This isn't like a gym membership where people join aspirationally. No one's new year's resolution is "I'm going to use o3 more often."
Do I read this correctly? Only 100 messages per week, on the pro plan worth a few hundred buck a month?!
Per their website: https://help.openai.com/en/articles/9793128-what-is-chatgpt-...
There are no usage caps on pro users (subject to some common sense terms of use).
I have a pro plan and I hammer o3–I’d guess more than a hundred a day sometimes—and have never run into limits personally
Wouldn’t shock me if something like that happened but haven’t seen evidence of it yet
Unless you use "free" GPT 4.1 like MS wants you (not the same as Claude, even with Beast Mode). And how long is that going to be free, because it feels like a design to simply push you to a MS product (MS>OpenAI) instead of third party.
So what happens a year from now? Paid GPT 5.1? With 4.1 being removed? If it was not for the insane prices of actual large mem GPUs and the slowness of large models, i will be using LLMs at home. Right now MS/Antropic/OpenAI are right in that zone where its not too expensive yet to go full local LLM.
As a pedantic note, I would say 'ration'. Things you hoard don't magically go away after some period of time.
Rationed/hoarded do imply, to me, something different about how the quantity came to be though. Rationed being given or setting aside a fixed amount, hoarded being that you stockpiled/amassed it. Saying "you hoarded your rations" (whether they will expire) does feel more on the money than "you ration your rations" from that perspective.
I hope this doesn't come off too "well aktually", I've just been thinking about how I still realize different meanings/origins of common words later in life and the odd things that trigger me to think about it differently for the first time. A recent one for me was that "whoever" has the (fairly obvious) etymology of who+ever https://www.etymonline.com/word/whoever vs something like balloon, which has a comparatively more complex history https://www.etymonline.com/word/balloon
Rationing suggests a deliberate, calculated plan: we’ll eat this much at these particular times so our food lasts that long. Hoard seems more ad hoc and fear-driven: better keep yet another beat-up VGA cable, just in case.
Counterexample: animals hoarding food for winter time, etc.
One could theoretically ration their rations out further... but that would require knowing the usage to the point to set the remaining fixed amounts - which is precisely whT's missing in the interface.
I just love this community for these silly things.
So, back to hoarding.
Reality is probably that there’s a backlog item to implement a view, but it’s hard to prioritize over core features.
It's even harder to prioritize when the feature you pay to develop probably costs you money.
Back to the conspiracy ^^
They should have an indicator, for sure. But I at least have been around the block enough to know that declaring “it would be easy” for someone else’s business and tech stack is usually naive.
One day a few of hours of prompting is fine, another you'll hit your weekly limit and you're out for seven days.
While still paying your subscription.
I can't think of any other product or service which operates on this basis - where you're charged a set fee, but the access you get varies from hour to hour entirely at the provider's whim. And if you hit a limit which is a moving target you can't even check you're locked out of the service.
It's ridiculous. Begging for a law suit, tbh.
What they could do is pay as you go, with pricing increasing with the demand (Uber style), but I don't think people would like that much.
Decided to give PRO a try when I kept getting terrible results from the $20 option.
So far it's perhaps 20% improved in complex code generation.
It still has the extremely annoying ~350 line limit in its output.
It still IGNORES EXPLICIT CONTINUOUS INSTRUCTIONS eg: do not remove existing comments.
The opaque overriding rules that - despite it begging forgiveness when it ignores instructions - are extremely frustrating!!
But at things I have no idea about like medicine it feels very convincing. Am I in hazard?
People don’t understand Dunning-Kruger. People are prone to biases and fallacies. Likely all LLMs are inept at objectivity.
My instructions to LLMs are always strictness, no false claims, Bayesian likelihoods on every claim. Some modes ignore the instructions voluntarily, while others stick strictly to them. In the end it doesn’t matter when they insist on 99% confidence on refuted fantasies.
Often they're better at recognizing failures to stick to the rules and fixing the problems than they are at consistently following the rules in a single shot.
This does mean that often having an LLM agents so a thing works but is slower than just doing it myself. Still, I can sometimes kick off a workflow before joining a meeting, so maybe the hours I've spent playing with these tools will eventually pay for themselves in improved future productivity.
