Cursor is already kind of useless for third party models unless you're willing to spend thousands of dollars. It's only worth it if you're going to use mostly grok/composer.

Shit, for most things, ive got a dev agent that reads change requests from a GDoc and communicates through email with me. I can develop in my mobile

Big reason we built https://boxes.dev around the model harnesses (Codex + Claude Code), so you can bring your own subscriptions.

On the other hand, subscriptions create lock-in in a way that API pricing doesn't.

I think a more likely end is that subscription value decreases over time because API pricing gets more reasonable, but subscriptions stay because they are a good way of getting money out of people consistently.

This is why "subsidized tokens" is possibly a misnomer. Money at lower variance is worth more than the same money at higher variance. Not "subsidy" so much as reducing risk and passing some of that to a consumer.

How long have you been in tech, outta curiosity?

What do you mean by that? If you're being sassy about a downvote then I'll remind you that when you reply to a comment, that person can't downvote your reply. Also this is an account with barely any comments or points, I'm pretty sure they can't downvote at all.

Just a stub of a conversation.

I don’t care about the points. Just the chat, which seems finished.

But it didn't come from the person you were talking to, so editing on a sarcastic "great answer" doesn't really work.

The person you were talking to just hadn't responded within the first hour. They still might.

Honestly, _this_ was their moat more than Composer to me.

Why? Written by does not mean designed by etc. There's a lot more to it.

I dont know how cursor works. Everytime ive tried to use, its super buggy and has memory leaks that make my very quite PC sound like a jet engine.

It is for 3rd party models (which you get double the amount of your sub price as usage).

Of course they wont charge you extra for BYOK. how would they???

> On Teams and Enterprise plans, third-party model requests include a Cursor Token Rate of $0.25 per million tokens. This rate applies on top of model API pricing for included usage, on-demand usage, and BYOK usage.

See the link I posted above or also this one: https://cursor.com/help/models-and-usage/token-rate

Additionally helpful where 2 apps/repos might require running at the same time, e.g. headless web apps, or, for plugin development where the plugin might be a dedicated repo but needs to run in another app to observe changes and make it re-test itself.

I also switched from terminal to the app not that long ago, I don't find it buggy, it has access to a browser which is really helpful.

However, what I see cooking at Cursor is way more promising than the others. Things like full system understanding, their own forge, multi-repo support. I'd put them as way more visionary in the Gartner magic Quadrant ;)

For the forges, what the sibling said.

In particular Haiku 4.5 is rubbish, Anthropic don’t have anything in the cheap/fast part of the market.

For people fortunate enough to still be on a subscription instead of per-token billing this is less relevant, but their time will come.

The cursor model is quite nice because it allows you to switch between cheap and expensive models for different tasks

Now, I’m no expert, because I was late to the game and have only ever used Pi. But I guess Cursor is some product that’s tied to your IDE? If so, then yeah. That’s just too restrictive. I still love my IDE, but I don’t want it to be my harness too.

Claude Code works with models from other providers too. Anthropic supports this. You can configure some Claude Code environment variables to switch: eg changing ANTHROPIC_DEFAULT_HAIKU_MODEL to point to GLM Flash or Luna, setting ANTHROPIC_BASE_URL to point to api.z.ai, and making ANTHROPIC_AUTH_TOKEN the API key for your alternative provider instead.

Some instructions here:

https://docs.z.ai/scenario-example/develop-tools/claude

That said, I've not actually tried this myself, opting to build my own harness instead. And you don't know what information Claude Code might be sending back to Anthropic about how you use competing models and which models you use. I don't know for sure that they do this, but after hearing about how they used steganography in the date of harness system prompts to identify the user's location, I don't entirely trust Claude Code anymore.

Why do you say that? I've used Claude Code with DeepSeek via OpenRouter just fine.

I basically never use the editor but the fact that there's a review UI for all the agent work is incredibly helpful.

For me, that makes the models much more usable. I've also been using GPT models a lot, as they're cheaper and less vomit inducing than Claudes text, so this is definitely bad news for me.

I think it's possible to have a better TUI, with something that isn't modeled after a shell prompt. Apparently, recent updates move in that direction. The GUI version of Claude Code seems to be coming to Linux, too.

Inference is cheap; Dax once said on a podcast that Opencode has a close to 90% profit margin on the openweight models it provides inference for (at 10x cheaper pricing). Those models are close in size and spec to models from large labs.

If either one of them stops doing that, customers will stop using them, because other models will become better.

In other words, there is no point in talking about inference cost in isolation.

When they outsource training to China they will be so profitable!

So this is OpenAI following suit after Musk admitted to distilling their models.

This was bound to happen once Cursor decided to sell itself to a competing model provider.

It'll be interesting to see if Anthropic applies their ban to Cursor, or if that datacenter deal they signed with Musk changes anything.

That thread you linked suggested they were using it with cursor as part of their development process, not for distilling.

>"Hi team, I believe many of you have already discovered that anthropic models are not responding on cursor. According to cursor this is a new policy anthropic is enforcing for all its major competitors."

And a month before the OpenAI block, they blocked Windsurf for the same reason: https://www.forbes.com/sites/johanmoreno/2025/06/05/anthropi...

So its pretty clear the TOS had this clause prior to Jan 2026.

They indiacte their development staff are using it, and they access it via API, then then go on to say all the nefarious things they "could" be doing, stopping before stating they actually are.

