The week of September 13, 2026 delivered a mixed bag for xAI watchers. Grok 4.7 slipped again — Musk’s September 12 target came and went with “needs a few more days to cook.” But the platform news was anything but stalled. Grok models landed inside Microsoft Copilot. Grok Bot shipped sales connectors for Salesforce, HubSpot, and Gong. Native iPad and Android apps launched. A 72-hour livestreamed startup build called Grok Bot Galaxy is set to prove the agent platform can run an entire company. And the $60 billion Cursor acquisition — now fully closed — continues to reshape the competitive landscape. Here is the xAI Weekly for September 13, 2026.
Grok 4.7: Slipped Again, and the Pattern Is the Story
September 12 was the target. September 12 came and went. Grok 4.7 is not here.
On September 11, Musk posted on X that the model “needs a few more days to cook.” The issue, according to Musk, is that reinforcement learning training penalized response length too heavily, causing the model to give up prematurely on hard problems and skip rigorous self-verification. He framed it as an RL tuning problem — not a hardware limitation or missing feature.
The Timeline of Slippage
| Date | Musk’s claim | Implied target |
|---|---|---|
| Jul 24 | “Grok 4.7 in 4 weeks” | ~Aug 21 |
| Jul 28 | “A few weeks” after 4.6 | Vague |
| Aug 12 | “Ready in 3–4 weeks” | ~Sep 2–9 |
| Sep 2 | “Comes out in 10 days” | Sep 12 |
| Sep 11 | “Needs a few more days to cook” | No new date |
This is the fifth missed target. Grok 4.5 slipped 46 days (Musk said “hopefully late May,” shipped July 16). Grok 4.6 slipped 5 days (“2 weeks” on July 24, shipped August 12). Grok 4.7 has now been “a few weeks away” for over seven weeks.
What Remains Unknown
Nothing about Grok 4.7 appears in official xAI documentation. No model ID. No pricing. No context window specification. No benchmark card. The xAI developer docs (docs.x.ai) still list Grok 4.6 as the current flagship, with Grok 4.5 as the coding/agentic alternative. Everything about 4.7 traces to Musk’s X posts — a pattern that makes independent verification difficult.
What We Do Know (From Musk, Unconfirmed)
- Parameters: ~2.1 trillion — 40% above Grok 4.6’s 1.5T
- Training data: SpaceX engineering records, telemetry, Starlink satellite data
- Pre-training: Completed August 12
- RL issue: Response length penalty too aggressive, causing premature task abandonment and insufficient self-checking
Current xAI Model Lineup (Official)
Grok 4.6 remains the flagship at $2/M input and $6/M output tokens with a 500K context window. Grok 4.5 serves the coding and agentic use case at the same price point. Grok 4 Fast handles the free tier. Grok 4.7 does not appear anywhere.
The lesson here is structural: when your entire product communications strategy runs through one person’s social media posts, timeline slippage isn’t a bug — it’s the platform. Enterprises evaluating Grok 4.7 for production workloads should plan around official documentation, not X posts.
Grok Models Land in Microsoft Copilot
On September 12, Satya Nadella announced that Grok models are now selectable inside Microsoft Copilot — the general-purpose assistant embedded across Microsoft 365. This is not GitHub Copilot (which already integrated Grok 4.6 on August 14). This is the main Copilot surface, covering Word, Excel, and PowerPoint across web and desktop.
How It Works
The rollout is gated through Microsoft’s Frontier Program — a focused release to participating customers rather than general availability. Key constraints:
- Disabled by default — tenant admins must manually enable Grok models in the Microsoft 365 Admin Center
- EU/EFTA/UK excluded from the initial preview due to data protection and regulatory review
- Models hosted outside Microsoft infrastructure, governed by xAI’s Enterprise Terms
- Customer prompts and data are isolated within tenant boundaries and not retained or used to train xAI models
The Full Microsoft Integration Map
This is not a one-off. The Microsoft-xAI integration has been building steadily:
| Date | Integration |
|---|---|
| Feb 19, 2026 | Grok 4.1 Fast in Copilot Studio (US preview) |
| May 2026 | Connectors for Outlook Mail, Calendar, Teams, SharePoint, OneDrive |
| Jun 16 | Grok add-in for PowerPoint (free for M365 subscribers) |
| Jun 18 | Grok add-in for Word (free) |
| Jul 20 | Grok add-in for Excel (free) |
| Jul 21 | Grok add-in for Outlook (paid X/SuperGrok only) |
| Aug 14 | Grok 4.6 in GitHub Copilot (8 developer surfaces) |
| Sep 12 | Grok models in Microsoft Copilot (general assistant) |
The trajectory is clear: Grok is becoming the non-OpenAI alternative inside Microsoft’s AI stack. For enterprises that want model diversity without leaving the M365 ecosystem, this is a meaningful option — especially given the tenant-level data isolation and admin controls.
