for a transcription QA checklist.
Primary source: xAI News (September 18, 2026 listing for Grok Voice Transcribe 2.0).
September 2026 update: Grok 4.7
xAI introduced Grok 4.7 on September 21 as its newest model for coding and knowledge work. The official announcement says it uses a larger base model and longer reinforcement-learning run, with more emphasis on multi-hour tasks, self-checking, longer context, and native understanding of the Grok Bot harness. xAI says it is available in Grok, Cursor, Grok Build, the API, and selected model routers and cloud platforms.
xAI lists starting API prices of $2 per million input tokens and $6 per million output tokens, with a faster variant at twice the price. It also publishes benchmark comparisons across coding, office work, legal, clinical reasoning, and cyber safety. These figures are xAI’s own reported evaluations and should not be presented as an independent ranking.
Safety and adoption limits
- xAI describes a new safeguard stack and lower risky-prompt pass-through; verify the model card and deployment controls before relying on that claim.
- Confirm which providers expose Grok 4.7, the exact effort settings, rate limits, retention, and data-use terms.
- Keep human review for production code, legal or clinical materials, and cyber actions.
- Separate evaluation traffic from customer data and set spend limits before enabling long-running agents.
Read xAI’s official Grok 4.7 announcement and developer documentation. Revisit our Grok review, ChatGPT vs Grok, and the previous Grok 4.6 update. No independent benchmark or first-hand test is claimed.
Grok Bot: the new agent layer under Grok
On August 11, 2026, xAI introduced Grok Bot as a beta layer inside the wider Grok ecosystem. It is not a separate foundation model or a replacement for the Grok chat page. Instead, xAI describes it as a team of always-on agents with a cloud computer of their own. A Bot can sign in to tools, apps, inboxes and websites, carry out a multi-step job, and return when it needs a human decision. This makes the product relevant to operations work rather than only question answering.
The announcement says Grok Bot is available to SuperGrok, SuperGrok Plus and SuperGrok Heavy subscribers, as well as Cursor Pro, Pro+, Ultra and selected Cursor Teams plans on desktop and iOS. Bot usage is separate from the ordinary Grok or Cursor allowance, and enterprise access is described as a waitlist. Confirm the entitlement, region and current terms in the account before making a purchase decision; the launch page is not a universal pricing table.
What the X connection adds
On August 29, xAI announced that a user can connect an X account and ask a Bot to search posts, read a timeline, check mentions or summarize activity. The first version can create a developer account when needed and gives paid Grok Bot users introductory X API credits. This is useful for social listening and a fast briefing, but it changes the permission boundary: the Bot can now ingest account data and the resulting summary can omit context or repeat an unverified claim.
Best fit and not fit
Best fit: small operations, sales and marketing teams that have repeatable browser-based work, can use a disposable account, and want a human approval step before a message, purchase, CRM edit or publication. Multiple Bots may be useful when a coordinator can route work to specialists.
Not fit yet: regulated workflows, production credentials, sensitive customer data, or unattended actions where an incorrect retry would be costly. The beta label, separate usage meter and broad connector scope mean buyers should not assume mature audit, retention, or recovery controls. Keep high-impact actions read-only until the account exposes the permissions and logs you need.
Our editorial assessment
We consider Grok Bot a meaningful new product surface under Grok, not evidence that Grok is automatically better at every task. The differentiator is persistent, computer-using execution: a Bot can continue while the user is away and can hand work between agents. Our decision rule is practical: choose it for reversible, well-scoped workflows with checkpoints; choose a conventional assistant when the job is a one-off answer or when you cannot grant a cloud agent access to the working environment.
Start with a low-risk task, limit connected accounts and domains, require approval for external actions, set a spend ceiling, and review the final artifacts rather than trusting a completion message. We have not independently tested Grok Bot, its X connector, or its enterprise controls. This page separates xAI’s documented claims from our editorial judgment and does not present the launch examples as benchmark evidence.
Primary sources and evidence limits
Product facts are drawn from xAI’s Introducing Grok Bot announcement, the Grok Bot and X integration notice, and the plans update. These primary sources document availability and intended behavior; they do not prove accuracy, uptime, complete X coverage, or safe unattended operation. Read our AI privacy checklist, then compare the workflow with ChatGPT vs Grok before enabling access.
September 2026 update: Grok Build is open on web and mobile
xAI’s August 19 announcement says Grok Build moved from an early beta limited to SuperGrok Heavy to availability on every plan, on the web, iOS and Android. A user can describe an app, game, website or dashboard in a Grok conversation and receive a working version. The announcement also adds publishing and sharing: each published app can receive a grok.me address, with access set to private, link-only or public, and an optional custom domain.
