Cursor, since Grok 4.5, has had an incredible deal for frontier level models, their subscription now goes way further than OpenAI or Anthropic. Even on their lower tier plans you can use a lot tokens on their of their first party models (Grok and Composer) and not really run out comparatively. Combine them with an orchestrator and implementor type setup and it goes even further.
>their subscription now goes way further than OpenAI or Anthropic.
Until it doesn't...
Honestly, this entire OpenAI reset credit fiasco this past week has convinced me to rip off the Codex and Claude Code bandaids and start building my own proper Pi Coding Agent running models that I select and pay for on openrouter.
And I am feeling a lot better about it now that I've finally got it working.
I don't get the point of this. We all seem to agree that these companies have almost no moat, if one stops being a good deal, you can switch to another. That doesn't invalidate the existence of a deal that is currently good.
My point was that chasing deals like this is just kicking the can down the road. You're going to have to reckon with harsh price increases sooner or later.
So I have resolved to avoid that future-dreading and fixed it, basically.
My entire digital existence for the last 20 years or so has been a parasitical relationship with VC funding. They keep throwing money at business models that involve building market share and I keep benefiting. It hasn't stopped working yet.
Based on your own comments, it sounds like it wasn't really that hard to address the problem though right? Let's be real, it takes like 10 minutes to download OpenCode and point it to DeepSeek V4 Flash. That's an amazing 80-20.
Seems weird to me to not take advantage of the great deal the frontier labs are currently giving for subscription pricing when there's such an easy fallback in the worst case.
switching cost is something you then have to pay everytime the deal changes.
the rest of the comment is about moving from expensive switching costs of what harness you are using to having the difference be closer to changing some config in openrouter
But still for US frontier you're paying 10-20x more per token compared to their limited subscriptions. For China frontier you'll be good though, and that might be the future anyway.
Relying on a single frontier model to just zero-shot all the work is so 2025.
Deepseek V4 Flash 0731 is surprisingly capable and cheap. [0]
Checkout pi coding agent. You can create as many different sub-agents as you wish, to specialize and understand and tackle or pass off any problem you like. It's refreshing, really. I feel like a coder in control again.
Your reason for the rate limit reset is speculative. Here's a history of them and tweets that correspond to when they happen. Others can be the judge if they believe the stated reasons or not: https://codex-resets.com/
Yeah, but it also makes budgeting hard. Normally you'd want to budget about 20% per day to use your whole week up in time. But since they reset usage sporadically, it becomes optimal to burn tokens as fast as possible to be as low as possible when they reset.
But since you don't know when resets are coming, it becomes kind of frustrating trying to game it out. One week they're resetting like crazy so you're trying to burn tokens as fast as possible. The next week they're not resetting at all so you have to adjust your workflow to be more conservative.
> use your whole week up in time [...] optimal to burn [...] trying to game it out [...] you're trying to burn tokens as fast
This whole thing sounds crazy to me, why are you so focused on making sure you hit 0% usage left when it's supposed to reset? Why can't you just use what you have and if it resets, it resets, and if it doesn't, it doesn't?
I've calculated I can spend about ~12% of usage every day, more or less, this is my "budget". If it resets, then the "12% per day" gets reset for that day, but that's it. Sounds crazy to me that I'd "invent work out of nowhere" just to spend more usage, why on earth would I adopt such a workflow? Sounds like you're burning tokens just to burn tokens???
If you spend $200 per month that's for X amount of work (represented as tokens) you can complete total. But if there are random resets then there's a different work maximization strategy. It's fine if you don't agree with it but it's not "crazy" behavior; it's completely rational to want to maximize a resource. It's not inventing work out of nowhere because there's more work to do than there are tokens in a week to do it. This is your assumption and why you've framed it as "crazy".
> it's completely rational to want to maximize a resource
Sure, in a video game where you have one attribute and you can minmax a strategy just focusing on that, but that's not how real-life works.
You can't just stack pending work on top of each other, expect yourself to be able to stay equally on top of everything and have the same results as if you didn't. With this comes the consideration about the tradeoffs of "produce mediocre but large body of works" vs "produce high quality but small body of works".
Who's to say what's more "rational" or not, it's not a straight-forward calculation which culminates in "Must consume all available usage to maximize resource usage" like some robot, as we are not.
It's quite literally inventing work out of nowhere as you wouldn't put the agent to do that and forcing yourself to be conscious about that work until it completes, unless you actually had the usage available. Asked another way, wouldn't your workflow clearly change if you had unlimited usage available? You'd probably attack tasks/problems that you didn't consider actually spending time/energy solving.
> With this comes the consideration about the tradeoffs of "produce mediocre but large body of works" vs "produce high quality but small body of works".
Not really. The AI will output the same sort of code at a certain skill level regardless of speed, it's not a human so the above is a false dichotomy. Also, it's sort of strange that you're saying it's inventing work out of nothing, you've never heard of a backlog? In companies that can grow very long and the more usage means the more that can be tackled.
> You'd probably attack tasks/problems that you didn't consider actually spending time/energy solving.
Yes? But that's probably after the backlog is complete unless they are truly low hanging or high priority fruit. So not sure how that reasons with your point.
> The AI will output the same sort of code at a certain skill level regardless of speed, it's not a human so the above is a false dichotomy.
