A 300 megawatt deal, a $180 a month decision, and the strange feeling of having a vendor solve your problem 12 minutes before you were about to pay them more.
I had the Anthropic billing page open. Two columns. Max 5x on the left, what I was paying. Max 20x on the right, what I was about to pay. The difference was $180 a month. The reason for the difference was that my Tuesday night Claude Code sessions had started bumping into the peak hour throttle, and the throttle was making the agent loop sluggish in the exact moments I needed it to be fast.
I was on the upgrade page because I had decided, two days earlier, that another $180 a month was a fair price for not having to feel the throttle. I had walked away from the page once already. I had come back. The cursor was hovering over the Confirm button.
A friend texted me a CNBC link.
I read the headline. I read the subhead. I closed the upgrade tab.
The headline said Anthropic had signed a deal with SpaceX for 300 megawatts of compute capacity at the Memphis Colossus 1 data center. Two hundred and twenty thousand NVIDIA GPUs coming online within a month. The blog post Anthropic put out alongside it said the new capacity would lift peak hour reductions on Pro and Max accounts and raise per request limits on Opus.
In other words, the throttle I was about to pay $180 a month to outrun was about to ease on its own. Within 30 days. For free.
This piece is about why I think that timing matters more than the headline lets on.
The throttle was telling us something
For the last 60 days, every Anthropic pricing headline read like a tighter limit. April 21, the brief disappearance of Claude Code from the $20 Pro plan. Late April, the weekly caps creeping down. Early May, the peak hour reductions that pushed me to consider Max 20x in the first place.
I had read those moves the way I think a lot of users read them: as repricing. As Anthropic discovering that the bundle did not work at current usage levels and starting to walk it back. The implicit story was about cost.
The SpaceX deal tells me the story was always about capacity. Not what the inference cost, but how many requests the existing fleet could serve at peak. When you cannot serve every customer who wants in, you do not raise prices. You throttle. The throttle looks like a pricing problem from outside. Inside, the throttle is a queue.
A 300 megawatt deal solves a queue. Repricing solves a cost problem. The fact that Anthropic chose to spend a number with three commas in it on capacity, rather than push a price increase to the existing fleet, tells you which problem they thought they had.
The implication for those of us watching the pricing page: the moves of the last two months were tactical, not structural. They are reversible. The SpaceX deal is the reversal.
What changes for the Max 5x holder
I am going to be specific because vague is what made me almost hit Confirm.
Max 5x today gets me roughly 5x the messages and Claude Code usage of Pro, with peak hour reductions kicking in around 4 PM Eastern on weekdays. The reductions are not a hard cap. They slow my agent loop down by about 30 to 40 percent during the affected window. On a Tuesday night when I am in the middle of a refactor, that 30 to 40 percent feels like sand in the gears.
Max 20x removes most of the peak reductions and lifts the message ceiling four times over. The price is around $200 a month, give or take depending on the billing cycle.
The SpaceX capacity comes online in 30 days. According to Anthropic's own messaging, peak hour reductions on Max accounts ease as that capacity lights up. The question for me, sitting on the upgrade page, became: am I willing to pay $180 a month for the next 30 days of unthrottled access, knowing that by month two the throttle on Max 5x will likely ease anyway?
I closed the tab. I am going to let the 30 days play out. If by mid June my Tuesday nights still feel like sand, I will revisit. If they do not, I just kept $180.
I am not telling everybody on Max 5x to do the same. I am suggesting everybody on Max 5x should pause before clicking Confirm. The new capacity is real and the timeline is short.
The wrappers should be more nervous than they look
Cursor, GitHub Copilot, Windsurf. They all resell Claude under their own pricing. They all spent the last two months absorbing the same throttle I was feeling. Their tactic: a pricing change of their own (premium request budgets, daily quotas, model dropdowns) that pushed some of the cost back onto the user.
When Anthropic's serving capacity doubles in 30 days, the wrappers lose part of their pitch. The pitch was something like: we route around the capacity issues so you do not have to. If there are no capacity issues to route around, the routing layer is doing less work, and the user is paying for that routing in a seat fee that just got harder to justify.
I do not think the wrappers go away. I think the conversation with their CFO gets harder in Q3. The numbers that mattered for the last 12 months were "we provide reliable capacity". The numbers that will matter for the next 12 are "we provide better routing than you can build yourself". The first one was easy to charge for. The second one is a much harder sell when the user can write 200 lines of TypeScript and have the routing themselves.
The orbital footnote
I want to mention this and then drop it, because I do not want it to swallow the piece. The SpaceX press release also said Anthropic had expressed interest in working with SpaceX to develop gigawatts of compute capacity in space. Not next year. Not anytime soon. The phrase: "expressed interest".
That sentence will spawn a thousand think pieces about orbital data centers. Most of them will be silly. The line shows up in the press release because the comparative cost of cooling at orbit gets interesting once you are spending hundreds of millions a year on terrestrial cooling, and SpaceX has the only credible path to mass orbit at a price that makes the math work.
I am not writing that think piece today. I just want the record to show that the line was in the announcement. In 18 months somebody will argue Anthropic hid it, and the receipts will say otherwise.
What I am going to do
Three things, all small.
Sit on Max 5x for 30 more days. Watch the peak hour throttle. If it eases, the upgrade decision is dead. If it does not, I revisit.
Log my Anthropic API spend daily, not weekly. The 30 day window is short enough that a weekly snapshot misses the inflection. Daily is cheap to set up. I want a clean before and after on the throttle change.
Send the upgrade page screenshot to the Glincker channel as a reminder. The thing I almost did made sense at the moment I almost did it. Twelve minutes later, the same decision read wrong. The lesson: pricing decisions made under throttle differ from pricing decisions made when the throttle is going away. Wait if you can.
The closing
I had the upgrade page open. The cursor was hovering. A friend sent a link. I closed the page.
The thing I want you to take, if you take anything: when a vendor announces capacity that solves the exact problem you were about to pay them more to outrun, wait. Not forever. Thirty days. The cost of waiting is one Tuesday night with a slow agent. The cost of upgrading is $180 a month, every month, until you remember to cancel.
Most of us are not going to remember to cancel.
If you are on Max 5x and were eyeing Max 20x for the same reason I was, the only honest take I can give you is the one I gave myself this afternoon. Watch the throttle through the first week of June. Decide then. The capacity will have arrived or it will not. Right now we do not know. Spending $180 to find out is not the right way to learn.
What is on your billing page right now that you were about to upgrade?
Written by **GDS K S* (thegdsks.com), building Glincker.*
If this was useful, follow me on X / @thegdsks. I write about the parts of the AI stack vendors keep off the pricing page.
Top comments (0)