Specifically the TOS violation was:

>customers are barred from using the service to “build a competing product or service"

So if you are building a competitive product, that alone is a TOS violation from their perspective.

Which is fine and their prerogative, but we shouldn't cast a dark cloud suggesting other violations were happening without stated evidence.

OpenAI publicly stated they can't "trust that xAI will use it within terms" not even that they've violated, not that there has been a violation from the Cursor team.

https://www.teslarati.com/openai-cites-distrust-of-spacex-in...

>OpenAI said the decision came down to trust, not technology. “We cannot be confident that SpaceX will use our technology within our terms of service"

It's hard to imagine a 1B / month deal wouldn't.

what's the moat for AI labs like OpenAI and Antrhopic as they seek trillion IPOs. a lot of people said Chinese models also distill US LLM models. if it is this easy to do. how do US AI labs justify asking for trillion?

It will be competitors distilling, or the government getting upset.

It's hard for me to see a different future.

Even if they have Astra or whatever completely to themselves, how do they deal with the -product- side? Or sales, or account management?

They've been focused on juicing model's coding capabilities, but it's absolutely -not- "gen ai" enough to be doing the whole thing in agents, even 3 or 5 years from now, if only because there's so much context we _can't_ give to these models, without reverse centauring ourselves with cameras and mics and oodles of compute everywhere, and so far that hasn't exactly been playing out the way the frontier labs and the singularity folk hoped it would (meta glasses? humane? really?)

[1] If anything, Waymo has a higher chance and global Waymo adoption is probably 20 years into the future, at least.

They're trying. I see Uber robotaxis almost daily.†

I haven't bothered to see if they're still in training or actually taking passengers.

† https://lucidmotors.com/stories/lucid-nuro-uber-partner

I'm substituting my own[0] vibe coding for buying[1] apps. Language mini-games to help with German? A few prompts. Fluid dynamics simulation for an airzooker? Vibed. A web app listening for a MIDI keyboard, upon which you can drop some .midi files, and get a rhythm action game to learn the piano? Vibed. Webcam for my Raspberry Pi? Vibed. Getting Marathon 2 (well, the open sourced and upgraded engine, Aleph One) working as a web app? Vibed. Isochrone maps? Vibed.

Half of this I can even get done with the free models.

> without reverse centauring ourselves with cameras and mics and oodles of compute everywhere, and so far that hasn't exactly been playing out the way the frontier labs and the singularity folk hoped it would (meta glasses? humane? really?)

Yeah, so fortunate that cameras are expensive and there aren't 6 on my table right now between laptops and phones. :P

Seriously though, what's saving humanity collectively from everything getting automated from the panopticon we'd already built before Transformer models got good enough for even the most basic of classification and translation tasks, let alone anything we now use them for, is that machine learning takes an obscene number of examples before getting competent. Any living creature that needed so many examples would starve to death before learning how to eat.

This difficulty is why, for all the billions of miles that Tesla cars have collectively driven, perhaps pushing trillions now, they're still not sold to the public without steering wheels. Tesla claim to make such vehicles now in the form of the Cybercab, but they're not for sale, and even then some of the pictures that get in the press still show steering wheels.

[0] if you can call anything vibe-coded "my own"

[1] or worse, given the popularity of subscription models in this era, leasing some SaaS

When do you have time for all of this? I don’t mean the vibing part but the using the app part.

I agree that agents are super good for one shotting throwaway code for tasks that would have required manual human actions previously but I would never have bothered buying an app for that.

Not having to deal with IT support for relatives is a win though since now I can just throw it at an LLM!

Mix of this being spread over more than a year, that I'm not doomscrolling because HackerNews and Telegram are my main social media presences, and being unemployed/prematurely retired (which one depends on what one thinks of €1k/month passive income and no rent).

I am also weirdly unmotivated by opportunities to spend money, which is both how I got this passive income and lack of rent, and why it's borderline enough for me. FIRE is very easy when all your working life, you only spend rent+50%, the rest of your paycheque going to savings and investments; but most people can't do this.

As a small business owner I see this. Random people contacting me to sell a software where I can instantly tell it is vibe coded. Subscription prices half or 1/4th of incumbents. Domain names registered in the last few months.

Just go ahead and do set it up. You’ll be my hero (I don’t have mental capacity to research which jurisdiction works for this right now, but obviously not in US or EU.)

Its actually very tone deaf to even pipe up and say anything about this given they did this already to Windsurf themselves, and they previously banned x.ai employees as well from using their models for basically the same thing Elon admitted they were doing with OpenAI.

I'm surprised they decided to say anything at all given their rocky past. Its the pinnacle of being a hypocrite.

Some of you guys hate Musk so much that you say blatantly nonsensical stuff. Like, how can you be on HN and authoritatively mouthing off and not have some idea about the xAI-Anthropic compute contract, or the fact that xAI took down its own Grok compute capability so it could sell that compute to Anthropic? And if Anthropic hated Musk/xAI as much as you are saying they do, why would they be paying xAI billions for compute? Think/research before typing.

Some of us are still employed mate

This was pretty pertinent to the meaning of that line, brother.

If you can take any book and turn it into a model, because it's "transformative enough", and "AI learns just like a person does", then surely a model distilling another model is transformative and fair use.

They tied themselves into knots fighting the letter of the law, and now, when they need the spirit of the law - that each creator deserves protection for their work - now we devolve to the law of the jungle. Maybe we'll even see LLM book curses, the way medieval scribes damned book thieves to blindness and worms.