Grok Build can also connect an app to xAI APIs for chat, images and voice. xAI says the app-level access can be revoked, and the release adds Remix, GitHub export, secrets storage and connectors for business data. These are important workflow changes because the product is no longer only a prototype inside a chat; it becomes a path from prompt to a shareable artifact that can continue in an editor or use connected data.
Best fit and not fit
Best fit: creators, educators, small teams and developers who want to prototype a lightweight interactive experience, publish a demo quickly, or explore a dashboard idea before committing to a conventional stack. GitHub export is useful when a prototype needs code review and a real repository. The best first project is disposable and has no customer credentials.
Not fit: production systems that require a formal threat model, stable deployment controls, or guaranteed data residency before the implementation has been reviewed. A working preview is not proof of secure authentication, accessible UI, reliable persistence or safe connector behavior. Keep third-party secrets outside source code, limit connector scope, review network calls and require a human approval step before publishing.
Our editorial assessment
Our view is that Grok Build is now a useful rapid-prototyping surface under Grok, not a replacement for a tested application platform. Choose it when speed from idea to a reviewable artifact matters. Choose a conventional builder when you need repeatable CI, team permissions, test coverage and predictable hosting from the first commit. For autonomous operations, see the Grok Bot section; for model and assistant trade-offs, read ChatGPT vs Grok. Use our AI privacy checklist before connecting business data.
We have not independently tested Grok Build, its publishing flow, custom domains, GitHub export or connectors. The examples in xAI’s announcement are vendor demonstrations and should not be presented as reliability or security evidence.
Primary source and evidence limits
Product facts come from xAI’s Grok Build on web and mobile announcement. It documents availability and intended features, but it does not establish uptime, app security, accessibility, cost, regional rollout or performance for every project. Verify current plan access, API billing, connector permissions, secret handling, export behavior and domain controls before publishing a real app.
September 22, 2026 case study: Grok Bot in customer support
SpaceXAI published a case study describing how it uses Grok Bot across the combined SpaceXAI and Cursor support operation. The company says support volume rose 175% after the combination and that it resolved tickets without adding headcount. It also reports a $0.20–$0.30 cost per resolution after optimizations, compared with the $1–$4 range it attributes to traditional AI support tools. These figures are vendor-reported case-study claims, not independent benchmarks, and should not be copied into a buying comparison as universal pricing or performance.
The operational pattern is more useful than the headline number. SpaceXAI says it started with a crawl, walk, run rollout: connect a small set of tools, restrict the Bot to internal notes, require human approval for every write, add traces and evaluations, and then expand from simple tickets to direct customer responses. The case study says Grok Bot can pre-investigate tickets, connect Plain, Linear and Datadog, reproduce an issue with video, monitor queue signals, and summarize recurring product feedback. It also describes refund automation and weekly quality summaries. None of these examples removes the need for access controls, idempotency, escalation and human review.
Who should use this pattern
Support leaders with repetitive, well-instrumented workflows may use the case study as a rollout template: begin with read-only investigation, define an approval boundary, log traces, evaluate accuracy and tone, and expand only when failure recovery is visible. It is not a fit for regulated support, irreversible refunds, production credentials, or sensitive customer data until retention, connector scope, audit access, and rollback behavior are documented for the actual plan.
Our assessment and decision rule
Grok Bot is becoming an operations layer under Grok, not merely a faster chat mode. Choose it when the work is repetitive, reversible, and measurable. Choose a conventional help-desk automation or a human queue when you cannot grant a cloud agent access to the ticketing, issue-tracking, monitoring, or billing systems. Start with a disposable queue, require approval for every write, and compare cost and resolution quality against a human-reviewed baseline. We have not independently tested the support workflow or verified SpaceXAI’s reported metrics.
Read the official SpaceXAI case study, then review our Grok profile, ChatGPT vs Grok comparison, and AI privacy checklist. The linked case study supports the product facts and vendor claims; our rollout advice and limitations are editorial synthesis, not independent testing.
September 28, 2026: Grok Team Bots add shared team context
SpaceXAI announced Team Bots as a public beta for Teams and Enterprise plans. The company describes a shared Grok Bot organized around a team role or workflow: members can use the same bot expertise, files, instructions, and skills, while the announcement says individual conversations and memories remain separate for each user. Team Bots can connect to applications such as Salesforce, Notion, and GitHub through plugins; teams may configure a connection for the group or let each person connect it. The announcement also describes API credentials for services without a plugin and access through a bot handle in Slack.