If I sit down for dinner at an all-you-can-eat buffet, I get to decide how much I’m having for dinner. I don’t mind if they don’t let me take leftovers, as it is already understood that they mean as much as I can eat in one sitting.
If they don’t want folks to take advantage of an advertised offer, then they should change their sales pitch. It’s explicitly not gaming any system to use what you’re paying for in full. That’s your right and privilege as that’s the bill of goods you bought and were sold.
I also find it hard to believe 5% of customers are doing that, though.
For example, I don’t mind that Netflix pauses playback after playing continuously for a few episodes of a show, because the options they present me with acknowledge different use cases. The options are: stop playing, play now and ask me again later, and play now and don’t ask me again. These options are kind to the user because they don’t disable the power user option.
I also wouldn't consider my usage extreme. I never use more than one instance, don't run overnight, etc.
As I said, I have trouble believing this constitutes 5% of users, but it constitutes something and yeah, I feel Anthropic is justified in putting a cap on that.
The dark pattern isn’t the usage limit. It’s the lack of information about current and remaining usage.
The human brain is stupid and remarkably exploitable. Just a teensy little bit of information hiding can illicit strange and self-destructive behavior from people.
You aren't cut off until you're cut off, then it's over completely. That's scary, because there's no recourse. So people are going to try to avoid that as much as possible. Since they don't know how much they're using, they're naturally going to err on the side of caution - paying for more than they need.
I haven't yet run into this limit...
I have zero doubt that this is working exactly as intended. We will keep all our users at 80% of what we sold them by keeping them anxious about how close they are to the limit.
Hover on it on a desktop, it’ll show how many requests you have left.
I know entire offices in Bangladesh share some of these accounts, so I can see how it is a problem.
but if you use an agent and it tries to include a 500kb json file, yeah, you will light cash on fire
(this happened to me today but the rate limit bright it to my attention.)
For brief, low context interactions, it is crazy how far your money goes.
Very good point, I find it unlikely that 1/20 users is account sharing or running 24/7 agentic workflows.
The stat would be more interesting if instead of 1 in 20 users, they said x in y of users with at least one commit per business day, or with at least one coding question per day, or whatever.
I suspect this could be a significantly higher percentage of professional users they plan to throttle. Be careful of defining Pro like Apple does if you market to actual professionals who earn based on using your product. Your DAUs might be a different ratio than you expect.
… if you run a SaaS, generally somewhere from 1 in 5 to 1 in 20 users are using you for real, while the others are mostly not using you.
I can personally think of a few internally licensed products, announced with huge fan fare, which never get used beyond the demo to a VP.
I imagine there are lots of people like me who have a subscription to be aware of the product and do some very light work, but the "real" users who rely on the tool might be badly affected by this.
the principle: let's protect against outliers without rocking the behavior of the majority, not at this stage of PMF and market discovery
i'd also project out just how much the compute would cost for the outlier cohort - are we talking $5M, $100M, $1B per year? And then what behaviors will simply be missed by putting these caps in now - is it worth missing out on success stories coming from elite and creative users?
I'm sure this debate was held internally but still...
They undercharged for this product to collect usage data to build better coding agents in the future. It was a ploy for data.
Anecdotally, I use Claude Code with the $20/mo subscription. I just use it for personal projects, so I figured $20 was my limit on what I’d be willing to spend to play around with it. I historically hit my limits just a few times, after ~4hrs of usage (resets every 5hrs). They recently updated the system and I hit my limits consistently within an hour or two. I’m guessing this weekly limit will affect me.
I found a CLI tool (which I found in this thread today) that estimates I’m using ~$150/mo in usage if I paid through the API. Obviously this is very different from my payments. If this was a professional tool, maybe I’d pay, but not as a hobbyist.
I’m guessing that they did, and that that’s what this policy is.
If you’re talking about detecting account sharing/reselling, I’m guessing they have some heuristics, but they really don’t want the bad press from falsely accusing people of that stuff.
my point is that 5% still a large cohort and they happen to be your most excited/creative cohort. they might not all want to pay a surchage yet while everyone is discovering the use cases / patterns / etc
having said that, entirely possible burn rate math and urgency requires this approach
The announcement says that using historical data less than 5% of users would even be impacted.