Just look at the pattern... they collude, they provide to them whatever is needed, and at some point, if this is not true yet, they will be the ones who will tell them what to do or not, bc you know how humans are, right... blackmailing, mess up the business or shames of people, etc.

It is just a matter of time. The state, as we know it, will collapse or will be greatly reduced, which, from a point of view, is positive, from another, Idk, bc if someone replaces that, we will be in the hands of someone, as usual...

Guy this was openly stated more than a decade and half ago by all the current oligarchs

Imagine the United States DoD, CIA, NSA training an LLM on all its top secret intel.

Oh my sweet summer child...

This has been the continual goal of the US defense and intelligence agencies since 9/11, which was blamed on a lack of information sharing. Why do you think Edward Snowden had access to everything? Because it was consolidated post-2001.

You think there is any access control at the upper levels of the dark intelligence agencies? They have access to the whole take.

And it used to be that was pretty useless unless you had an actual lead. You'd need a East German Stasi level of labor to read everyone's secrets at scale... Now you don't, thanks to AI.

https://enwp.org/Total_Information_Awareness

So much for the all-powerful cabal.

Why fight someone who is so open to corruption? The important thing is that they got everything they wanted, and all they had to do was throw a few million at the clown, attend his parties, and maybe take down some diversity programs they didn't believe in anyway.

China is a different beast altogether, though.

Because they know it will work. You don’t have to watch Trump speak for very long to know he’s not very intelligent. Similarly, you don’t have to be much smarter than him to know how easily he can be manipulated with adulation.

Lmao

I'm curious if we'll ever get a form of civil war where corporate militias (corporate _robot_ militias in the case of xAI, of course) draw fire on good old meat policemen coming to bring the CEO to court.

Sure, it may happen, but I suspect the alternatives (bribing, sending a scapegoat to jail, buying the elections, etc... and focus on the "making money" part) will stay preferable for a while.

This has already been broken, and they've become so much more bold since then.

"For the general election of 5 March 1933, the Nazis were allied with other nationalist and conservative factions. At a secret meeting on 20 February, major German industrialists had agreed to finance the Nazis' election campaign."

https://en.wikipedia.org/wiki/Enabling_Act_of_1933

I think that if you study socialism in the 20th century you cannot say this seriously when you compare the alternatives.

Good thing you reminded us, otherwise we might've had critical thoughts about the current system.

Doubt.

The networking between people in power at the top will, in my opinion, most likely assure that they remain in power, because at the end of the day that is what they want most.

I expect government will seize control of the most advanced models (if they haven't already), and the rest of us will be throttled, and status quo will be maintained. I do not expect that either super advanced agents or the owners of the hardware they run on will be able to pull off the kind of coup you describe. At all.

Who do you think will rule at that point? They just cannot fight that, they do not have the technology these companies have.

Usually, the number of bodies who will comply is the proxy for power, but with automation, it could be different in the near future.

https://xkcd.com/538/

Regulation alone cannot control if it cannot be made effective.

You think all those lawyers and bar tenders and whatever MTG was serving in Congress are out there tilling fields and sewing shirts? Nope. Without that tech they are f...

"They"

Who are you talking about? Because if you haven't noticed the people in charge now will die.

Most of the Fortune 500 of the 1900s have been gone for decades. The olds are vastly outnumbered by youth who either hate them or wouldn't mind just getting to beat someone who can't fight back up

With measles and food contamination going wild now it's just as likely the rich and pols kill themselves.

There is no beating physics and Elon, Thiel, and such rely on a lot of people to do the work for them. If society implodes all bets are off. Their wealth is coupled to legitimacy of America. Those body guards aren't going to abandon families for a broke former billionaire

There's no upside for the rich is things implode or the divide is not believed to exist; they become common rabble and get a pick axe too

The rich and politicians are just stupid meat suits too. They misread a room and screw up alot

And that's just the most obvious one. There are many other kinds of threats which are much more subtle.

That is how governments traditionally stay in charge, and I expect the pattern to continue. Citizens simply don't have access to the levels of destruction the state is capable of, by design. They can have all the advanced technology they want, so long as it doesn't threaten our leaders' monopoly on violence.

I have no doubt there are agency operatives influencing discussion online, from recorded precedence. Calling the concept 'nonsense' is overly dismissive. Regardless, using that kind of claim to deflect from the premise isn't a compelling reason to invalidate the concept. Ideas can stand on their own.

Also, copyright has always been privately enforced anyway?

Enforcement ultimately happens through law and the legal system.

Otherwise, it's just a standard case between two private parties resolved through our legal system; e.g. Linkedin vs Hi5.

This is a gross distortion. Standard contractual rules bind the parties that signed the contract and the remedies are proportional to the damages and bounded. Copyright is tort law, the state binds the world to respect the rights of creators and the damages on infringement are punitive and can far exceed the actual commercial damages - to the point of bankrupting the infringer.

The key to torts is that the state is not neutral, there is a social good here it's protecting. Crucially, copyright, like some other torts - securities, antitrust, environmental, battery - also has a criminal enforcement regime, where, for particularly serious offenses, the state actually invests public resources to put the criminal infringer behind bars with little to no involvement from the original rights holders.

In the particular case of US, there is an entire state apparatus dedicated to enforcing US copyrights, a foreign affairs policy to shutdown "Notorious markets for counterfeiting and piracy" in other countries, international enforcement of DMCA etc.