SpaceXAI’s examples include account briefings, engineering triage, marketing review, and read-only analytics from approved warehouse tables. These are examples reported by the company, not independent customer evaluations by AI Tools Manual. The announcement’s distinction between shared bot expertise and private individual conversations is important, but it should be verified against current workspace settings and privacy documentation before a business uses sensitive material.
Who should evaluate Team Bots—and the controls to check
Best fit: teams with recurring work that depends on shared instructions and approved tools. It is not a safe default for a team that cannot administer plugin scopes, credentials, and membership, or that expects a beta feature to provide mature audit and governance controls. Our assessment is to start with a non-sensitive workflow, connect only the minimum required data, keep destructive actions behind approval, and confirm how shared files, team skills, per-user memories, and Slack conversations are separated.
Because this release introduces shared context and third-party connections, verify the current enterprise terms, retention controls, connector permissions, and account eligibility before moving confidential or regulated data into a Bot. See our ChatGPT review for an alternative assistant workflow. Evidence note: this section summarizes SpaceXAI’s Team Bots announcement; no security audit or hands-on test is claimed. Next review: October 15, 2026.
October 2026 Grok API and Grok Bot changes
xAI’s developer release notes say grok-voice-transcribe-1.0 reached end of life on October 2, 2026. Requests using that model slug are routed to grok-voice-transcribe-2.0 at the same price, which xAI describes as more accurate. This is an API migration detail for developers; it should not be confused with a change to the consumer Grok chat experience. Teams with pinned model IDs should update their configuration and regression-test transcripts, timestamps, and downstream parsing even though xAI says the old slug is routed automatically.
The official Grok Bot changelog also lists October 2 desktop-app changes: the marketplace was renamed Connect Apps, connector cards show the connected Google account, and approval cards explain when a rule paused or allowed an action. An earlier October 2 entry says a primary Bot can proactively offer work. These changes apply to Grok Bot and connected-app workflows, not necessarily to every Grok plan or API endpoint. They make permission review and accurate account selection more visible, but do not remove the need to inspect actions before allowing an agent to send or alter information.
Who should act: developers using the retired transcription slug should verify their request configuration and test representative audio before relying on the route. Grok Bot users should check which account is connected and review approval rules. Users who do not use the voice API or desktop Bot do not need to change their workflow on the basis of these notes.
Our assessment: the voice-model transition appears designed to preserve existing requests, but automatic routing is not a substitute for application-level validation. The Bot’s clearer connector and approval messaging is operationally useful; it is not a guarantee that every proposed action is correct. For a different general-purpose assistant, see our ChatGPT review. Evidence note: the endpoint lifecycle and Bot behavior above come from xAI’s API release notes and Grok Bot changelog; no independent API test was performed. Next review: October 25, 2026.
October 12–14, 2026: X Subscription bundles Grok, while Grok Voice 3.0 reaches the API
SpaceXAI says X Subscription combines Grok, Grok Bot, Cursor and X Premium. Its official plan table lists Free at $0, Plus at $8/month, Premium at $30/month, Super at $100/month and Ultra at $200/month. Paid plans share Grok-model usage across Grok apps, X and Cursor; included usage is listed as 4x, 15x and 50x the Plus allowance for Premium, Super and Ultra. API usage is billed separately at API list prices through top-ups. Confirm country, tax, eligibility and current terms before comparing costs. The bundle changes how users may compare separate subscriptions, but it is not a like-for-like replacement for every standalone plan. Read the official X Subscription announcement and confirm current country, tax, cancellation and entitlement details before making a cost comparison.
SpaceXAI also announced Grok Voice 3.0 for the Voice Agent API at $0.08/minute, the same stated rate as Voice 2.0. The company says `grok-voice-latest` was switched automatically and that Voice 3.0 powers Grok Bot voice mode. Those are material model and billing changes for API callers; details are in the official Grok Voice 3.0 post and its linked API documentation.
Our assessment: the bundle may simplify access for people already paying separately for X, Grok or Cursor, while API users need to distinguish per-minute voice billing from consumer plan allowances. Check the exact included usage, overage behavior and whether your existing account is on a compatible tier. We have not tested voice quality or independently validated SpaceXAI’s performance claims. If the task is persistent agent execution rather than chat, compare Grok Bot with Manus. Next review: November 16, 2026.
Human review required: re-check all plan prices, shared usage allowances, rollout and account eligibility, API per-minute price, automatic model migration and any billing implications before publication.