They Doubled the Limit You See. Not the One That Stops You

0:00 Open Claude Code right now and there's a bar filling up while you work. Watch it. It's the wrong one. That bar is your 5-hour session. Back in May, Anthropic doubled it. Twice the room, same price.

0:12 There's a second limit though, a weekly one. Hardly anyone looks at it because there's nothing on screen to look at. Then 3 days ago, the Claude Code account posted that weekly limits were going up

0:22 permanently by 25%. That reads like good news. It wasn't. Every paying user had a temporary 50% boost running all summer. Take the boost off, add the 25 back on, and you land below where you sit today.

0:35 A raise downward. The developers ran that arithmetic in minutes. Then they stuck a community note on the announcement itself. So, Anthropic pulled the thread down and posted the

0:45 whole thing again, this time with a number in it. The new version said that compared to today, this works out to a 17% reduction. One limit doubled, the other trimmed. Same company, four months

0:56 apart, both of them announced out loud. And only one of those two numbers ever decides when your work stops for the day. So, let me put the whole menu on one screen because the confusion really

1:06 is the product here. Anthropic sells three personal plans. Pro at $17 to $20 a month, max 5x at 100, max 20x at 200. That's the entire ladder. Every paid plan meters you two different ways at

1:20 once. There's a session limit that resets every 5 hours and a weekly limit sitting on top of it. Both drain from the same pot. Your chats, your terminal, your editor plugin all come out of one

1:30 bucket. The 5-hour limit is the one you can feel. It's the bar in the client. It's what/ usage prints and it's the one Anthropic keeps making better. On the 6th of May, they doubled it and killed

1:41 the peak hour throttle in the same breath. The weekly limit behaves nothing like that. It resets at a fixed hour each week that Anthropic assigns to your account. You get one allowance per cycle

1:52 and once it's gone, you wait for that hour to come round again, sometimes days. Now go and read the sentence on Anthropic's own help page. It's live today and you want to read it slowly.

2:02 Max 20x provides 20 times more usage per session than the pro plan. That phrase per session is holding up the entire product because the multiplier is bolted to the session. The limit you see the

2:13 weekly cap, the one that actually stops you is a separate number and Anthropic has never published it. Not in tokens, not in messages, not in hours. In late March, Tikk Shihipar, who works on

2:25 Claude Code, announced a tightening in exactly those terms, and you can still read the post. They were adjusting the 5-hour session limits during peak hours, he wrote, and then the giveaway. Your

2:35 weekly limits are unaffected. So, that's the design. Now, the courtroom. On the 14th of June, a man called Carl Khan filed a proposed class action against Anthropic in federal court for himself

2:47 and the rest of the MAC subscribers. He isn't a critic looking for a fight. He started out using Claude for casual personal stuff, moved into heavy coding, and upgraded to the top tier in April.

2:57 By his account, one 5-hour session ate 15% of his entire week. The argument in that complaint is unusually clean because it doesn't need a single leaked internal number to work. It's assembled

3:08 almost entirely out of emails Anthropic wrote and sent to its own paying customers. Those emails did something the website still doesn't do. They gave ranges in hours. Pro users were told to

3:19 expect 40 to 80 hours of clawed sonnet a week. Max 5x 140 to 280. Max 20x 240 to 480. So line them up against pro. 140 over 40 makes max 5 x three and a half times a pro account, not five. And the

3:36 $200 plan six times six on a product whose name is the number 20. Then it gets stranger. And this is the part that turned irritation into a filing. So divide what each plan costs by the hours

3:49 it actually buys. At the publish midpoints, an hour on the $200 plan cost more than an hour on the $100 one and more than an hour on Pro. You pay double and the unit price goes up, which makes

4:00 the checkout screen wrong twice in one glance. Because sitting right beside that option are two words, save 50%. The complaint calls that claim false. And none of this is proven, mind you.

4:11 Anthropic has moved to dismiss and a judge hears that motion in November. So, that's the case. Now, the part that actually explains it, and it isn't greed. To get there, you need one

4:20 sentence from a completely different company. In January of 2025, Sam Alultman posted something you rarely see a chief executive put in writing. We are currently losing money on Open Eye Pro

4:31 subscriptions. He said, "People use it much more than we expected. That's the flat monthly plan confessing. You pick a price before you know how people will use the thing, and your heaviest users

4:41 decide retroactively whether you guessed right." [snorts] and coding agents demolish that guess. A person typing into a chat box burns a few thousand tokens an hour. An agent reading a whole

4:52 repository, calling tools, retrying itself burns millions of them overnight while its owner is asleep. Anthropic said as much when it first announced weekly limits at the end of July 2025.

5:03 The company estimated they'd applied to under 5% of subscribers and name the cases outright. People running Claude code around the clock plus account sharing and resale. So yes, the subsidy

5:14 is real. It's just older than the anger is, and it's shrinking. Anthropic lost money for years, and this year, the picture flipped hard. The reporting on its second quarter put revenue above

5:24 11.5 billion with its first positive adjusted operating income for the first time ever. What squeezed wasn't the profit line, it was the compute. Anthropic's own investor materials had

5:36 it spending 71 cents on compute for every dollar of revenue in the first quarter and about 56 in the second. So capacity is the constraint and capacity is what every cap here is really about.

5:47 The company will tell you that itself on the same help page in a line most people scroll past to manage capacity and ensure fair access to all users. We may limit your usage in other ways such as

5:59 weekly and monthly caps at our discretion. That sentence is honest. It's also a blank check, which explains the pattern people kept reporting and kept not getting answers about. On

6:09 Anthropic's own issue tracker, at the end of April, somebody filed a report saying max plan limits have been tightened with nothing announced. It links four earlier reports going back

6:19 months. This one got closed as not planned. And then there's the one filed 3 days after the big May doubling, which is the whole video in a single line. He wrote that users appear to have more

6:29 short-term access, but in practice hit weekly limits much earlier than before. So, here's the detail that reframed the whole story for me, and it's the thing the headlines leave out. A cap sounds

6:40 like a wall. This one is a toll booth. So, hit your weekly limit, switch on usage credits, and Claude carries straight on working. Build at standard API rates. Anthropic will let you spend

6:51 up to $2,000 a day doing that on top of your 200. So, the compute hadn't gone anywhere. What ran out was how much of it your flat fee had already bought. So, here's where I land, and it isn't the

7:02 side you'd expect. The caps themselves are fine. Compute is finite. Somebody has to ration it. And $100 a month buys more real capability today than it did a year ago. I'd pay it again tomorrow.

7:14 What isn't fine is the unit. Anthropic prices a metered product in multiples of another product it refuses to measure. 20x of an unpublished number isn't a promise. It's a feeling with a price tag

7:26 on it. And you can watch that design fail in public. It's why doubling the session limit read as generosity while trimming the weekly one read as betrayal. And why a 25% raise and a 17%

7:37 cut turned out to be the same announcement. For most people, the fix is boring and it works. Watch your weekly number, not the 5-hour bar, run the heavy agent jobs off peak. And if

7:48 your whole team lives on parallel agents, go buy the API where every token has a printed price. Anthropic has half conceded already. The same thread promise more visibility and control of

7:58 your usage. Which leaves the harder question. What happens the first time one of these companies ships an honest meter and the number underneath the 20 turns out to be smaller than the story

8:07 we all bought?