The frontier race is no longer “who has the biggest base model.” It is who owns the loop: model, editor, persistent computer, and the right to keep working after you close the laptop.

A few facts. Then the bet.

What is actually on the board

xAI shipped Grok 4.6 on August 12. Same-day availability on the API, Cursor, Grok Build, and partner gateways. Context window: 500k. List price: $2 / $6 per million tokens. On Artificial Analysis it lands around Sol, not above Fable.

Grok 4.7 is not out. On August 12 Musk said initial training was done, supplemental training on “a massive amount of SpaceX company data” was starting, and the model should be ready in three to four weeks. Earlier he described 4.7 as a 2.1T step up from the 1.5T 4.6 line. Treat the date as a founder target. Treat the direction as real: more engineering data, more agent runtime, not another chatbot skin.

OpenAI has not shipped Astra yet. After internal evals in early August, the company said it cannot rule out that Astra meets the Critical cybersecurity threshold in its own Preparedness Framework: functional zero-days on hardened systems, or an end-to-end attack plan from a high-level goal, without a human in the loop. A significant number of Astra and cyber workloads were paused or moved behind harder isolation, monitoring, and alignment. That is not marketing. That is a lab hitting its own red line.

Then OpenAI split Daybreak into Blue and Red. Blue gives approved defenders a frontier general-purpose model, GPT-5.6 Sol, with safeguards tuned for defensive work. Red unlocks purpose-trained cyber models, including GPT-5.6-Cyber, for authorized vuln research, exploit validation, and red teaming. Official story: arm defenders before attackers get the same stack. The practical pattern already exists. Operators with Daybreak access run it as a subagent under Sol. The public model stays inside the consumer guardrails. The gated body does the dual-use work. That is how you get the full stack without shipping the raw thing to everyone. If Astra ships the same way, Daybreak is the licensed organ under the app, not a side demo.

SpaceX closed the Cursor acquisition on August 14. Sixty billion, all stock. Cursor is no longer an independent IDE that happens to route Grok. It sits in the same house as Grok, Grok Build, Grok Bot, and Colossus.

OpenAI answered two weeks later. On August 28 it notified SpaceX it will wind down the contract that feeds OpenAI models into Cursor, proposed cutoff November 12, the maximum notice in the change-of-control clause. Stated reason: trust. OpenAI says it cannot be confident SpaceX will stay inside the terms of service, and it will not put Astra into that stack. Translation without the press-release polish: they do not want a rival with a giant GPU fleet and a coding product sitting on top of their next model. Existing GPT-5.6 family models stay until the cutoff. Astra does not go in. Michael Truell said OpenAI is about 5 percent of Cursor traffic. The number is small. The principle is not.

This is not abstract. In April, under oath, Musk said xAI had “partly” used OpenAI models in distillation and validation, and called it standard practice. Labs call this “standard practice” when they do it and “theft” when someone else does. Cutting Cursor is what a lab does when the student buys the classroom.

Anthropic did the opposite. Tom Brown said Cursor has been a partner since Sonnet 3.5 and that Anthropic will increase compute to keep Claude inside Cursor. That sounds generous until you notice the wiring. In May Anthropic signed a compute deal to run Claude on Colossus. The landlord owns the editor. The tenant supplies the model that still makes the editor useful. If you want those hours, you stay in the product that now belongs to the cluster owner. That is not “Anthropic is being forced to distill Opus 5.1.” It is worse: the incentive to remain inspectable and deeply integrated is structural.

Meanwhile the consumer surface is a mess. X: Basic, Premium, Premium+. Grok: Lite, SuperGrok, Plus, Heavy. Cursor: Start, Pro, Pro+, Ultra. Grok Bot bundled into some of those and not others. On August 30 Nima Owji posted the obvious complaint: too many overlapping plans for one empire, and asked for a single SuperX. Jacob Witt, now building at SpaceXAI, replied “Working on it.” In a separate thread the same night he wrote: “Agree this is too confusing. Working on cleaning up all the different subscriptions and paths.” When a company starts saying that out loud, the product map is already decided. The billing map is late.

The missing piece is not another model. It is a computer that does not die when you close the lid.

Thibault “Tibo” Sottiaux has been blunt about this. Laptops were designed for one human, a few apps, and a typing speed. An agent that keeps a hundred contexts open, spawns workers for days, compiles, tests, browses, and retries is not a local chat window. It is a cloud machine. Light local command. Heavy cloud execution.

Grok Bot is that experiment in public: early beta from August 11, a persistent cloud computer with browser, filesystem, and terminal; named bots that keep working after you leave. One detail the launch copy often gets wrong: the machine is per account, not per bot. All of your bots share the same VM. Cursor Cloud Agents are the coding-shaped version of the same idea: isolated VMs that clone a repo and come back with a PR. Different skins. Same physics.

If Astra is the long-horizon swarm OpenAI wants, it will need exactly those VMs. Grok Bot looks less like a finished consumer product and more like a field test of the runtime the whole industry is being forced into.

What I think happens next

Codex-app is already a complete product. Tibo just said ChatGPT Work and Codex together hit 25 million active users. Editor, agents, review, GitHub surface, computer use. That stack exists. Astra is the missing organ. If it arrives with a persistent VM of the same class Grok Bot is testing now, Codex-app is not “a coding IDE with a strong model.” It is the full monster: backend that does not sleep, UI work in the same loop, and cyber on top if you are cleared to run Daybreak as a subagent under the same roof, the way Sol already does it today.

That is the object a Grok-App has to clone. Not “Codex but friendlier.” The future Codex-app monster, whole.