The idea that a private TOS has the same level of public protection as copyright is downright childish.

On the third hand, the settlement Anthropic famously had to pay was for copyright infringement because they didn't actually have the right to even access some of the training data they used, so I can see how this might be compatible with the law.

On the fourth hand, I'm saying that as someone who absolutely isn't a lawyer and sometimes gets surprised when reading about copyright cases that sure sound like they ought to have been trademark cases given my limited understanding.

For those who chose to give it all away, the ship has sailed, but they did choose to give it away for free to anyone so they can't complain that they succeeded.

If buying one copy of a book entitles the ai company to train on that data and redistribute information derived from it in perpetuity, then why shouldn’t a rival ai company be allowed to train on tokens from say OpenAI and redistribute information derived from the OpenAI model also in perpetuity? The rival ai company paid for the tokens, after all.

I mean it quite likely is sort of in the same legal bucket. It won’t stop them suing but it is going to be the legal equivalent of two biologically-related warlords making their champions fight with their hands tied for sport.

Big models are built by scraping the recorded thoughts of everyone, so giving everyone a chance to run a distilled small model is just going full circle.

Obviously the underlying motivations aren't 100% altruistic, but I'll still take it.

This is not a necessary end state. It is the byproduct of the disease of sociopathic MBAs.

(Source, worked for multiple tech companies that were laser focused on delighting customers until money people came in and ruined it, to the point they would prefer devs sit idle than work on things the MBAs didn’t have on a priority list).

you only need people's money if you can't control their labor.

did a plantation owner in the 1600s need the money of his slaves ?

if you control a robot that can fight and take over land, enslave workers for things robots are bad at, harvest food, create buildings etc. what do you need other people for ? your entertainment? robots can do that too.

For the record, with the exception of Amazon’s Jassy, the CEOs of the top 5 companies are all engineering types with engineering credentials.

The CEO credentials of the top 5 AI companies are all science degrees.

Absolutely not.

Sundar's education is: IIT-Kharagpur Bachelors in Eng Stanford Masters in Eng Wharton MBA

Maybe you're thinking of the last one

He is an MBA. And his reign has been full of managerial bullshit

Sure, the guys focused on tech don't usually think with an empathy of a nurse (novadays not even nurses are guaranteed to be empathetic), but I know quite a few devs and IT people who are extremely pro-social.

It's the managerial, financial-oriented mindset that is the greatest predictor of sociopathic character.

The more power one has in an organization the more socio- and psycho-pathic they tend to be, simply because it's easier to get to the top if you are an amoral person, who is not holding back for any reason other than having power. That is why absolute power corrupts absolutely.

MBAs are bred to be sociopathic money-grabbers - I know that from experience.

I once made the mistake of going into an MBA program. On the very first lecture the lecturer asked each participant for his/her motivation for being there.

Most of them said 'money' and those who didn't were asked again until they caved-in and said 'money' as well or they were laughed out by the lecturer.

MBAs are expected to be sociopathic or you as a company shareholder won't be able to motivate them easily to do your bidding.

It's like public/private schooling, but worse - those programs are designed to create mindless drones for the institutions that use them for their own policy enforcement.

Yes it is, in the legal/copyright sense of fair use. That's why they ban it in their TOS. Which customers agree to when signing up for the service.

So… my TOS doesn’t matter, but theirs does?

No, you don't need to agree to TOS. I'm sure you never actually agree to them and you're not going to jail for it. If you believe that, I hope you never visited cnn.com, for example. Their Terms of Use is 11,000 words. Are you sure you have agreed to that when you clicked on some random news link? Are you sure you're happy to to surrender your legal defense and its expenses to them if they make a claim against you? You're OK that they have no liability for sharing your PII when they're not authorized to? You won't complain if you pay for a subscription and they don't give you access to the subscription content, don't refund your payment, and don't even tell you why?

You accept it. You pay consideration for it. If your website has a TOS dickover, that requires someone attest with their legal name and pay you $1, yes, it may be enforceable under some circumstances.

Without a licence you can’t wget -r a website and make copies of the materials on it, alter them, and so forth.

The GPL established this rather clearly. Copyright law doesn’t require consideration.

(The licence itself is a basic BSD licence, so it just requires attribution including in marketing materials, which obviously hasn’t happened.)

Within this context, I don’t think so. I can’t make a website that buries some shrink wrap that requires everyone who reads it become vegan.

Wilhoit Conservatism: In-groups protected by contract law but not bound by it, alongside out-groups bound by contract law, but not protected by it.

1. Copyright violations (can put you in jail) 2. TOS violations (will be a fine at worst)

Companies do get away with drive-by legal shittiness way too often and frankly the practice needs to be reined in, but at the end of the day the only damages are the financial ones you can prove in court.

Incidentally, this suggests that once an LLM is capable of accessing and distilling a competitor's LLM without human intervention, then any legal argument about TOS violation is moot. But somehow I doubt that will fly in court.

It's possible that distillation for other reasons, with no intent to harm the vendor you distill from, would have been ruled to be fair use. But in law, intent matters.

Copyright (C) 1998, 2015 by Greg Egan

First Night Shade Books edition 2015

All rights reserved. No part of this book may be reproduced in any manner without the express written consent of the publisher, blah blah (it felt very ironic to transcribe that bit in particular to make this point)

For example, never will you see printed in a book something like "this book is for the exclusive use of the original owner and you cannot lend, resale or otherwise make available to other people" - if such a thing was possible, like most software EULAs do, publishers would be all over it.