It'll output the same code given the same prompts yes, but you don't just accept whatever it puts out, it requires iterations before it's actually ready to be committed as none of the agents write perfect code on their first try. So, it's not a "false dichotomy", I'm just looking at larger things than "LLM does inference"
Ah, I see the confusion now. You are assuming one looks at the code at all and adjusts it to fit whatever style guide is needed. The parent you were initially talking to is talking about vibe coding, where whatever the agent spits out is accepted on the first try as long as it works for a given feature. Therefore the more usage one has the more features one can build.
Can you explain what you mean? These days courtesy of an addictive reset game OpenAI is playing, I can't find anything with frontier intelligence that's more cost efficient...
If they didn’t constantly reset, they’d be about the same as Anthropic.
Right now, I find that Grok offers better value, uses fewer tokens per turn, and makes better code. I haven’t tried Cursor because I don’t want to change editors again, but maybe I should try it…
The benchmark article we're replying to shows that Grok token usage is at least on par with the latest OpenAI models [1], and significantly cheaper per token:
Yeah, even without the resets, chatgpt subscription currently goes quite a bit further than an equivalent anthropic plan. The main reason to have an anthropic plan is to get access to Fable 5 if you feel the quality of output makes it worth it.
Not that I know of. AA's token use metrics (mentioned in this article) are indicative, however. They say explicitly here that the Grok models are notably token efficient. This is my experience.
I'm willing to pay 2x for a 10% smarter model. Intelligence matters that much (because 10% smarter probably saves, on average, several hours of human time).
For building a full stack custom CRM and media pipeline tool with video conversion, transcription, and indexing. Supabase, AWS, Meili, NextJS, GCS - lots of surfaces and planes.
4.8 basically couldn't do it, I abandoned the project as the fallback was, "current business processes".
With F5 it's been 4 weeks and almost ready for production release.
I have the same quality results with Fable. With just a brief prompt, it created a great static website with a beautiful animation of a workflow. Gemini's output was so poor that I closed the chat. And with Codex, the results were bad, so I discarded them.
Agree. Opus 4.8 could make nice little toy and demo apps. This app had a lot of surfaces and pipeline, Postgres, vector search, S3 -- couldn't handle that. Fable 5 is still a lot of work and you have to check it, but it really does perform at senior eng level. Shipping good size features daily.
I believe they are the only western provider that has Kimi K3 on a subscription plan today as well. I would love to ditch Anthropic and be on Kimi if there were a subsidized plan like that with ZDR
I believe it's one of the models you have to go in to your settings on and enable Chinese providers for to use. Could be mistaken. I wish there was a clear list on this.
I'm pretty sick of having to address incorrect points masquerading as being well-informed, helpful, or knowing. Please: stop it! Stop it at once. You are wrong. Stop talking like you know anything, because you have only mislead and provided incomplete or outright bad information. Stop hallucinating into this thread! It's rude and bad.
On the OpenCode Go workspace there is a huge box:
Providers
Control which providers are used for routing.
Enable models hosted in China <toggle-button>
If you turn this off, DeepSeek Flash & Pro both stop working:
Error: Provider request failed with HTTP 403: The latest version of this model is only available hosted in China and requires explicit opt in: https://opencode.ai/workspace/wrk_msh_is_a_liar_stop_bullshitting_please_be_real_example_url/go
Kimi K3, MiMo V2.5 Pro, MiniMax M3, Qwen3.8 Max all work (remarkably!) without the box ticked! That was a strong surprise for me.
My request stands: I would like to have up-front information on what models only have Chinese providers. This information is not available except by buying a plan and experimenting, currently. And I had to try each one to find out, which I wish could be avoided. I want to end this thread where I started it (now that I have done the work to cut through the din and noise and misinformation), with my original request: please OpenCode Go make the information about which providers have Chinese-only/non-China hosting readily available on your website.
My answer is still correct. You asked if you had to turn on Chinese models to use this. And no you don’t, the models are available by default. If they work when you manually turn off Chinese models is a different question.
There seems to be something strange going on with how it plays with the GHCP harness: I've experimented on a variety of inputs (code, plain text, literature search), and more than half the time it falls into an infinite text/tool call loop a la GPT-2. Which is a bit spooky as you're still being billed for those infinite loops! But when it works it works, and responses do seem a good deal cheaper than OAI/Anthropic equivalents, so hopefully they'll get it ironed out.
They don't do request based pricing anymore. Its just token based (1 credit = $0.01) plus some bonus credit based on which plan you subscribe. So for example a $39 plan get $70 of credits.
Yeah, I know, but "credit" translates differently because the models bill at different rates, which gets turned into "multipliers" (or at least, it did).
Have they converted entirely to transparent API rates + base allocation now? One of the reasons I left was that if I was going to be billed at API rates anyway, I'd just rather use the APIs. The value proposition still sucks for individuals now, when the other major providers are bundling at below-API rates.
I’d love a subsidized Kimi subscription too. The official Kimi subscription is always out of stock and doesn’t have great limits, while the K3 allotments on OpenCode and Cursor don’t seem to last very long either.
Kimi is expensive . Cursor with subscription is cheaper , grok 4.5 per task paid per tokens ( no subs ) is also cheaper .
If you willing to share to no zdr, meta is waaaaaay cheaper vs Kimi.
With recent offerings from spacex and meta , I hardly imagine why would you pay money to any Chinese vendor it’s not as cheap and it’s not as intelligent neither .
Maybe deepseek is an exception , but it’s only good for narrow use cases that probably goes into modal.com and other gpu + fine tune me easy vendors , not vanilla dumb but cheap model .
When I last used Cursor their subscription covered usage of ~$20 per month. Have they switched to a subsidized subscription model like ChatGPT and Claude?