A Grok-App is the logical answer. Cursor’s desktop (files, terminal, inline edits, repo) plus Grok Bot’s VM (the job keeps running when the laptop sleeps), local and cloud together, one Grok 4.x family, one subscription. First-generation harvest (three apps, overlapping plans, grok-bot credits gone in two days) dies the moment 4.7 is actually cooked. Keeping the pieces separate after that is irrational. Timing is the open variable. I would not bet the unification on 4.7 day one. I would bet it on the next cycle once the harvest is finished: more Claude 5.* inside Cursor while Anthropic still needs Colossus, then a Grok-only harness when the house model is good enough. And after that I would not be surprised if Musk kicks Anthropic off SpaceX compute, further weakening a competitor.

The personality split is the point. Codex speaks code. Anthropic still owns the careful-colleague register: legal, finance, math, verification. On Vals, Claude Opus 5 leads Legal Research Bench at 55.3 percent. Fable 5 is second at 49.5. Grok 4.6 is third at 48.1, tied with GPT-5.6 Sol on score and much cheaper per test. On Harvey’s Legal Agent Benchmark, Grok 4.6 is third at 15.8 percent all-pass. Legal is no longer the hole. Finance and mathematics still are. A later Opus or Fable point release is the obvious thing to drink from while the window is open. SpaceX data is the thing nobody else can pour on top: engineering logs, real constraints, not another internet crawl. Astra, from the public math results and the rumor cycle, looks strong at research. Whether that holds against a later Grok 4.8 or 4.9 is open. That fight is not 4.6 versus Opus 5 today. It is who owns the overnight computer when both stacks are complete.

OpenAI keeps the monster and the gate: Astra (this week?), Daybreak as the licensed body, Codex-app as the respectable full product. Anthropic keeps converging on corporate and regulated work, where “most aligned” and audit trails beat “ships Friday.” xAI tries to own the individual builder with a clone of that future Codex-app: humanized the Anthropic way, cheaper, fewer refusals, SpaceX data as the unfair layer.

If that Grok-App ships clean, Anthropic does not lose the enterprise. It loses the indie default. The person who today runs Codex-app on one screen and Claude Cowork on the other (Codex as implementing staff, Claude watching documents and drifts) does not pick a tribe. They keep the monster that builds, and they swap the coordinator. Grok-App is that swap: same overnight computer, fewer refusals, SpaceX data on top.

From where I sit the board is set. The labs are no longer competing on leaderboards. They are competing on who is allowed to sit on whose computer overnight.

Sources

[xAI / SpaceXAI, “Introducing Grok 4.6,” August 12, 2026. https://x.ai/news/grok-4-6 @SpaceXAI, )

xAI / SpaceXAI, “Introducing Grok 4.6,” August 12, 2026. https://x.ai/news/grok-4-6 @SpaceXAI,

Elon Musk (@elonmusk) on Grok 4.7 timing and SpaceX data, August 12, 2026.

OpenAI (@OpenAI), “Responding to the next frontier of critical cyber capabilities,” August 7, 2026. https://openai.com/index/responding-next-frontier-critical-cyber-capabilities/

OpenAI (@OpenAI), “Pacing model development in an era of cyber-critical capabilities,” August 18, 2026. https://openai.com/index/pacing-model-development-cyber-capabilities/

OpenAI (@OpenAI), “Expanding Daybreak as the Cyber Defense Window Narrows,” August 10, 2026. https://openai.com/index/expanding-daybreak-as-the-cyber-defense-window-narrows/

OpenAI (@OpenAI), Daybreak. https://openai.com/daybreak/

Anthony Ha, TechCrunch (@TechCrunch), “SpaceX officially closes its Cursor acquisition,” August 15, 2026. https://techcrunch.com/2026/08/15/spacex-officially-closes-its-cursor-acquisition/

OpenAI (@OpenAI), “Our decision on Cursor following its acquisition by SpaceX,” August 28, 2026. https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/

Maxwell Zeff, Wired (@WIRED), “Elon Musk Seemingly Admits xAI Has Used OpenAI’s Models to Train Its Own,” April 30, 2026. https://www.wired.com/story/elon-musk-distill-openai-models-partly-xai/

Anthropic (@AnthropicAI), “Higher usage limits for Claude and a compute deal with SpaceX,” May 6, 2026. https://www.anthropic.com/news/higher-limits-spacex

Tom Brown (@NotTomBrown) on keeping Claude in Cursor after the SpaceX acquisition, August 29, 2026.

Nima Owji (@nima_owji) on overlapping SpaceXAI subscriptions, August 30, 2026.

Jacob Witt (@jacobwitt6), “Working on it,” August 31, 2026.

Jacob Witt (@jacobwitt6) on cleaning up subscriptions, August 31, 2026.

Grok Bot early beta, August 11, 2026. https://x.ai/bot @bot

[Vals AI (@ValsAI), Grok 4.6 on Legal Research Bench (48.1%, #3 of 49) and Harvey Legal Agent Benchmark (15.8%). https://www.vals.ai/benchmarks

)

Vals AI (@ValsAI), Grok 4.6 on Legal Research Bench (48.1%, #3 of 49) and Harvey Legal Agent Benchmark (15.8%). https://www.vals.ai/benchmarks

Tibo Sottiaux (@thsottiaux) on local command and cloud execution for agents, interview with Matthew Berman (@MatthewBerman), August 2026.

Tibo Sottiaux (@thsottiaux), 25 million active users celebration and usage reset, August 31, 2026.

"Astra, Cursor, and the coming (?) Grok-App

Job breakdowns

Job breakdowns

Command Menu