However, if content is “licensed” instead of “sold” …

Asking for all Anthropic employees who dream big.

ChatGPT 3 was released nearly six years ago, and the models are staging increasingly aggressive breakouts now. Where are they going to be by 2030?

No, the AI companies merely figured out a way to spin gross negligence into a PR win. Any idiot can build a Murderbot which "goes rogue" - it can be as simple as taping a knife to a Roomba. The harm it does is not in any way related to its "intelligence" or "sentience".

We're seeing "breakouts" because the AI companies are being rewarded for their incompetence. You don't have incredibly lax security standards and zero form of oversight resulting in fully-automated felonies which should result in jail time, you instead have a "powerful near-sentient cybersecurity model" and should be given hundreds of billions of dollars!

The government level lobby against foreign competitors is not contractual but just another form of reinvention of criminal injunctions against infringers.

This is a battle for power, let us not be naive...

My employer let us evaulate all the big providers for a month each before having a vote on which one we'd like to keep. Cursor won almost unanimously. (This was before the xAI takeover was even on the horizon, otherwise I would've voted differently myself.)

Ironically, as a Cursor and Claude subscriber, but not an OpenAI subscriber, this will push me back to Anthropic. I don't feel good about giving OpenAI money directly, but I've found switching to GPT 5.6 Sol now and again in Cursor to be useful. Now I'll probably just forget about OpenAI models.

> as a Cursor and Claude subscriber, but not an OpenAI subscriber, this will push me back to Anthropic

Instead of deciding based on the merits of the model, they are being pushed by the actions of the companies. You should be able to use the best model, or whatever model you like, independently of what any company thinks or does.

I think maybe you’re suggesting private companies should be forced to support all models which is a position that isn’t technically feasible or one that most would agree is even a sensible ideal.

I'm suggesting you shouldn't let a company dictate which models you use. Putting yourself in a position where they can is detrimental, because you miss out on using the best model for the job.

https://developers.google.com/style/highlights

They're commodity fetishism maxxing.

Cursor/SpaceX/Grok is a different beast. I think they will cheat other companies but I'm not convinced yet that they have it out for users. We'll see. I'm certainly not some kind of Cursor lifer or something.

We should all move to OpenCode anyway. :)

If his teams will handle the data of US citizens with such disregard, he'll almost certainly treat his customers (many of whom are those same US citizens) the same.

The other AI companies obfuscate their goals much better.

Grok is a workhorse, not the smartest yet, but produces good code and follows the plan. Grok 5, which has been training since beginnig of this year, will probably be a monster. (10T/6T Param model variants)

I have a $20 Cursor sub still personally and would say that the usage for Grok models is much higher than what $20 gets you with Anthropic. I can't speak for the $100 plan but I would assume it scales similarly.

I rarely run out of my usage monthly even with the $20 plan.

The best bet is Cursor Ultra ($200), which is the equivalent a month plan that you get from OpenAI and Anthropic. I get a LOT more capacity out of it because I almost always use the "Auto" mode, which auto picks between Opus 5, GPT 5.6, Grok 4.6, and Composer 2.5.

I like Auto mode so much actually. I use it for my Cursor work account as well because it's about 10x faster than just using Opus 5 or GPT 5.6 and about 90% is good. I switch back to the bigger models if I get in trouble, but that's pretty rare.

Vibe coding requires some discipline. You need understand what it's trying to do. If you have a good mental model you can keep it out of trouble. If you don't, it doesn't matter how good the model is. You're gonna end up in a mess.

This provides a method, but the data looks stale and perhaps a bit thin compared to say, Cursor, or even AntiGravity data.

It has been more of an OpenClaw or Hermes alternative than a coding agent like OpenCode or Pi, so it’s likely to do well given less context bloat.

Cursor will be fine after this, as they have their own first party models and can potentially use open weight models via API or their own hardware, which have gotten good enough to Opus and Fable quality now for many use cases.

The model providers are in a difficult position. With an added danger of open weight models that are within reach of large enterprises running themselves.

I think Sam Altman has played a clever business move here, they can use this excuse immediately after the Hugging Face leaks and it's not going to be dismissed willy-nilly. Whereas long term they probably wanted to end their partnership with Cursor anyway. I bet Dario will be happy about this too, but maybe the compute deal prevents him from doing similar anytime soon.

Very interesting times we live in...

Burn.

This move seems to protect Astra against distillation. I wonder whether Astra will be released before or after 11/12.

codex growth at 20mill now, prob more cost effective to allocate that compute to people directly paying now

The man is an outright existential threat to democracy.

Do I smell a cursor-analogue from openAI?

I wonder if Anthropic will follow suit given their campaign against distillation or if they'll stay quiet given their reliance on SpaceX (among others) for hardware.

i work in qa so i'm naturally wary of adding another critical dependency to a development workflow, if the model you actually want can just disappear because two vendor change ownership or start argueing then the wrapper has to be really good on its own.

cursors harness and its own models may be good enough that users stay, and if they do cursor has a moat at least. if everyone follows the openai models out the door, it will just become a polished ui.

Maybe if Claude joined along it would be more impactful. For better or worse though, Xai seems like its lining themselves up to handle being on their own for cursor.

Is there a way to work around this (like using your OpenAI limits directly?)

The first party cursor models (initially composer and now grok-cursor) and price very competitively for their quality and have very generous usage limits included in the base plans, and then per token pricing afterwards.