That seems kind of clear: The majority of users will never notice.
that 5% is probably the most creative and excited cohort. obviously it's critical to not make the experience terrible for the 95% core, but i'd hate to lose even a minority of the power users who want to build incredible things on the platform
having said that, the team is elite, sure they are thinking about all angles of this issue
But those power users are often your most creative, most productive, and most likely to generate standout use cases or case studies. Unless they’re outright abusing the system, I’d lean toward designing for them, not against them.
if the concern is genuine abuse, that feels like something you handle with escalation protocols: flag unusual usage, notify users, and apply adaptive caps if needed. Blanket restrictions risk penalizing your most valuable contributors before you’ve even discovered what they might build
that's exactly what they have done - the minority of accounts that consume many standard deviations above the mean of resources will be limited, everyone else will be unaffected.
correct me if I'm wrong, it's not like we have visibility into the token limit logic, even on the 5hr window?
I regularly hit the the Pro limits 3 times a day using sonnet. If I use claude code & claude its over in about 30 minutes. No multi 24/7 agent whatever, no multiple windows open (except using Claude to write a letter between claude code thoughts).
I highly doubt I am a top 5%er - but wont be shocked if my week ends on a wednessday. I was just starting to use Claude chat more as it is in my subscription but if I can not rely on it to be available for multiple days its functionally useless - I wont even bother.
Can you share what you're doing? I've been experimenting with Claude Code and I feel like I have to be doing a lot with it before I even start seeing the usage warning limits on the $20/month plan.
When I see people claiming they're getting rate limited after 30 minutes on the $100/month plan I have a hard time understanding what they're doing so different.
For what it's worth I don't use it every day, so maybe there's a separate rate that applies to heavy and frequent users?
And I guess it'll go downhill from here. Anthropic, I wish you the best. Claude is a great tool at good value. But if you keep changing the product after my purchase, that's bad value.
Huh?? grep and sed do this for free; you don't need A.I. for that.
Unlike some other AI coding IDEs, Claude Code doesn't either keep your entire codebase in context, or use some sort of vector representation and vector search of your code. It basically uses grep to find a keyword it's looking for then reads the file (it can also just read a few lines of a file too) into its context.
It seems to be the "magic" of Claude Code..it's very UNIX like, uses existing tools to do what it needs to do instead of trying to re-invent the wheel, so to speak.
You very well might be a top 5%er among people only on the Pro rather than Max plan
Low danger task so I let it do as it pleased - 30 minutes and was maxed out. Could probably have reduced context with a /clear after every file but then I would have to participate.
I usually use Tasks for running tests, code generation, summarizing code flows, and performing web searches on docs and summarizing the necessary parts I need for later operations.
Running them in parallel is nice if you want to document code flows and have each task focus on a higher level grouping, that way each task is hyper focused on its own domain and they all run together so you don’t have to wait as long, for example:
- “Feature A’s configuration” - “Feature A’s access control” - “Feature A’s invoicing”
So, if rate limits are based on an overall token cost, it is likely that one will hit them first if CC reads a few files and writes a lot of text as output (comments/documentation) rather than if it analyzes a large codebase and then makes a few edits in code.
Just to nitpick: When the limit is a week, going over it does not mean losing access for a week, but for the remaining time which would assuming the limits aren't overly aggressive mean losing access for at most a couple of days (which you say is more palatable).
I wouldn't say you're too reliant, but it's still good to stay sharp by coding manually every once in a while.
Is it really necessary to use GPT to write a simple comment like that?
If I had used an LLM, maybe I wouldn't have misspelled "losing" not once but twice and not noticed until after the edit window. <_<
FYI many input methods (including my own) turn two hyphens into an em dash. The em-dash-means-LLM thing is bogus.
Well, not the entire week, however much of it is left. You said you probably won't hit it -- if you do, it's very likely to be in the last 36 hours (20% of a week) then, right? And you can pay for API usage anyway if you want.
Nothing about using an LLM removes skills and abilities you already had before it.
And yes, the goal might be to only use it for boilerplate or first draft. But that's today, people are lazy, just wait for the you of tomorrow
Just because you state it, it doesn't make it true. I could tell you that taking buses or robotaxis doesn't change a bit your ability to drive.
