People should realize by now that they aren't giving out thousands of dollars worth of compute in a $20 subscription out of the goodness of their hearts. The base level plans aren't meant for any kind of serious work beyond the equivalent of Google searches. At best treat it like a trial.
Your cheap cynicism makes OpenAI's lower prices sound like a reason to avoid using them, but that's backwards! It's a better strategy to use them now while prices are low and switch later if necessary. Nobody knows who will have the best prices later, so better to decide later.
Aware of what, though? We don't actually know OpenAI's profit margins on inference. Do we need to know?
It's enough to know that the market is competitive and new models with better prices are released often. This means you should have a way to switch models.
But there's open models that are (pessimistically) 1 year behind? It's not like if openai decided to rugpull everyone, we'd be going back to writing code by hand and all the developers who forgot how to fizzbuzz would be screwed.
Open AI rents compute. It seems unlikely given the huge investments they have already spent that they can profitably match someone who is just selling GPU time.
There's an escape valve: I've had Claude stand-up AI-enabled features in my app, so we're much less dependent on Claude itself. The app uses cheap API calls to check code quality and run other "lessons learned" sweeps. Some of these become regular code in the end as well.
Same. Claude "loves" writing tools to make checks deterministic so much that I actually have to pull it back sometimes and point out that some things need to keep using LLMs. But even then, I have it "farm out" a lot of things to scripts using dirt-cheap models for things I don't need a Sonnet or higher level model for as well. The proportion that runs on the SOTA models keeps dropping.
Or to just give users enough usage to find categories in which it is useful for them.
I pay the $200/mo., and don’t regret it for an instant - I have a project manager, an executive assistant, a business analyst, a software developer and an international accountant, for an absolute song.
The audience for this sort of thing is currently quite limited. It may grow more in the future, especially if prices get out of control, but for now, it's a relatively small subset of people.
Sorry but this is a pretty delusional take. No, you cant run the same level or even close to that level of model on a mac, not even on a 512gb studio. You can run good models sure, but not models anywhere close to the capabilities of these ones yet, despite what idiots on Twitter keep spouting.
The "run locally" thing is a red herring. The fact that anyone can be an inference providers and offer them as a service is what we need to focus on. If they are "good enough" then market forces will stop Anthropic/OpenAI from charging monopoly rents.
> People should realize by now that they aren't giving out thousands of dollars worth of compute in a $20 subscription out of the goodness of their hearts. The base level plans aren't meant for any kind of serious work beyond the equivalent of Google searches. At best treat it like a trial.
I've said this countless times already; if OpenAI or Anthropic were serious about world domination they would offer an infinite $500 to $1000/month tier subscription. No limits; eat as much as you'd like buffet. Maybe limit concurrent connection (say 10 conns max in parallel) to limit abuse. This would actually allow regular users to run 24/7 agentic loops, massively speeding up deployment. Win win for everyone.
Once they capature enough marketshare prices and restrictions are both going to skyrocket. The difference between the subscription usage and API pricing are stark.
This is interesting to think about. There's so much competition among the frontier labs and from outside via open source it just doesn't seem obvious to me they could maintain elevated prices
You’re going to get multiple tiers (as we already are). You’re always going to pay top dollar for frontier models, but you’re going to find things highly discounted if you’re willing to move off the frontier. See GLM 5.3 Flash, for instance.
Ya, setting aside any judgement about what's "right", this strategy doesn't seem profitable. There's no real moat between one model/provider to another, switching is relatively easy. I don't understand how this is supposed to work for them.
It strikes me as similar to UPS / USPS / Fedex -- everyone uses the mail, and they mostly use whichever is cheapest for their requirements. I don't think there's much loyalty to specific services, and people are happy to switch between the options
just dont fall in love with goofy memory style features and its likely gonna be fairly easy to just plug and play whatever model for a ton of use cases.
I dont see a lot of love for weird memory like features on HN, but on provider subreddits its constantly talked about.
Say things are going to suck and they suck. You look brilliant.
Say things are going to suck, and they don't suck. Nobody cares because it doesn't suck.
Say things are going to be good, and they're good. A few attaboys for getting it right, but nobody cares.
Say things are going to be good, and they suck. You look like a moron.
Because of negativity bias everyone is leaning towards predicting DOOOOOOOOOOM. People aren't even consciously doing this, it's just a factor of the medium, because looking like a moron hurts way more than a few attaboys.
I think in about a year, we're going to see a scad of these ASICS like chat jimmy running year old models on dedicated hardware. Imagine racks and racks full of Astra but running at 15,000 tokens a second or whatever? Imagine swarms of them running the models we have today essentially for "free." That's where we're going to be. The bottleneck will be production, tbh, not demand.
"Hey Astra-Silicon, solve the Goldbach Conjecture!" Sure, it might take a few hours and be totally un-readable to a human being, but the 6m lines of Lean or whatever will be correct. Then what? What can we start doing then?