Occasionally I’ve used some of the Anthropic models in it but found their cost massively outweighed the benefits for the type of usage we do.

I was using Sonnet 4.5 and 4.6 for a while but the newer models are a lot more expensive at the per-token pricing.

Using OpenAI through codex (cli) is the way you should use it.

> After Musk acquired Twitter, now part of SpaceX, the company broke (opens in a new window) the terms of our contract (alongside many others).

I have never seen one large company throw another large company under the bus without any legalese filters like this before. Wondering if OpenAI site got pwned. Or maybe we in some strange, alternative reality. I doubt even their LLMs would be this frank.

https://www.theverge.com/ai-artificial-intelligence/921546/e...

[1] https://www.forbes.com/sites/antoniopequenoiv/2026/04/30/elo...

https://news.ycombinator.com/item?id=49428835

If Claude pulls away from Cursor as well then Cursor is basically done.

I'm not a fan of Elon, nor do I use Cursor, but OpenAI and Anthropic aren't the only models.

Also, it's a signal more than anything. OpenAI pulls out of Cursor. Then Anthropic and Google does as well. Then you no longer have "Cursor" you have Grok.

Grok doesn't have the best reputation in general, regardless if you're tech savvy or not.

The movie Goya’s Ghost explains this in great detail.

It's gunna to happen anyway.

We are sailing into the age of vertically integrated intelligence.

> we’ve decided to hold the contract cancellation to the latest date we can while not providing future models to Cursor.

Seriously though, when the api providers start blacklisting cursor and they’re forced to swap to open weight models…

If that happens, what even is the difference between openrouter and cursor?

Strategically, OpenAI long-term will want to own the direct relationship with all their customers. As will most major AI labs.

https://cursor.com/docs/models-and-pricing

The quality/price ratio is clearly superior when running Qwen, GLM, or Deepseek.

To make a product more exclusive won't hide the fact that product is becoming more and more irrelevant.

i only tell VC type people or others from whom i want something "oh yeah grok is cool" if they mention it, because they're always sucking off Elon

But this is kind of like banning GPU exports to China. It will encourage them to develop their own technology. Yes, it will take them longer, but in the long-run China will be more successful with their own GPU tech.

The same could be true of SpaceX/xAI

You had me until:

>The same could be true of a company run by a drug addict.

Elon Musk is nowhere near as capable as millions of intelligent developers in China.

Is there really a high demand for racist chatbots with no safety controls?

And even if there is, will governments even allow it to keep operating?

Using both OpenAI and Anthropic models, Cursor is cheaper than having two separate subscriptions, and their free tier of Composer is good enough. I will now have to reconsider my options.

Then you can let it run, and when it stops, point a fresh session to the ledger to pick up the work from where the previous one left off.

You can get more fancy by using something like GitHub issues for coordination too and have multiple sessions create issues and epics and others to work on that in isolated worktrees, but the spec-plan-ledger approach usually works great for small to medium large projects.

There are also a bunch of newer startups in this space. I'm building one myself, https://boxes.dev -- we're very early but building for this exact use case. Some other ones worth a look are Factory Droid and Amp Orbs. Those two build their own agent harness (like Cursor), whereas with boxes.dev we run the native codex and claude code harnesses directly.

    We are making this choice because 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.

For a company named "Open" AI, they've been hardly open at all and the slew of resignations of top tier execs in the company is really the proof.

Make no mistake, this isn't a moral decision. This is just killing the competition using morality as an excuse.

Last year, Anthropic cut off Windsurf the same way [2]. Fortunately, unlike Anthropic, OpenAI allows you to use your subscription with other harnesses, including Cursor.

Of course today, Anthropic is taking the "high road" and is OK with with SpaceX [3] including their models in Cursor, because they need SpaceX's compute.

[1] https://www.forbes.com/sites/antoniopequenoiv/2026/04/30/elo...

[2] https://techcrunch.com/2025/06/03/windsurf-says-anthropic-is...

[3] https://x.com/NotTomBrown/status/2093541294027280657

Not doing business with them is a completely normal and unremarkable decision.

Anyone else surprised that the maximum notice period allowed is about 3 months?

I would have thought 3 months is minimum notice period, with provisions to stop faster if any provable violation occurs.

This clause probably has a expiry date and they are saying “we are trying not to be jerks here by waiting until the last second to trigger it”.

I took the 50% discount instead of cancelling - but with this I'll likely not renew.

So sad to watch them go from great to mediocre in such a short time.

Virtue signaling in glass houses...

i loved switching between sol / fable / grok 4.5.

... and something tells me fable might not be available on cursor for much longer :/

- fast because the code is already indexed no need to start every prompt rediscovering the world with rg

- in editor/vscode review much better than relying on git diff at the very end. You can click on next button to jump there. Quick edit in place. Or a quick prompt (seconds) to fix something small

- in editor completion / quick edits, when you do things by hand.

It seems that everyone loves agentic Claude code things these days but I don’t understand how you can review what it did and remain as much in the flow as you do with cursor.

Also I find that incredibly slow. By the time Claude finished a prompt I could have done a few in cursor.

And mind you, I’ve been wanting to use zed instead but the clunky vscode fork called cursor. But cursor has just a better UX for AI integration even though it’s slow and guzzle up RAM.