Grok Bot Microsoft 365 Plugin (September 1)
Separately, the Grok Bot Microsoft 365 plugin shipped on September 1 with deeper capabilities than the Copilot integration:
- Outlook: Read, write, search, draft, send, reply, forward, and organize emails
- Calendar: Look up events, check availability, create and modify meetings, accept or decline on behalf of the user
- OneDrive: Browse folders, read documents, upload Grok-generated content
- Permissions: Mail.ReadWrite, Mail.Send, offline_access
- Availability: Grok Business and Enterprise accounts only
- Operation: 24/7 cloud-based — no local device or server required
The Grok Bot plugin is more aggressive than the Copilot integration — it can take actions on your behalf, not just answer questions. That makes it more powerful but also demands more governance attention.
Grok Bot Platform: The Agent Stack Matures
While Grok 4.7 stalled, the Grok Bot platform shipped at a remarkable pace. Two distinct waves landed this week.
Quality-of-Life Package (September 9–10)
- Inline message drafting — compose messages for approval before sending, directly in chat
- Form filling and logins — Bot can fill forms and handle logins with password manager support (1Password and similar)
- Performance improvements: 10–35% more effective usage from a more efficient harness (average 10%, heavy users up to 35%)
- Desktop app: Cold start 247ms faster (18%), LCP 425ms faster (23%), FCP 234ms faster (16%)
- One-click account switching between Grok Bot accounts
- Native iPad app launched
- Native Android app launched
- 20+ language support
- File sharing directly with Bot on mobile
Sales Connectors (September 10–11)
New integrations with the biggest names in go-to-market tooling:
- Salesforce — CRM integration for pipeline and contact management
- HubSpot — Marketing and sales automation
- Gong — Revenue intelligence and call analysis
- Clay — Lead enrichment and data operations
- Granola — Meeting notes and knowledge management
xAI went further: the company shared its own sales team’s Bots as installable templates — pre-configured with context, connectors, and routines. Musk amplified the announcement on September 11. This is a distribution play: instead of expecting every customer to configure their own sales agent from scratch, xAI provides battle-tested templates ready to install.
Why This Matters
The sales connector wave is the clearest signal yet that xAI sees Grok Bot as a horizontal agent platform — not just a developer tool. Salesforce, HubSpot, and Gong collectively cover a huge percentage of B2B revenue operations. Making Grok Bot interoperable with these systems, and providing pre-built templates from xAI’s own sales org, lowers the activation energy for enterprise adoption dramatically.
Grok Bot Galaxy: 72 Hours to Build a Company
On September 15–17, three SpaceXAI staffers will attempt to build a company from scratch using Grok Bot as their primary worker — live, no script, 72 hours. The event is called Grok Bot Galaxy.
The Format
- Participants: Matt Palmer, Lauren Tan, and Roshan Sadanani (all SpaceXAI employees)
- Format: In-person at The Howard, 661 Howard Street, San Francisco
- Hours: Daily 8:30 AM – 6:00 PM Pacific
- Livestream: Free globally via Luma registration — no Grok Bot subscription required to watch
The Schedule
| Day | Focus Areas |
|---|---|
| Tue Sep 15 | Grok Bot 101, Engineering, Product Managers, Founders |
| Wed Sep 16 | Sales Engineering, Sales, SDRs, Customer Support |
| Thu Sep 17 | Marketing Operations, Post-Sales, Marketing, Final Showcase |
Why It Matters
This is the most public real-world stress test of agentic AI to date. The timing is deliberate — it lands three weeks after the $60 billion Cursor acquisition closed and days after Grok Bot’s sales connector wave. xAI is making a clear argument: Grok Bot isn’t just a coding assistant or a chatbot. It’s a platform that can run multi-department work across engineering, sales, marketing, and operations.
The free livestream is an audience-building play. Expect demos, failures, and real-time problem-solving. If Grok Bot can handle sales, engineering, and marketing work live for 72 hours, it changes the conversation about what AI agent platforms are capable of — from theoretical capability to visible execution.
The $60B Cursor Acquisition: Context for the Week
The SpaceX acquisition of Anysphere (maker of Cursor) provides the backdrop for everything this week.