I mean, kind of yeah? But like, there are already people working on the chips, there are already people designing the next hardware, etc. I'm sure we're going to get to a plateau in capability soon-ish? Exponentials are actually all sigmoids. But what does that look like? If the plateau in capability is 100s of times smarter than the average person (arguably we're already there in many many but not all domains) then in 10 years time last year's reasoning model etched into an ASIC or some crazy monstrosity built out of FPGAs but for LLMs is probably way more than enough for 99.999% of use cases?
But yeah, doomerism is the dominant narrative of the day here right now. There's a sort of eschatological poisoning that's happening presently. Nobody can even seem to imagine a world where things get better. It's crazy. Maybe it's because I recently went through a major illness, maybe it's because I hit my head one-to-many times along the way? But I've never been more optimistic about the future than I am now.
This would need to be a collusion across most of the providers (which I think is likely to happen) otherwise OpenAI would get ditched for Anthropic or vice versa, depending on who raises the prices more. If it happens across the board then enterprises that can use Chinese models won’t have that many options but to pay up.
Session limits are obnoxious and make Codex much less useful for me. I tend to code in spurts when I find some time and the $20 weekly limit was reasonable for my side projects.
I'm now hitting the session limits which means I can either purchase another plan, upgrade, or move off of Codex.
The equivalent alternative is not "no hourly limits", it's simply less overall weekly allowance for Plus users, or higher prices overall. It's a way to balance a scarce resource, and maybe the least frustrating one?
People have a fair chance to learn resource management in 5 hour chunks, while not being limited to silly models, and burning through all tokens in the first hour of the week. Seems mostly good.
Can you talk more about what changed in the last two months? I use open router for when I run out of limits, and I’ve done better than buying an additional sub for my uses. I thought about looking at a more niche player, but it’s not a priority for me yet.
First of all, it still has the 5 hour limits, which is what this thread is about and what I think what most in this thread want to avoid.
Second, the offerings are subject to change at random. They advertised $60 of usage for $10/month (knowing most users wouldn't reach that). This was consistently the case up until early August, when the per-model "usage multipliers" started taking over. Some models give $15 of usage per month, others $30, still others remain at $60, and apparently one at $100 now?[0] Either way, I don't want to expend the mental effort to track which model is the best deal for capability and usage.
Right now I use Hyper[1] for $20/month and give $100/month to OpenAI. Not OpenAI's biggest fan, but the value is good right now, and that's what matters.
What does Hyper give you that OpenCode doesn’t? Is it just the removal of the 5 hour limit? I looked at Hyper’s docs but it wasn’t clear exactly how the fine points work. Is it $20 for as much as you want of any model they support? They don’t say that but they don’t say no limits anywhere that I could find. Been using OpenRouter with API prices and choosing cheap models to get work done. It works but a $20 all you can eat provider with access to good models like Kimi 3 and GLM 5.3 who is not in china and doesn’t retain or train on data would be perfect for me.
Hyper gives $12.50 worth of usage every 24 hours, totaling $375 per month. It resets at the same time regardless of when you start, unlike the 5 hour limits on other providers. Yes, it's only $20/month. It's also insanely fast; I've never gotten 75+ tokens/sec on K3 from any other provider.
My goodness, the complaining... Just get all your devs a $200 ChatGPT Pro (20x) plan. Yes, you lose the "team" component, but you gain so much more. And what's $200 against the salary of a good developer? It's absolutely inconsequential, even in far cheaper non-US salary regimes.
I dislike this approach. In my opinion it is much useful a 1x/2x billing based on hour like Deepseek does. Being unable to use it at the hours I need it makes me want to remove the service, not upgrading it.
This is one of the reasons I’ve built a resident daemon on top a local CC/codex. 5h sessions is a [lazy] way to shift the scaling responsibility onto a user, so they are staying with us for a while
I think it's reasonable for the Plus plans to have this. If you're doing any serious work you should be on the Pro plan. If you're a casual user, 5 hours is more than plenty.
In all of them I've seen it's a window that starts when you haven't been using it in a while and begin work OR when your previous window expired. The window starts with a set number of tokens and cuts you off when they're used, it is fully reset at the end of the window (it's not a sliding window).
When you haven't used any tokens for an extended amount of time it reverts to the point where whenever you send your first token is when the window begins.
Hm, it seems like it'd make sense to automate a script to send one small prompt at, say, 7 or 8am so that you get your reset around noon instead of having to wait until 5h or more into the workday.
I never quite understood why they picked 5h, it seems oddly arbitrary
Clearly, the prices and limitations will increase with time if they don't find an efficient way to perform training and inference of LLMs. It would be great to see those issues being seriously addressed and, eventually, being fixed for good. THAT would definitely make an important and practical difference. Not only economically but also scientifically.
I think the people that are saying that mean that the inference is massively profitable. From what I've seen, the margins are between 50 and 90% on an API call. It's the training that's very, very expensive!
As far as I understand, for both OpenAI and Anthropic, currently the majority of their GPUs are being used for training, with only a smaller portion being used for actual customer serving.
