Anthropic optimiert die Nutzungsbeschränkungen von Claude, um die Kapazität zu verwalten
https://www.theregister.com/2026/03/26/anthropic_tweaks_usage_limits/
Anthropic optimiert die Nutzungsbeschränkungen von Claude, um die Kapazität zu verwalten
https://www.theregister.com/2026/03/26/anthropic_tweaks_usage_limits/
12 Kommentare
„Tweaking“ is putting things lightly. Earlier this week some Claude users noticed a DRASTIC change in how much they were able to use the platform, with some capping out their 5 hour allotted usage in as little as a single prompt. Most assumed it was a bug but as days passed Anthropic refused to acknowledge it, leading some to question if it was being done on purpose and others to claim that the entire affair was just user error, people using the system wrong and wasting their usage all on their own.
Anthropic finally admitted today that there was no bug and they had in fact been messing with things without telling anyone, apparently running a/b tests against certain users absolutely tanking their usage, and in some cases eating through unsuspecting users extra paid credits in the process.
The general consensus from most of the platforms community has been outrage, both at the fact that Anthropic not only hid the entire test from users but insisted that there was no such issue at all while doing it, and at how the new limitations make the platform unusable when people actually need to use it.
Not surprising. If you pay for claude code usage by the token you can easily burn $50-100 in an hour. Their max plan is $200 and probably allows for somewhere around $5k in raw token usage per month. Only power users that use their entire token budget reach that though. The alternative is losing customers to gpt codex.
This subsidisation cant go on indefinitely. At some point, they need to focus on profits.
Anthropic is hostile toward its customers. No refunds even when they’re responsible for burning through your tokens with errors, bugs, regressions, and now A/B tests.
90’s vibes when the dial up Internet was cheaper at night hours
literally paid for Pro last week the first time to try things out and have been wondering wtf was up with Opus being literally unusable, can fully confirm:
– the single message (through Opencode, 0 MCPs/SKILLS) burning up the whole 5-hour limit
– support saying ‚it works as intended‘
quite glad that I paid the subscription with a temp 1 month card so I don’t even have to worry about canceling
OpenClaw chews through tokens at an astonishing rate. Even with prompt caching and compacting.
You don’t have to use their API or services fortunately to use their models. Plenty of other providers and ways to get access to them with more generous limits. For example if you want to use claude code just use it with your own aws bedrock instance, super easy to configure
I just started using Claude and I feel like it’s pretty darn good with coding. Who else has similar or better results since this is seemingly gonna be a problem?
Man this annoyed the hell out of me. Stopped me from working on projects multiple times this past week.
+ to increase profits
AI companies waking up to the reality that their margins are more similar to retail companies than tech companies is quite funny.
It’s like going to McDonald with $5 and they tell you that it’s up to them to decide how much fries you’ll get