It's always been clear to me that Cursor's business model of reselling others' APIs had its days numbered. Not necessarily because the providers would pull the plug, but because you wouldn't be able to compete with subsidized plans.
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.
The TUI version of Claude Code just isn't very good, though. The default keybindings do not work in the majority of terminals (because they generate the same byte sequences for key sequences assigned to different commands). Editing large prompts is rather painful because there doesn't seem to be in-prompt search commands. There is no source code browser, so source code references cannot checked by the users without a separate tool. Diffs cannot be copied directly because they are indented by spaces (even though some tools strip indentation from patches, it's still ugly). Switching between conversations is somewhat discouraged by the interface, which makes it harder to keep the context clean.
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.
The tide is turning though. I was using the old non agent view in cursor every day for about a year when I built my startup with it. Then Claude Code blew it out of the water. But since two month or so I much prefer Codex and Cursor with the agent view.
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 ;)
Ingesting the whole codebase in their cloud and then maintaining indexes of where everything lives and what already exists. Not sure if that's where they are going but that's what I feel looking at some of the beta functionality and announcements.
I saw that happen culturally. But never saw that much different in performance. And I really dislike terminal development. And their desktop app is really buggy.
Out of interest when did you last use the desktop app? I am a huge daily user of it and apart from very minimal bugs, it works perfectly for me. Wasn't always the case, so they've definitely made improvements
Given you can use the same models on both, you're right about performance.
However I've found the cross-session context and memory in claude code much more helpful which can lead to better outcomes faster.
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.
> Claude code is a decent harness but then you have to use Anthropic models...
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.
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.
I think the point is that CC was the first general use, terminal based harness. I think. Now you can use PI, or OpenCode, or whatever, and use any model you like.
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.
I've recently moved to a company that uses Cursor, and I'm actually quite fond of it.
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.
Why do people keep insisting that subscription plans are subsidized? They are not. API prices are outrageous, designed to milk enterprise users. They probably have an 80-90% profit margin on subscription plans.
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.
OpenAI leaked financial data (2025) suggests this - it has $5.7 billion for marketing which 44% of revenue. It’s hard to explain where these billions go other than into subsidised subscription plans and a free tier.
Because using chinese models on OpenRouter is way more expensive than the equivalent codex sub, unless you use something clearly inferior like deepseek flash.
If the Opus/Fable harness is superior and it's written using Opus/Fable then it will cease to be a moat. Otherwise the premise for coding using LLMs is essentially untrue.
> 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.
cursor adds $0.25/M to your token bill for 3rd party services. sounds small but it's insanely big on cached inputs, which are an insanely high % of use
Honestly, Cursor's AI harness worked best with the Anthropic models. Could just be my bias though. Evaluating frontier LLM's all seem to be by vibes. I think the industry is understanding it's both the harness and the model that makes the difference; can't half ass either, and some harnesses do better with different models. I'm good to leave this "career." Senior software engineers, make your bag.
On one hand, yes, it doesn't make sense to give out subsidized tokens forever.
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.
Right, subscriptions are not only cheaper because they "create lock-in". They also let you do things like forecast demand and plan your capacity for it, borrow against it sometimes if you need to, and keep somewhat less around because your cashflows are predictable. It is also valuable because it keeps customers around but that's by no means the only reason recurring revenue gets a higher multiple.
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.
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.
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.