I’ve not heard anyone say that the whole thing is profitable. Companies like Anthropic have said that inference is profitable, but I assume that’s only on a steady state basis, which nobody in the industry has ever reached at this point. As far as I know, everyone is still burning cash with data center buildouts and overall training costs.
Controversial take but this works better for me than just the weekly limit. I have about 1-2 hours of sustained focus per 5 hour period so if I run out of tokens then I can take a break and start again in a few hours. Or just use the top ups that I’ve paid for in case I want to push through.
Having just the weekly limit meant I could blow through too much within the first couple of days. Fortunately, resets were raining down during that time.
(I know I could just vibe code a tracker that splits up my usage into arbitrary periods and lets me know when to take a break. Maybe if they remove the 5-hour limit again.)
I definitely see the pricing model. I can afford $100, maybe $200, but not $3000.
It's enough to know that the market is competitive and new models with better prices are released often. This means you should have a way to switch models.
I wouldn't be surprised if every $10 you add to the subscription cost halves your audience.
I pay the $200/mo., and don’t regret it for an instant - I have a project manager, an executive assistant, a business analyst, a software developer and an international accountant, for an absolute song.
Their goal is to create AGI and replace humans with something resembling the plot of Horizon Zero Dawn and its sequel but more extreme.
I've said this countless times already; if OpenAI or Anthropic were serious about world domination they would offer an infinite $500 to $1000/month tier subscription. No limits; eat as much as you'd like buffet. Maybe limit concurrent connection (say 10 conns max in parallel) to limit abuse. This would actually allow regular users to run 24/7 agentic loops, massively speeding up deployment. Win win for everyone.
It strikes me as similar to UPS / USPS / Fedex -- everyone uses the mail, and they mostly use whichever is cheapest for their requirements. I don't think there's much loyalty to specific services, and people are happy to switch between the options
I dont see a lot of love for weird memory like features on HN, but on provider subreddits its constantly talked about.
On the internet you have 4 possible outcomes.
Say things are going to suck and they suck. You look brilliant.
Say things are going to suck, and they don't suck. Nobody cares because it doesn't suck.
Say things are going to be good, and they're good. A few attaboys for getting it right, but nobody cares.
Say things are going to be good, and they suck. You look like a moron.
Because of negativity bias everyone is leaning towards predicting DOOOOOOOOOOM. People aren't even consciously doing this, it's just a factor of the medium, because looking like a moron hurts way more than a few attaboys.
I think in about a year, we're going to see a scad of these ASICS like chat jimmy running year old models on dedicated hardware. Imagine racks and racks full of Astra but running at 15,000 tokens a second or whatever? Imagine swarms of them running the models we have today essentially for "free." That's where we're going to be. The bottleneck will be production, tbh, not demand.
"Hey Astra-Silicon, solve the Goldbach Conjecture!" Sure, it might take a few hours and be totally un-readable to a human being, but the 6m lines of Lean or whatever will be correct. Then what? What can we start doing then?
But yeah, doomerism is the dominant narrative of the day here right now. There's a sort of eschatological poisoning that's happening presently. Nobody can even seem to imagine a world where things get better. It's crazy. Maybe it's because I recently went through a major illness, maybe it's because I hit my head one-to-many times along the way? But I've never been more optimistic about the future than I am now.
I'm now hitting the session limits which means I can either purchase another plan, upgrade, or move off of Codex.
People have a fair chance to learn resource management in 5 hour chunks, while not being limited to silly models, and burning through all tokens in the first hour of the week. Seems mostly good.
What's not good for the customer is the constant change of rules. It is bait & switch.
Second, the offerings are subject to change at random. They advertised $60 of usage for $10/month (knowing most users wouldn't reach that). This was consistently the case up until early August, when the per-model "usage multipliers" started taking over. Some models give $15 of usage per month, others $30, still others remain at $60, and apparently one at $100 now?[0] Either way, I don't want to expend the mental effort to track which model is the best deal for capability and usage.
Right now I use Hyper[1] for $20/month and give $100/month to OpenAI. Not OpenAI's biggest fan, but the value is good right now, and that's what matters.
[0] https://opencode.ai/docs/go/
[1] https://hyper.charm.land
A better approach if they want to balance the load is having higher usage or lower usage consumption at different times of day.
The 5h limit hurts the most on the $20 plan because the limit is already very small (5h is ~15% of your weekly).
When you haven't used any tokens for an extended amount of time it reverts to the point where whenever you send your first token is when the window begins.
I never quite understood why they picked 5h, it seems oddly arbitrary
As far as I understand, for both OpenAI and Anthropic, currently the majority of their GPUs are being used for training, with only a smaller portion being used for actual customer serving.
I've had few weekends where I spend the full week credits over night then have nothing to do for a week
Having just the weekly limit meant I could blow through too much within the first couple of days. Fortunately, resets were raining down during that time.
(I know I could just vibe code a tracker that splits up my usage into arbitrary periods and lets me know when to take a break. Maybe if they remove the 5-hour limit again.)