EDIT - AI says:

Cursor seems to have intentionally dropped semantic/embedding-based codebase indexing. Their explanation is that newer coding agents are now good enough at searching repos directly. They can grep in parallel, inspect directories, read likely files, and refine the search themselves. Cursor says this now works as well or better for most cases. [1]

So the architecture shifted from something like “chunk → embed → vector DB → retrieve” to “local search index/ripgrep → agent explores.” There is still an index, but it’s a conventional local search index. No embeddings are involved in that retrieval path. [1]

The interesting part is that Cursor was still reporting a ~12.5% accuracy gain from semantic indexing in January 2026. [2] So this looks less like “semantic search never worked” and more like agentic search improving fast enough that the extra infrastructure stopped being worth it.

[1] https://forum.cursor.com/t/what-do-you-think-about-cursor-re...

[2] https://cursor.com/blog/secure-codebase-indexing

but the the review what the AI did by jumping directly into the file with accept/reject/go next/next file with the ability to edit in place was nowhere near as good in zed vs cursor last I checked.

The features you mention are in Antigravity IDE v1, which I am trying as I still have a Gemini subscription.

No idea how long it will be around for, as Google Antigravity2 (more like Claude Code) is clearly the favoured way forward, plus the new Antigravity for VSCode Extension*, which I didn’t like as much when I tried it. It offers the same accept/reject hunk workflow, but I don’t know about editing in place. It doesn’t feel as if it has the tight integration of Antigravity IDE v1, which feels as if it was designed to challenge Cursor.

Having said that, much of the time I use Google Antigravity2 alongside a Jetbrains IDE. You can still review all diffs within Google Antigravity2 and attach comments to any line on order for a followup prompt to make amendments. But I still prefer the review and accept/reject hunk workflow.

* https://marketplace.visualstudio.com/items?itemName=Google.g...

There are things I will probably miss from Cursor, as they used to iterate on DX goodies quite quickly. For frontend tasks, I liked the idea of Cursor Browser being able to do a devtools-like selection of an element tree to be included in the chat context. I’ve been told there are similar extensions available for vanilla VSCode, so will have to see how good they are.

I don’t like relying on GIT for that and making things as good/done as I go. But with these agentic tools that’s what I have to do. Stage everything I like, manually edits a few bits.

But what do you when you find something where you want to say the agent: I don’t like this for reasons x and y redo as z. I’m like one shortcut away in context inside cursor to say that to my agent? You have to change tab in herdr and then what? Type the file name and line numbers? Is there a way to do this that does not seem very frustrating?

Also open code, codex, etc. super slow. They grep the code. Every. Single. Time. Cursor has a semantic index

I know this is change of context, but I have full screen terminal toggled with keyboard shortcut and with middle mouse paste (Linux) I do not have to move cursor anywhere

The in-editor context review is something that does not seem long for first world. Sadly. Maybe I’ll be forced to make a zed extension or something, maybe on top of jujustu…

  gdf() { git diff '*'$1'*'; }

  gdf fileName (doesn't need full file name)

  gdf folderName (all files in a folder)

iterm2 on mac also has a similar mode, but it is a bit finicky to set it up (it is really hidden in the settings).

For instance, in Claude Code you can have it bring up this TUI overlay after every edit it makes (it literally is an overlay, rather than a separate tab in your terminal) and you can annotate line-by-line anything you want changed or rejected. Then once you exit the overlay those suggestions are automatically piped back into the agent.

I tried a few of these TUI-based code annotators, and this one is by far the best. It also works for files not under version control.

I'm currently not using it as I also prefer something more akin to "text editor with agent built in", but I wanted to mention here in case you were curious.

The GUI slightly better than Claude. And it’s a million times cheaper than Claude.

The way I review it is to let it finish a PR sized slice and then I review the whole diff in a separate terminal as if it’s a PR from someone else. Then I very impolitely list the shit that needs fixing, let the LLM grind out a new batch and repeat the process. When the code is good enough I throw it into a commit and push a PR out for other humans to review.

Clicking around in a UI would just waste time and if the change is too large to review in a terminal window the change is too large for another human to review.

I sometimes see people reviewing LLM code leaving the same review comments multiple times, or worse, after some time they just give up reviewing.

The part where you add instructions into a document is a bit of hit and miss. If they’re general enough it works but at my employer they have gone all out and have LLMs writing instructions that are committed to repos. There’s also a bunch of tools enforcing security instructions that usually just make the model confused (it starts arguing against the instructions randomly in the middle of performing a task). The agents are “stuttering” on most tasks because of the bloated instructions.

It's funny because we're once again speed running what the best practices already are!

Look at the finished piece of work and then treat the agent as if it’s that colleague you really hate and nitpick the shit out of the change it wrote. Usually the agent writes some over-engineered garbage which you then “help it” whittle down to something you would be okay approving if another human wrote it.

Insanely inefficient. It's 10x more productive to watch the thinking traces and edits in real time and steer the model appropriately.

If your workflow is typical no wonder my team members who use claude code are so much less productive.

If I work with one agent / few subagents on one feature I can steer it as soon as I notice it drifting into the direction of waste. This way I only review the total diff 1.5-2 times. And I also don't waste my own mental energy on context switching between tasks agents are producing diffs for in parallel.

If that’s what you got from my comment then you need to review it again.

When thinking first started and you would still see the whole "thinking process", I thought it was a ploy to 10x token use because it was just the most inane bullshit. "But wait, the user is asking me to" in loops.

