Keep Working on the Plan

The first plan I saw in Cursor was a numbered list. It was not code. It was what the agent intended to do, file by file, step by step, and then it waited. I could edit it. I could argue with it. I could leave it overnight and come back to it in the morning. Nothing ran until I said so.

I started where a lot of people did. Tab completion first, then GitHub Copilot in VS Code. Autocomplete was useful. It was not a way of thinking. The plan was.

I am cancelling Cursor. I paid for two things: a plan I could keep working on before anything executed, and the freedom to hand that plan to the best model for the job. Plan mode is still there, and it is still my favorite thing in any AI tool. The choice is being taken apart so Cursor can be a Grok harness, and a plan is only as good as the model that executes it.

The Plan Before the Code

My brain does not hold a long invisible sequence well. If the steps live only in my head, some of them fall out, and the anxiety fills the space where they were. I wrote about that move in Filling in the Blanks: the blank is intolerable, so something fills it. AI fills it toward the average. My brain fills it toward catastrophe.

A plan on the screen takes the blank away. I can see the whole shape of the work before any of it happens. I can catch the step where the model assumed something about my repo that is not true. I can cut a step, add a constraint, ask why. The plan is where I do my thinking, and the model does the typing after I have finished. That is the difference between a tool I use and a tool that uses me.

Then the second half. Once the plan was right, I could pick who executed it. Opus for the hard debugging. A cheap, fast model for the boilerplate. Composer for everyday work, when it was the best workhorse in the building. And when I did not want to decide, Auto mode decided for me, and it was brilliant at it. One decision I do not have to make is a real accommodation. Nothing else on the market put those pieces in one window.

What Changed

SpaceX agreed to buy Anysphere, the company behind Cursor, on June 16, 2026, in an all-stock deal that valued Cursor at $60 billion. The merger closed August 14. SpaceX already owned xAI, the lab that makes Grok. Cursor announced the same day that Grok 4.6, “which we released Wednesday,” was “an early look at what we can now build together.”

Here is what it looks like from my chair.

New chats open on Grok 4.7. Users on the Cursor forum have reported for weeks that new agents reset to Grok even when they configured something else. Cursor staff explained that this is intentional when the default is set to “Cursor Default,” which follows whatever model Cursor is promoting. One user put it plainly: “A default that ignores my usage history and moves new agents to Cursor’s currently promoted model is not neutral.”

The new plans are built around Grok. The X Subscription launched October 12 and bundles Grok, Cursor, and X Premium. Cursor’s own pricing page says the $8 Plus tier includes Grok models only. On Premium and above, third-party models such as Claude are still available, but at API list price from prepaid credit, with no included allowance. The included pool covers Grok and Composer. Existing subscribers can keep their old plans for now.

OpenAI is leaving. On August 28, OpenAI notified SpaceX that it would wind down its contract with Cursor, with a proposed shutoff of November 12, and that it would not provide future models. The reason was in the second paragraph: “we cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk’s companies violating contracts.”

Composer has not had a new version since 2.5 shipped in May. Grok got three in the same stretch. I do not know what changed around Composer. What I know is my own sessions: the same mistakes over and over, no matter what the prompt says or how specific the rules file is.

And Grok. My experience vibe coding is more frustrating now than it has ever been. Simple mistakes. Instructions read and then ignored. Work I have to check line by line, which defeats the point of a plan.

That is not only my mood. Cursor runs its own benchmark, CursorBench, built from real Cursor sessions. Grok 4.7 at its highest effort setting scores 46.3 percent and costs $6.01 per task. Claude Opus 5.5 at High scores 56.0 percent for $3.97. Grok is cheaper per token. It is not cheaper per finished task, and it lands nineteenth on the board its own owner publishes. Grok 4.7 does top a few leaderboards, including Artificial Analysis’s cyber index and an open benchmark about rebuilding retired programs. I am not doing either of those. I am asking it to follow the rules in my repo.

Elon Musk does not dispute any of this. In an interview with China Media Group this month, he called Grok 4.7 “a solid workhorse of a model” and said, “It’s not as good as, say, Opus 5.5 that just got released.” Then: “I would expect that we probably catch up to Frontier sometime next year, most likely.”

Next year. So why force it on everyone now? If the model will be competitive in a year, wait a year. Let Grok earn the default. Instead, the product got worse on purpose, today, for a model its owner admits is behind.

Here is the part I cannot get past. On October 7, Musk posted about Grok Bot, SpaceX’s own agent product: “Going forward, @SpaceX will use the best back end model for any given task, including Claude Opus 5.5, MidJourney, Suno and other leading APIs. Whatever is most likely to give you the best outcome.”

His product gets the best model. The product he bought gets Grok.

Yet another business killed by Elon. I made the longer case about the distance between what he says and what he does in He Can’t Have It Both Ways; this is one more entry on that ledger.

Rent Seeking

Economists have a term for this: rent seeking, capturing value through control of something rather than by making something better.

That is what it looks like to me. People chose Cursor because it let them pick the model. Musk bought it and used ownership of the harness to steer those people toward his own model. It spends Cursor’s goodwill to pad Grok’s usage numbers. The honest version would be the slow one: make Grok good enough that Cursor users pick it on their own, the way they picked Composer. That would benefit Cursor. This benefits xAI.

I cannot see inside the deal, and I am not going to claim I know why he bought it. I can see the steering, I can read the benchmark, and I can read Musk saying his own model is behind.

Musk has also told people how to judge companies. On May 1, 2022, he gave his followers investing advice: “Buy stock in several companies that make products & services that you believe in. Only sell if you think their products & services are trending worse.”

In this case, at least, he cannot claim he makes a good product. He bought one and made it worse. By his own rule, Cursor is a sell. Cancelling is how a subscriber sells.

What I’m Trying Next

I have been inside the Cursor ecosystem so long that I stopped looking around. That is the good news here. I get to look.

I have not tried most of these yet, so this is a list of candidates and one test, not reviews.

Devin. It is installed. I have barely touched it. From what I can tell, it is the closest thing to Cursor.
Claude Code. Anthropic’s own agent, with Anthropic’s newest models and no middleman deciding which ones I can afford.
OpenAI Codex. Same idea from the other lab. OpenAI’s own help page points Cursor users toward the Codex extension after November 12.
OpenCode. Open source, and built to let you bring the model.
My own VS Code extension. I have thrown this idea around more than once. If nothing gives me the plan the way I need it, maybe I build the plan.

The test for each one is the same two questions. Can I see the plan and keep working on it before anything executes? Can I choose the model? Anything that answers yes to both gets a real trial. Anything that answers yes to one gets a fair look.

That Cursor

I loved Cursor. That Cursor is dead, and I do not see myself coming back. Good luck to the team. They built the tool that made AI make sense to me, and I do not feel good about where it is going.

The last plan I wrote in Cursor is still there, unexecuted, waiting for me to say go. The model picker above it says Grok 4.7. I did not pick it.