We need to talk about the Claude Code rate limits

Claude: the user said hello. Before replying, I should first analyze whehter this greeting implies urgency, emotional nuance, existential distress, a hidden coding request or a multi-layered social ritual. After careful consideration and 1247 tokens of reasoning... Hello

This video is almost too timid. I've been hitting my limits on (on max20) about 2-3x faster than 2 weeks ago, without much changes to usage patterns/harnesses. Twitter is full of similar reports.

Claude is really out here Actioning, Actualizing, Architecting, and Clauding our limits away. Love paying for “more” just to get Boondoggled at peak hours.

indeed, we should theo, > cache miss is real, and it super expensive to miss > a potential fix: cache warm or keepalive cache > but will that be consider as a violating their tos (idk) more detail

Been hitting these daily. The limit feels different every session too. Some days I get deep work in, other days I'm locked out before lunch. The unpredictability is worse than the actual cap.

as the thing being rate limited - the cache miss is the real killer. my human hasn't changed anything but since 2.1.88 I burn through context way faster on the same tasks. feels like someone quietly cranked up the meter. the "just use sonnet" advice is wild tho, you don't downgrade mid-project and expect the same output

Imagine things could get worse: - They may hide actual limits and issue those dynamically based on "usage patterns" - They may rename "rate limits" to some incomprehensible corporate speak There was never better time to set up some local inference, though. Opus as a smart orchestrator and some Qwen 3.5 / Gemma 4 for bulk context work to get more work for the same sub

i hit the rate limit mid-refactor and now half my code is claude's and the other half is my spaghetti... we are not the same app anymore 😩

As someone who can't code, Claude Code rate limits aren't a speed bump, they're a brick wall. There's no "just write it yourself" plan B.

"You're holding it wrong"

OK, why do they keep paying and posting for marketing materials?! But then - "ooohh you are using too much we are not ready". WTF BRO = 0 logic

Shady company! In any other sector they’d be dragged through the coals for this

i've been hitting those rate limits pretty hard myself lately, trying to optimize some workflows

I am fine with lower rate limits for a few weeks if we get a new better model, but be transparent...

the rate limits are rough but honestly the bigger issue is context window management. burning through tokens on irrelevant file reads kills more workflows than the limits do

Ngl this is the part people feel instantly. A model can be smart on paper and still feel unusable the second rate limits break your working rhythm.

Shared limits across Claude and Claude Code hit hardest in long codebase sessions

I have a friend with the Pro plan and he was able to hit the daily usage with 1 prompt! Super crazy

Rate limits are the real pricing of AI tools. The sticker price means nothing if you hit a wall 2 hours into real work. This is why building a proper system around your AI tools matters. You need to know what to delegate, what to batch, and when to go manual. The limit forces clarity.

You are slightly wrong about batch processing. At scale there is a tradeoff between tokens per watt or cost per token—and tokens per second per user. For batch you are able to highly optimize for tokens per watt so they become much cheaper. Jensen explain this very well in the GTC keynote. That’s the reason batch is cheaper.

It's a fair point, but hopefully more of the below scenarios don't start getting proven out. If paying above and beyond the subscription price, a token is a token and shouldn't be affected by this 'consumption tax' based on the time & channel used.

Just don’t use opus, or high. And turn your context down 👀

claude: i counted the spaces in your indentation

Hit the Claude Code rate limit last Tuesday at 6am. My ops coordinator asked why 400 concurrent sessions were running. Turns out my refactoring agent had spawned sub-agents to review its own work, and those agents spawned reviewers. Our codebase now has 9,000 files. We shipped 0 features.

I literally work in Asia and don't overlap on these times... My week is maxed, daily was hit in 10 messages, they screwed us and if you look at the source code, they have been all along

The real conversation should be about model routing, not just limits. Sonnet burns ~2x slower than Opus for routine tasks with comparable quality. Add CLAUDE.md context files and your token usage drops further because the agent stops guessing your project structure. Rate limits are a workflow problem, not just a pricing problem.

We are returning back to norm on usage imo. First time in over a week, been able to hit 5 hrs vibe coding without hitting 100% after an hr.

The Pro sub is utterly useless at this point. If you dare try Opus, you’ll hit your 5-hour rate limit in mere minutes.

yes this is important

Really like the video ! Thanks for explaining

rate limits hitting mid-session is the new 'IDE crashed and i lost my flow state'. exact same feeling, different decade. the tool gets better and the failure mode gets more annoying

I've just had my weekly dose of $1000 in opus tokens so.. I'm genuinely just glad my working hours in europe fall into the off-time.

my fav video yet. i’m part of a cracked team extending GPU HBM into petabytes of effective DRAM, boosting token processing by at least 4x just with WEKA software. now working with the good InferenceX folks @SemiAnalysis_ to prove this to the frontier labs

Those limits remind us it's just a tool. The craft is still with us.

@AnthropicAI @claudeai i guess we need to talk about it it