Taking a bus sometimes is fine, but that is missing literally the entire point that I made.
Now, maybe that is the future (no more/extremely little human-written code). Maybe that's a good thing in the same way that "x technological advancement means y skill is no longer necessary" - like how the advent of readily-accessible live maps means you don't need to memorize street intersections and directions or whatever. But it is true.
There was research about vibe coding that had similar conclusion. Feels productive but can take longer to review.
the moment you generate code you don't instantly understand you are better off reading the docs and writing it yourself
Funny story: The widespread of Knorr soup stock already made people unable to cook their own stock soup, or even worse, the skill to season their soup from just basic, fresh ingredients.
Source: my mom.
And just as with cooking: most people won't care - and the same goes with LLMs. It can be good enough... Less efficient? Meh - cloud. AI slop image? Meh - cheaper than paying an artist. LLMs to get kids through school? Meh - something something school-of-life.
I look around and see many poorly educated people leaning hard into LLMs. These people are confusing parroting their prompt output as knowledge, especially in the education realm. And while LLMs may not "remove skills and abilities you already had before it" - you damn sure will lose any edge you had over time. It's a slippery slope of trading a honed skill for convenience. And in some cases that may be a worthwhile trade. In others that is a disaster waiting to happen.
It will change. There will be FOSS models, once it no longer takes hundreds of millions of dollars to train them.
just three possible examples
But indeed it always struck me that some developpers decided to become Apple developpers and sacrifice 30% of everything they ever produce to Apple.
I would argue that it might a bit different though, because when doing iOS development it's possible that you don't lose you core skill, which is building software, and that you can switch to another platform with relative ease. What I think might happen with LLM is that people will lose the core skill (maybe not for the generation who did do LLM-less development, but some devs might eventually not ever know other ways to work, and will become digital vassals of whatever service managed to kill all others)
In exchange for 500% more paid users
Apple largely makes paying users exist that otherwise wouldn’t be there.
Apple is your business partner, doing marketing and distribution for you, and shares its user base. Bloomberg terminals provide real time data and UI to non-technical finance people. Github provides you Git hosting service so you don't need to setup and maintain servers. MATLAB (although there are Octave, Python and open alternatives) sells numerical computation environment to non-CS engineers. Xilinx is sells its hardware and dev tools. Game devs use Unity because they want to focus on gameplay and not game engine development.
These are all the examples of Division of Labor. This time, however, you have to pay for your core competency, because you cannot compete with a good AI coder in the long run. The value you provide diminishes to almost nothing. Yes you can write prompts, but anyone, even a mediocre LLM can write prompts these days. If you need some software, you don't need to hire SW engineers anymore. A handful of vendors dominate the SW development market. Yes, you can switch. But only between the 3 or 4 tech giants. It's an Oligopoly.
If we have FOSS alternatives, at least we can build new services around them and can move on to this new era. We can adapt. Otherwise, we become a human frontend between the client and the AI giants.
Sure, devs can still work without AI.
But if the developer who uses AI has more output than the one that doesn't, it naturally incentives everyone to leverage AI more and more.
And note that I objected online services, local LLM don't have the same issues.
Think of driving a car. If the shortest path (in term to time of travel) is through traffic jam, and there is a longer path where you can drive must faster, it's very likely that most people will have the feeling to be more efficient with the longer path.
Also the slow down of using LLM might be more subtle and harder to measure. They might happen at code review time, handling more bugs and incident, harder maintainance, recovering your deleted DB ;)...
I can see the impact on my own input both in quantity and quality (LLMs can come up with ideas I would not come up to, and are very useful for tinkering and quickly testing different solutions).
As any tool it is up to the user to make the best out of it and understand the limits.
At this point it is clear that naysayers:
1) either don't understand our job
2) or haven't given AI tools the proper stress testing in different conditions
3) or are luddites being defensive about the "old" world
```
The fundamental requirement for the LLM to be used is: don’t use agents or things like editor with integrated coding agents. You want to:
* Always show things to the most able model, the frontier LLM itself.
* Avoid any RAG that will show only part of the code / context to the LLM. This destroys LLMs performance. You must be in control of what the LLM can see when providing a reply.
* Always be part of the loop by moving code by hand from your terminal to the LLM web interface: this guarantees that you follow every process. You are still the coder, but augmented.
