It’s really hit and miss for me. Well defined small tasks seem ok. But every time I try some “agentic coding”, it burns through millions of tokens without producing anything working.
You mean SWE-1? I used it like a dozen times and I gave up because the responses were so bad. Not even sure whether it’s good enough for autocomplete because it’s the slowest model I’ve tested in a while.
Not my experience for slowness. For smartness I am typically using it for simple "not worth looking that up" stuff rather than even feature implementation. Got it to write some MySQL SQL today, for example.
At this point I'm afraid to ask, but I will do I anyways:
How do Claude's rate limits actually work?
I'm not a Pro/Max5/Max20 subscriber, only light API usage for Anthropic - so it's likely that I don't really understand the limits there.
For example, community reports that Anthropic's message limit for Max 5 translates to roughly 88k token per 5-hour window (there's variance, but it's somewhere in this 80-120k ballpark based on system load; also assuming Sonnet, not Opus). A normal user probably won't consume more than 250k token per day with this subscription. That's like 5M token for a month of 20 active days - which doesn't justify the 100 USD subscription cost. This also doesn't square with Anthropic's statement that users can consume 10000+ USD of usage on the Max 20 tier.
I'm clearly misunderstanding Claude's rate limits here. Can someone enlighten me? Is the 5-hour window somehow per task/instance and not per account?
With Anthropic's Claude subscriptions - while many people appear to use tokens as an idea of the usage limit, I doubt that's what is really used by Anthropic. Why do I say this? Well, there are multiple models, Haiku, Sonnet and Opus, we all know that Opus is the most expensive and burns through the usage limit of the subscription the fastest of all. I'd theorise that Anthropic have some kind of internal credit value (perhaps as simple as $ USD) which they allocate with some variance based on things like overall system load.
Anyway, my personal experience on Max 20x is that, with Opus at least, on a busy day in the past I can burn through between 150 to 200 million tokens in a day using Claude Code for development stuff. Split that up into 5 hour windows, and assume I'm possibly using 2 or 3 windows in a day, that still works out to a lot of tokens, well into the millions. So, the 88k tokens per 5-hour window on Max 5x, I'm not sure if it's really as small as that. Maybe the apparent reductions recently in usage limits have made it drop to around that ballpark. Originally I saw Max 5x as a heavy usage Sonnet plan, with Max 20x being a heavy usage Opus plan, however with the new and additional weekly usage limit happening on August 28th I think I'd see the plans as potentially moderate to heavy usage Sonnet for Max 5x, and heavy usage Sonnet with multiple concurrent agents for Max 20x.
TLDR: I strongly imagine that Claude subscription usage limits are based on some kind of internal credit value, perhaps $ USD, not specifically tokens, and depending which model you use is how fast this "credit" will be depleted.
The usage limits are currently for an account, based on a 5-hour window, from the first message that was sent in a new 5-hour window. From August 28th there's an additional weekly limit which looks like it will primarily make Opus usage restricted.
It's probably best to look at it as credit based, which map to a certain scale of particular tokens (ie. an Opus token takes 5x the credits of a Sonnet token).
For example, ISPs are not able to crawl your traffic if it's via HTTPS. I've worked on data sets gathered by major ISPs and it's scary how much they know about their users (especially if they also have a mobile phone with the same company). ISPs use such intelligence for personalised marketing (either for their own product catalogue or 3rd parties)
The domain isn't, the full URL is. (But content size etc probably still allow identification of an individual page on a small site, and the context of the domain is already valuable)
> Other reason, they are gaining money mainly because there are enough fools around running big corps, who can conveniently pass the buck if something goes wrong.
Well, this can be said for any big business software company. It's the same story with SAP, IBM and friends. Business software is simply not about spending money well.
People say that a lot, but I work with customers all the time who are very happy with their solutions. I don't disagree that some people hate them, or hate this or that ui, but I can think to many companies/interactions with people that love the products.
That's never really expressed places like here. It's just "not spending money well", when I think there are legitimate positives to a lot of business software.
I don't claim that I know the ultimate truth, but I worked for many years for one of those big business software companies. But from my experience, only the people who don't have to interact with such products in any form like it. (Who are often also making the decision to purchase it)
It happened only a few times that users of such products (analysts, admins, support, devs, ...) told me that they like this or that. And I still remember each one of those guys and girls. (Weren't that many)
I know what those companies charge for their software and maintenance. I've been involved in the development of those products for a couple of years. And my very personal opinion is: they're not worth their money. No matter what a customer tells the press how much easier their business is after purchasing product X from company Y.
As I said, it's my personal opinion on that matter.
Big companies are often are complicated mess, and I can emphasize with desire to simplify and streamline operation. There's some benefit with setting your infrastructure in such a way that there's a known party you can pay to fix things quickly, and that you can sue if they can't.
Often they also like to stay understaffed in IT, so having consistent vendors, and similar systems allows them to keep just a few guys around that have any clue how it works, and supplement with consultants and vendor support.
Yeah, I see your point. But I would emphasise dahart's argument that we, and especially law enforcement in a democratic society, should have higher standards.
Well trained and educated members of the police force shouldn't act like bullies in a high school - no matter if they think their job is important or not. Like everyone else at their job, they need to be professional. Probably more professional than most of us sitting in front of a computer.
I'm just staying that if you start with "I don't understand why our law enforcement standards are currently so low!" then you're probably not going to be helpful in fixing the problem.
Gotta diagnose first. Diagnoses comes with empathy and understanding.
I think that's not really the point here. If the story turns out to be true, then Microsoft is abusing its market power to lock down devices from vendors like Lenovo and friends.
Since Microsoft got into trouble with the EU before, they should know that this could potentially cause another investigation. Hands down, a big company is actively trying to make devices unusable with anything but their software. This is not in the interest of consumers, therefore probably illegal in some regions.
That being said, I think everyone has probably different expectations and workflows. So if that’s what works for them, who am I to judge?