Confused on the Max 5x vs Max 20x. I'm on the latter, and in my email it says:

> "Most Max 20x users can expect 240-480 hours of Sonnet 4 and 24-40 hours of Opus 4 within their weekly rate limits."

In this post it says:

> "Most Max 5x users can expect 140-280 hours of Sonnet 4 and 15-35 hours of Opus 4 within their weekly rate limits."

How is the "Max 20x" only an additional 5-9 hours of Opus 4, and not 4x that of "Max 5x"? At least I'd expect a doubling, since I'm paying twice as much.

▲ deviation 4 days ago | parent | next [-]

This makes sense if we compare compute cost instead of hours.

Transformer self-attention costs scale roughly quadratically with context window size. Servicing prompts in a 32k-token window uses much more compute per request than in an 8k-token window.

A Max 5× user on an 8k-token window might exhaust their cap in around 30 hours, while a Max 20× user on a 32k-token window will exhaust theirs in about 35 to 39 hours instead of four times as long.

If you compact often, keep context windows small etc, I'd wager that your Opus 4 consumption would approach the expected 4× multiplier... In reality, I assume the majority of users aren't clearing their context windows and just letting the auto-compact do it's thing.

Visualization: https://codepen.io/Sunsvea/pen/vENyeZe

▲ thomasfromcdnjs 5 days ago | parent | prev | next [-]

Would love more feedback on this, I will definitely downgrade from Max 20x if it is the case. Cost me $350 a month in Australia...

▲ akmarinov 5 days ago | parent | prev | next [-]

I upgraded to 20x because i was constantly running against Opus limits and now it seems the 20x is almost equal to the 5x in that regard

▲ lvl155 5 days ago | parent | prev | next [-]

This is why I stopped using the MAX. Downgraded to Pro and started using o3 and others via API. I really don’t need that many hours to game plan in the beginning. At most it will cost me $10 between o3, Gemini, and Opus per project. There are new model releases every couple of weeks and I’d hate to get stuck with just one provider.

▲ ImaCake 4 days ago | parent | prev | next [-]

The ambiguity here is awful marketing practice. This bitter pill would be much easier to swallow if it was a hard number instead of these vague ranges. It would serve Anthropic better too - telling people they only get 300hrs vs between 240-480 (which they will naturally evaluate as 240hrs) will mean less users leaving the platform.

▲ dawnerd 4 days ago | parent | prev | next [-]

They really need to to just a limit so you can see how much you've used, not some vague hours per week or whatever. Github copilot will tell you, you have 300 requests with sonnet a month, makes it really easy to know when you're blowing past without having to worry about how long something has run.

▲ yobid20 4 days ago | parent | prev | next [-]

Someone should do a study then file a class action if their marketing material is false.

	▲	gabriel666smith 4 days ago \| parent [-]
		I've been tracking usage in my first month of "20x" Max (which was, unfortunately, this month). Depending on how this usage is amortised (working days, which is what matters to me, or 5 hour periods, or I guess, now weeks..?) their marketing material has been varying degrees of false. This has ranged from 'a bit false' to 'extremely false'. That is true both on a relative scale ("20x") compared to my previous use of the $20 plan, but - more dishonestly, in my opinion - absolutely false when comparing my (minimal, single-session, tiny codebase) usage to the approximate usage numbers quoted in the marketing materials. The actual usage provided has regularly been 10% of the quoted allowance before caps are hit. I have had responses from their CS team, having pointed this out, in the hope they would _at least_ flag to users times that usage limits are dramatically lower so that I can plan my working day a little better. I haven't received any sort of acknowledgement of the mismatch between marketing copy and delivered product, beyond promised "future fixes". I have, of course, pointed out that promised and hypothetical future fixes do not have any bearing on a period of paid usage that exists in the past. No dice! I'm, unfortunately, a UK customer, and from my research any sort of recourse is pretty limited. But it has - without question - been one of the least pleasant customer experiences I've had with a company in some time, even allowing for Anthropic experiencing extremely high-growth. Claude Code Router has been a Godsend for my usage level. I'm not sure I can justify the time and effort to care and pursue Anthropic's functional bait-and-switch offering more than I already have, because being annoyed about things doesn't make me happy. But I completely second this: it's not acceptable to sell customers a certain amount of a thing - then and deliver another - and I hope US customers (who I believe should have more recourse) take action. There are few other industries where "it's a language and compute black box!" would be a reasonable defence, and I think it sets a really bad precedent going forward for LLM providers. One might imagine that Anthropic's recent ~$200m US gov contract (iirc) might allow for a bit of spare cash to, for example, provide customers with the product they paid for (let alone refund them, where necessary) but that does not seem to be the case. It makes me sad to see a great product undermined like this, which is, I think, a feeling lots of people share. If anyone is actually working towards wider recourse, and would find my (UK) usage data useful, they're very welcome to get in touch.