```
Not sure about you, but I think this process, which your source seems to present as a prerequisites to use LLM efficiently (and seems good advice to me too, and actually very similar of how I use LLM myself) must be followed by less than 1% of LLM users.
The deciding factor is not speed. It is knowledge. Will I be able to dish out a great compiler in a week? Probably not. But an especially knowledgeable compiler engineer might just do it, for a simple language. Situations like this are the only 10x we have in our profession, if we don't count completely incapable people. The use of AI doesn't make you 1000x. It might make you output an infinite factor of AI slop more, but then you are just pushing the maintenance burden to a later point in time. In total it might make your output completely useless in the long run, making you a 0x dev in the worst case.
So far almost no code I got from LLMs was acceptable to stay as suggested. I found it useful in cases, when I myself didn't know what a typical (!) way is to do things with some framework, but even then often opted for another way, depending on my project's goals and design. Sometimes useful to get you unstuck, but oh boy I wouldn't let it code for me. Then I would have to review so much bad code, it would be very frustrating.
If you meant goes down for good, then I'm sure it would be annoying for a few weeks for the FOSS ecosystem, just the time to migrate elsewhere, but there is not much GitHub specific we would really miss.
The 'take me from A to N' is a pretty broad problem that can have many different solutions. Is that comparable?
We can all see this end up in a oligopoly, no?
If it's affecting 5% of users, it might be people who are really pushing it and might not know (hopefully they get a specialized notice that they may see usage differences).
Given that I rarely hit the session limits I’m hopeful I won’t be affected, but the complete and utter lack of transparency is really frustrating.
Now your vibes can be at the beach.
Sorry, I'll just be "that guy" for a moment. Assuming that access is cut at a random time during the week, the average number of days without Claude would be 3.5. That's not reasonable as it's dependant on usage. So assume that you've always been just shy of hitting the limit, and you increase usage by 50%, then you'd hit the limit 4.67 days in. Just 2-3 hours shy of the weekend - a sort of reward for the week's increased effort.
Have a blessed Thuesday.
And... well, I am worried that I could do something stupid or accidental on Monday and Tuesday and then loose access with more than half of the week left, especially since, as everyone else has noted, Claude doesn't show you how close to the limit you are until you've almost hit it. (That said, I also appreciate that it doesn't warn me too early, or I would be constantly watching the usage tick up well before I actually need to worry about it.)
A couple of weeks ago, Claude got into some kind of weird loop where it just kept saying something like "now I need to add the code continuation" over and over and over again. I had auto-accept turned on, and it was chugging for several hours before I realized something was wrong and stopped it. Who knows how much usage that burned! Luckily, it happened late in the evening, so I knew my usage would reset the next day anyway. But IIRC it was also a Monday evening...
(2) I interpret this change as targeting people who are abusing the single Pro account, but using it more like a multi-developer business would maximizing the number of tokens (multiple sessions running 24/7 always hitting the limits). Anthropic has a business interest in pushing those users to use the API (paying per token) or upgrade to the $200/mo subscription.
(3) While I fear they might regularly continue to push the top x% usage tier users into the higher subscription rate, I also realize this is the first adjustment for token rates of Claude Pro since Claude Code became available on that subscription.
(4) If you don’t want to wait for the next unthrottling, you can always switch to the API usage and pay per token until you are unblocked.
I'm pretty sure they calibrated it so that only the people who max out every 5 hour window consistently get hit by the weekly quota.
as many other services did, and even some tangible products are implementing, the introduced limit will later on be used to create more tiers and charge you more for the same without providing anything extra. #shrinkflation
Do we even know the Anthropic financials? My guess is that they're probably losing money on all their tiers.
These limits will only get worse. Let this be a wake-up call for you to not put all your development eggs in the A.I. basket.
I feel like someone is going to reply that I'm too reliant on Claude or something. Maybe that's true, but I'd feel the same about the prospect of loosing ripgrep for a week, or whatever. Loosing it for a couple of days is more palatable.
Also, I find it notable they said this will affect "less than 5% of users". I'm used to these types of announcements claiming they'll affect less than 1%. Anthropic is saying that one out of every 20 users will hit the new limit.