Wasn't there a study recently that even found a model's performance was sometimes better when the "reasoning" was nonsense? As in, no clear correllation between what the reasoning says in a human's interpretation, and how the model actually did with the task.

I sometimes read the traces out of boredom or morbid curiosity. My favourite bit with Claude is how, even if you give a very comprehensive prompt in complete sentences, almost every trace will contain a variation of "the user asks me to X, but their thought cuts off mid-sentence."

Fever dream.

I use Zed to review the code.

CLI is great for most things. Some times, when I am reviewing a PR, I just tell Codex to do a diff between source and target branches in PR, and generate a HTML report. Then I read that, and also review relevant pieces of code in Zed.

I tried using Cursor, many of my colleagues use it, but stopped using it due to clunkiness. And of late, even in there, many of my colleagues use agentic coding. Instead, I prefer the snapiness of CLI (Codex coded in rust, takes less memory per session) and Zed (also coded in rust,) and blazingly fast. Zed is the new Sublime Text if you will.

I just retried Zed and yeah, it's much much better than last time I gave it a shot!!!

I turned on the review stuff, and yeah I can do more or less what I could in Cursor. Check something. Do a quick inline promtp for the three lines I dislike and move on.

And of course the editor is so much nimbler/nicer.

No plan mode, but for big edits I'd probably use codex or something just like you. If you have any advice for integrating for that flow, I'll take them.

I guess whatever advantage cursor has, it will die once "Sol" is no longer integrated, as it was my favorite heavy work model. But I sure will miss the snappy auto/composer stack of Cursor.

[1] https://opencode.ai/

[2] https://github.com/oraios/serena

Separate terminal tab with git diff/show, Zed for edits - rarely, there is little reason to change code yourself. Opus at low effort is faster than you can switch windows.

This is a big quality of life thing. Sure, I should just go on a walk without having an agent going, but oftentimes it’s not intruding on it - it’s the thing that actually ENABLES it.

For the past three or four months it's performed very well, even the cheaper Gemini models. No serious complaints. But if you go back a year it was pretty terrible. My work style is to keep the coding agent on a short leash so YMMV.

I love using it for the chat question style where I ask it things like:

For these frontend pages with X options collect the database tables for each.

This is of course something easy to do manually, but I like it stepping through a couple functions much faster than me searching around manually and maybe getting lazy and not checking all of them.

I mainly use Cursor for another reason though: all of my work is remote over SSH and I don't like keystroke latency, so I prefer code editing / agent conversations to feel local. (Everything else is done in the terminal.) But I would also prefer to use something less propriety. I may go back to VSCode with plugins to achieve a similar experience.

I haven't tried it yet but it looks promising. He would be thrilled if people give it a go!

Slop website indicates vibe-coded software, and I wouldn't let anything like that even look at my code.

It’s not an editor in that sense but a tool optimized for reviewing PRs produced by AI.

The products are basically equivalent but JetBrains is a profitable bootstrapped company that has been building IDEs for 20 years and Cursor is a flash in the pan startup with a squillion dollar valuation to support. Sooner or later they’re going to do something that will fuck their users over.

And selling to SpaceX was it!

If your coding tool depends on other companies’ AI models, one business decision can suddenly change the whole product

All of these companies, and I truly mean all, stole data to train their models.

Although I wouldn't take an Openai announcement literally, as Sam Altman is also not trustworthy.

If a company can provide the access to such a model via AWS Bedrock or similar enterprise compliant platform with a decent IDE/harness they can capture billions of dollars of market.

The closest we have to this is Claude Code but I think it can even be better and specialised. Cursor has an enormous amount of data on coding and developer usage and maybe they should use that to build a proper platform

I'm surprised how clear they are about this.

Related:

Can highly recommend Zed as an alternative, it's better than VSCode with the Codex extension as managing multiple chats is much nicer in Zed.

Zed: https://zed.dev/

Since the introduction of the Cursor Token Rate, my company tells us to use Claude and Codex directly with Anthropic (or Vertex AI) and OpenAI, respectively.

https://cursor.com/docs/models-and-pricing#cursor-token-rate

I'm now using Grok 4.6 most of the time. Sonnet some times. But that's it. Cursor-native models 99% of the time. I'm pretty sure Anthropic will follow OpenAI soon in pulling their models out of Cursor because it's now a true competitor.

How can you ban a competitor from using a model that you openly provide access to?

I understand that the OpenAI ToS may have been violated by X, but has anyone ruled on distillation as fair use? OpenAI believes distillation of nonAI produced content is fair use, so they don’t appear to have a leg to stand on.

there are plenty of alternatives, 100s, so I don't think antitrust is on the table, think of this more like a boycott of the world's richest man for being a twat

Distillation is a threat to frontier model companies, but that doesn’t make it dangerous or wrong. OpenAI can (and likely will) use the same logic to ban you from using open source harnesses.

That is to say, this isn’t a fig leaf imo.

Right now I think GPT 5.6 Luna is by far the most cost effective model, so this change is effectively kicking me off of Cursor.

I did try Kiro which seems to have the same philosophy but it's just not there. Most of the features I like in Cursor are missing.

Some other commenters mentioned that VS Code + Copilot would be better these days. Dunno.

OpenKnowledge seems to maybe fill that niche somewhat (it also supports annotating Markdown with comments to feed into an agent), and has agent integration within the GUI, but one killer feature that seems to be missing is Remote SSH support (short of hosting its web server on the remote and having the IDE live in the browser).