▲ foota 5 days ago | parent | prev [-]

You're paying for prioritization during high traffic periods, not for 2x usage.

▲ strictnein 5 days ago | parent [-]

That's not what they claim:

https://www.anthropic.com/pricing

   > Max
   > Choose 5x or 20x more usage per session than Pro*
   > Higher output limits for all tasks
   > Priority access at high traffic times

That first bullet pretty clearly implies 4x the usage and the last one implies that Max gets priority over Pro, not that 20x gets priority over 5x.

▲

foota 5 days ago | parent [-]

That is sort of what it implies, but I don't think that's what's actually happening on the backend. I was looking at this yesterday though and I agree that it's all a bit hand-wavy. I feel for them somewhat though because it's hand-wavy because it's a difficult problem to solve. They're essentially offering spot instances.

▲

jjani 4 days ago | parent | next [-]

It's not "sort of what it implies" - it's literally what it says.

> Choose 5x or 20x more usage per session than Pro*

If a recruiter tells you you'll be getting "20x more money per hour" at this new startup, and you go there and you get only 6x, you're going to have a very different tone than "you sort of implied 20x".

▲

jjani 4 days ago | parent | prev | next [-]

Or go here: https://claude.ai/upgrade/max. What does it say on top, on the actual pricing page?

Max plan

5x more usage than Pro $100.00/month + tax

Save 50% 20x more usage than Pro $200.00/month + tax

Especially with the "save 50%", if they're not actually offering 4x that of 5x, that's easily illegal false advertising in half the territories Anthropic's customers are located in.

▲

foota 4 days ago | parent [-]

I want to say again that I don't think their plan pricing is straightforward, but (at least when I was looking at it the other day) I came away with the (correct, imo) impression that the 5x and 20x were just marketing terms and I should take it with a grain of salt. I agree it's not literally what it sounds like.

I think the disconnect here is that the 5x or 20x is true within a single session (and you'll see their website seems to always say this, clearly their legal team went over it with a fine tooth comb). The above about weekly quotas etc., isn't within a single session so the 5 or 20x no longer applies.

▲

jjani 4 days ago | parent [-]

I'm sorry but that's gaslighting in the face of literally being shown the payment page that shows otherwise. "5x more usage than Pro" and "20x more usage than Pro" cannot be reasonably interpreted in the way you're doing. It doesn't matter how much you preface it with "I don't think it's straightforward".

> clearly their legal team went over it with a fine tooth comb

If their legal team in fact did, they'll have said "yup, that payment page won't fly if you'd get sued in Germany/Australia/anywhere with remotely decent consumer protections" and Anthropic just decided to roll with it anyway as they did while breaking millions of websites' ToS, pirating half the world's books, academic papers and more. As the chance they'll actually get sued and receive meaningful punishment is negligible.

You're either incredibly naive, are doing similar things and justifying it, or have a personal stake in this. There's no other explanation. You can look at my comment history and see that I'm generally nice and don't comment in this manner, but you're leaving no other option as this can no longer be framed as good faith.

	▲	foota 4 days ago \| parent [-]
		Are we reading different pages? Their pricing page: "Choose 5x or 20x more usage per session than Pro*". Per session is the important bit here. I do see that their sign up/registration page doesn't include the per session text, instead saying just "5-20x more usage than Pro". I do think the per session but is the part that makes it true, but I also don't think you're being skeptical enough. If something says 20x more I'm obviously going to look for 20x of what, and I think it's on the reader to realize what they're signing up for. I don't have any stake in this, anthropic, or AI; but I don't like it when people flame something that I think is understandable. It's certainly a bit of spin, but I don't wouldn't go so far as to say it's unethical etc.,.

▲

buzzerbetrayed 5 days ago | parent | prev [-]

Some say hand wavy where others say dishonest. You’re justifying their dishonesty because telling the truth would cost them customers.

Gross.