Deal Facts
- Acquirer: SpaceX (NASDAQ: SPCX, publicly traded since June 12, 2026)
- Target: Anysphere, Inc. (Cursor)
- Value: $60 billion, all-stock
- Announced: June 16, 2026 | Closed: August 14, 2026
- Structure: 389,289,254 SpaceX Class A shares; Cursor cofounders each received ~$2.4B in stock
- Largest acquisition of a venture-backed startup on record
- Cursor annualized revenue: ~$4B at close (~15x revenue multiple)
- Antitrust termination fee: $40 billion
The OpenAI Response
OpenAI announced it will stop supplying models to Cursor on November 12, 2026. Cursor was one of OpenAI’s five largest customers, representing over $1B/year in potential revenue. OpenAI framed the decision around trust concerns with Musk’s companies. The practical effect: Cursor users relying on OpenAI models need to migrate to Grok or other alternatives by mid-November.
The Combined Entity
The corporate chain: Cursor → Anysphere → SpaceXAI (AI division) → SpaceX (parent). Grok Bot was the first product born from the merger, going live days after the August 14 close. Grok 4.6 was partially fine-tuned on Cursor’s developer session data — a training corpus that no competitor can replicate, built from billions of lines of real developer code interactions.
Grok 4.6 Biosecurity Win and Grok 4.5’s Benchmark Position
Two resurfaced stories this week add context to xAI’s competitive positioning.
Grok 4.6 Biosecurity Benchmark (September 1)
xAI highlighted LatchBio’s independent BioSecBench-Refusal analysis, where Grok 4.6 was the only model above 50% on both metrics:
- ~59.2% refusal rate for disguised hazardous biology requests
- ~64.8% completion rate for routine biological work
- ~62.1% harmonic mean — balancing safety with usefulness
This is a meaningful differentiator. Models that refuse everything are safe but useless; models that comply with everything are useful but dangerous. Grok 4.6’s balance, driven by model reasoning rather than API-level filters, generalizes better to novel inputs.
Grok 4.5 on τ³-Banking
Grok 4.5 (launched July 8) ranked first on the τ³-Banking agentic tool-use benchmark at 33%, ahead of GPT-5.5 and Claude Sonnet 4.6. It sits 4th overall on the Artificial Analysis Intelligence Index (54/168 models). At $2/$6 per million tokens with a 500K context window, it remains the value play in xAI’s lineup for coding and agentic workloads.
Tesla + Grok: Voice in the Driver’s Seat
Tesla’s Summer 2026 Software Update (version 2026.26) continues rolling out with Grok Voice Think Fast 2.0 as a genuine vehicle controller. Grok can adjust climate, toggle modes (Dog, Camp, Bioweapon Defense), open the glovebox, fold mirrors, control wipers, search for Superchargers, navigate, and more — all by voice.
The next milestone: voice input reaching FSD’s planning layer within approximately three months. This would be the first time speech directly influences driving behavior. A conversational “supervisor mode” for robotaxi is in development — imagine saying “drop me at the entrance, then go park” and having the car comply.
The broader physical AI ecosystem continues to expand: Optimus humanoid robots use the same Grok models, Starlink handles 15,000+ inbound support calls daily with Grok Voice, and the Cybercab robotaxi launched September 3 in Austin.
What to Watch
- Grok 4.7 launch: No new date. Watch for official xAI documentation — model ID, pricing, context window, benchmarks. Until it appears on docs.x.ai, treat any timeline as aspirational.
- Grok Bot Galaxy (Sep 15–17): The 72-hour livestream is the most public agentic AI test to date. Watch for what works, what breaks, and what surprises.
- Microsoft Copilot rollout: The Frontier Program is gated. Watch for general availability timing and EU regulatory progress.
- OpenAI cutoff (Nov 12): Cursor users on OpenAI models need migration plans. Watch for migration tooling and whether Grok captures the displaced demand.
- Grok 5: End-of-2026 target per Musk’s monthly cadence promise. No details yet.
That is this week’s xAI Weekly. The strategic picture is sharpening: SpaceXAI is executing on platform integration while the flagship model slips. Grok models inside Microsoft Copilot, sales connectors for the biggest GTM tools, and a 72-hour livestreamed company build are all distribution and adoption plays. The $60B Cursor acquisition is bearing fruit — Grok Bot is the first product of the merger, and the OpenAI cutoff on November 12 creates a forced migration event. Meanwhile, Grok 4.7’s repeated delays are a reminder that model development doesn’t follow social media timelines.
If your organization is evaluating Grok for enterprise workloads, planning a migration from OpenAI models on Cursor, or exploring Grok Bot for sales and operations automation, connect with Kevin on X or book a strategy session with Big Hat Group — we help enterprise teams navigate AI decisions with confidence.
Check back next week for the latest.