Well This Was Unexpected...
Sunday, Aug 30 · 27 min · about 8 min to read
Theo-brain watched all 27 min
After SpaceX bought Cursor, OpenAI is ending Cursor's direct access to its models on November 12. Theo thinks the real trigger is Astra, OpenAI's next model, which OpenAI fears rivals could train their own models on. His lesson: for top models, pay the AI labs directly and use tools you control, not a middle layer that a company fight can cut off.
Briefing is enoughYou can skip the video; this briefing gives you what you need unless you want Theo's full commentary.
Tap any time to jump the video there.
Key takeaways
After SpaceX acquired Cursor, OpenAI said Cursor's direct access to its models ends November 12. Cursor subscriptions will no longer include OpenAI models.
Under our proposal, Cursor's direct access to our models would end on November 12th. Yeah, this is real. Your Cursor subs will no longer allow you to use OpenAI models. So, instead of
Theo says the real trigger is Astra, OpenAI's next model. OpenAI thinks it is good enough that it fears distillation, meaning a rival training its own model on Astra's outputs. So it won't let companies it distrusts reach Astra under existing contracts.
This is because Astra apparently is a better enough model that OpenAI is concerned about distillation of Astra and as such is making sure any company that isn't compliant with the terms of service is not allowed to get Astra in within their existing contract. This
On model choice, Fable 5 is still Theo's preferred model for writing and landing code in real projects. But Soul (GPT-5.6) costs under $1 per benchmark run, against over $3 for Fable, largely because it uses far fewer tokens (units of text a model reads and writes).
between Fable and Soul feels very real in my day-to-day usage and Fable 5 is still my preferred model for actually writing and landing code in real projects. That said, we're not here to
when you look at it because the fable tasks are over $3 per run and souls under a dollar per run. So, it's a over 3x difference. A big part of that
Theo expects more of this. He predicts Anthropic, OpenAI and probably Google and SpaceX will keep cutting off access to rivals.
this. I have a feeling this is not a one-off thing. If anything, it's the start of the end. We're going to see more and more moves like this from Anthropic, OpenAI, and probably even companies like Google and SpaceX.
Two skills here are new to you.
Keep GPT models in Cursor through the side doors New
- What it is
- OpenAI says two routes stay open after the cutoff: using your own OpenAI API key in Cursor (an API key is a password-like code that bills your own OpenAI account), and using OpenAI's Codex extension inside Cursor.
- Why Theo cares
- These are the routes OpenAI says will keep working. Theo jokes that he has questions about anyone using the Codex plugin inside Cursor, but confirms it will still work.
- How to try it
- Create an OpenAI API key and add it to Cursor.
- Or install OpenAI's Codex extension inside Cursor.
know should work in the future. First, we will continue to allow using your own openi API key and similarly we will continue to provide access through our IDE extensions for cursor. So if you're using the codeex plugin and cursor I have questions but it'll still work cool
One open-source app over several lab subscriptions New
- What it is
- Apps that put Claude Code, Codex and other agents in one interface while you pay each lab directly. Theo names his own T3 Code, plus Gene from the Coolify developer, Superset, and Herder. All are open source.
- Why Theo cares
- Switching between Claude Code and Codex is 'obnoxious', and one unified place is great. He also says an OpenAI contact told him Codex subscriptions will keep working in T3 Code.
- How to try it
- Try T3 Code, Superset, Gene or Herder as one window for your agents.
- Sign in with your own Claude Code and Codex subscriptions rather than a bundled plan.
lots of other awesome options like Gene from the developer of Coolifi or Superset, an open source terminal focused solution, or even Herder, which is kind of a new flavor of a T-Mux type panel manager that is also open source. Yes, believe it or not, all three of these options are open source, which is
will say I had a contact at OpenAI hit us up making sure we knew that this will not affect T3 code and that they plan to continue supporting projects like ours indefinitely going forward. So if you're concerned that your codec sub is going to break in T3 code, I have been told you have nothing to worry about. To be
Advice
If you want frontier performance, buy the labs' own subscriptions. Theo says a $200 Claude Code or $200 Codex plan beats anything an all-in-one service can offer.
Fable and Soul have changed how I work. And if you are addicted to that, I'm going to be real with you, you probably need to use the subscriptions from the official labs. The value you get for a $200 cloud code sub or $200 codec sub just shits all over what you can get anywhere else. You can get a surprising
Theo's own numbers: a handful of subscriptions across both labs cost him about $1,000 a month, and he does upwards of $100,000 a month in inference (model usage).
now, I've done the numbers for my own usage. My handful of subscriptions across both providers cost me like a grand a month and I'm doing upwards of a hundred grand a month in inference. So
Own your tools. Deal directly with the products you rely on, not through middle layers that can get caught in companies' fights.
things. So yeah, if you want to avoid getting hurt by companies beefing with each other, make sure you own your tools and that your relations with the products that you rely on are direct and not routed through other layers like
Hot takes
Paying to use OpenAI's or Anthropic's models and training on the outputs is reasonable. Both labs trying to ban people for it is absurd.
His reasoning: OpenAI and Anthropic collected web data early, before the web noticed and locked it down. The data SpaceX's AI can't get is inaccessible because of those labs, so blocking it through the models too is bad.
early advantage by collecting all the data before the web realizes was happening and companies started locking it down. Getting the data from OpenAI and Anthropic by paying them to use their models is in my opinion totally reasonable and it is absurd that either company is trying to ban people from
fact, I'd go as far to say the data that SpaceX AI can't get is only inaccessible because of anthropic and open AI. So banning them from getting it the other way, which is through the models, is bad. Let's get to the final details of
Executives at OpenAI and Cursor arguing in public over how usage was counted is childish and not a good look.
His reasoning: Neither executive is really affected by the change, but customers are, so the public fight only makes things worse.
Seriously though, like I don't think it's a great look for people high up at these companies to be the ones doing this beef public because this is affecting users. Neither Tibo nor Michael are meaningly affected in their day-to-day work or their livelihoods even by these changes, but they have a massive negative impact on customers. So
Grok 4.5 and 4.6 are usable for a lot of real code work but are not frontier. After Fable and Soul, going back feels like the stone age.
His reasoning: With Grok you have to orchestrate the models yourself: decide how the work is split and where it goes, and audit it carefully.
models way better. And let's be real, Grock 4.5 and 4.6 are totally usable for a lot of realworld code work. They are not Frontier though. So, if you've already gotten a sip of the Kool-Aid and have felt the power possible with models like Fable and Soul, going back to Grock feels like a step back to the stone age
feels like a step back to the stone age in a lot of ways where you have to orchestrate the models yourself, so to speak. You choose how the work is broken up and where it goes and do a lot of careful auditing of that work directly.
How Theo thought about this
Cursor said OpenAI models were only 5% of its traffic. Theo asked what was being counted: users, requests or tokens. OpenAI models use about a third as many tokens per task, so a 5% token share is more like 15% of the actual work.
cursor user traffic, it is not specified whether or not that 5% is of requests, of users, or more likely of tokens. Because if anthropic models are three times less token efficient, then the 5% should be thought of more like 15%. I
Theo said the Astra detail makes OpenAI's move more understandable, but not better or worse in his view; it only makes clearer why it is happening. He kept the reason for a decision separate from his judgment of it.
does make why OpenAI is doing this more understandable. It doesn't make it better or worse in my opinion. It is just clearer why this is happening because of that. If Astra does turn out
Sign in to mark what you've watched and keep notes.
Have a question about this one? Ask about this video →
Theo Daily is unofficial and not affiliated with Theo or t3.gg. Written by Claude Opus 5.5 from each video's transcript.
