Claude Code Limits: What Changed and 15 Ways to Save Usage
Claude Code limits are the thing I get asked about most, and I've felt them personally: I use Claude Code every day to build StrideOps, a multi-tenant SaaS with over 2 million lines of code.
Claude Code limits are a five-hour session window plus a weekly cap. In March 2026 they collapsed for three reasons: tighter peak-hour limits, a promo ending and a caching problem. Anthropic has since fixed the bug, doubled five-hour limits and walked back the OpenClaw ban. The 15 habits below still save usage.
I read about 12 GitHub issues, went through the Reddit threads and tracked every public response from Anthropic staff for this video. Here's what happened, what changed after, and what to do now.
How Claude Code limits work
Two limits matter.
- The five-hour session window: the percentage bar on your dashboard. The timer starts with your first message.
- The weekly limit: total usage across all sessions and models. If you burn through it, you're locked out even with a fresh five-hour window.
That second one is where I was when I recorded this. My weekly limit was gone, so my open five-hour window was useless, and I'd gone through over $100 of extra usage credits in a couple of hours.
The plans in the video were Pro at $20, Max 5x at $100 and Max 20x at $200. Notice what's missing: the actual token numbers. Anthropic publishes multipliers, not limits. 5x of what? That's one of the main community complaints.
What happened in March and April
Ask yourself: from October to February, did you ever really hit your limits on the $100 or $200 plan? Probably not. I rarely did, and I was coding from the time I woke up to the time I went to bed. Then three things landed at once.
| Date, 2026 | What happened |
|---|---|
| March 14 | Anthropic announces double usage outside peak hours for two weeks |
| March 26 | Five-hour limits tightened during peak hours, 5 to 11 a.m. Pacific on weekdays |
| March 28 | The double usage promo ends |
| March 30 | Claude Code's product lead says they're investigating limits draining too fast. That post hit 4 million views |
| April 2 | Anthropic says tighter peak limits and bigger 1 million token sessions explain most of it, and that no bug was overcharging anyone |
| April 3 | Subscriptions are announced to stop covering third-party tools |
| April 4 | The cutoff lands at 12 p.m. Pacific, with less than 24 hours of notice |
The promo. My read: the 2x promo masked a lower base limit. When it ended, "back to normal" felt like a new normal nobody had seen before, and there was no explanation.
Peak hours. Anthropic said this only touched 7% of users. But 5 to 11 a.m. Pacific is when most US developers work, and 7% of that user base is a huge number of people.
The caching. Claude Code caches the parts of a long conversation that haven't changed, which makes them 90% cheaper. Users were reporting cache misses on almost every request, and costs rising 10 to 20x. One developer burned a five-hour window in 19 minutes on the $200 plan.
The OpenClaw ban
The head of Claude Code announced that from April 4, your subscription would stop covering third-party tools: OpenClaw, OpenCode, any outside harness. It had been a gray area. You'd need pay-as-you-go extra usage instead.
To soften it, Anthropic offered a one-time credit equal to your plan price, claimable until April 17, plus discounts of up to 30% on pre-purchased usage. On the $20 plan, that credit goes in about five minutes.
15 ways to save Claude Code usage
These are from the video, with notes where the product has changed since. I've rewritten each one so it's correct today.
- Update first. Run
claude update. Old versions were paying for old bugs, and the caching fix alone landed in 2.1.101. - Use `/clear` between tasks. Message 50 costs dramatically more than message one. One task, one session.
- Start fresh on big idle sessions. Don't reopen a huge conversation that has gone cold. It reprocesses all that context.
- Keep junk out of context. I suggested a .claudeignore file in the video, and said skipping node_modules, build output and lock files could cut tokens 40 to 60%. I can't find that file in the current docs, so use deny rules in your settings instead.
- Lower your thinking. I said to set
MAX_THINKING_TOKENSto 10,000. Newer adaptive models ignore that budget, so use `/effort` instead, and drop it when you don't need deep reasoning. - Prefer CLIs to MCP servers. In April, one MCP server alone could add 15,000 to 20,000 tokens to every message, and a CLI typically gave you a 70% reduction. MCP tool definitions are deferred by default now, so that overhead is far smaller, but Anthropic's docs still call CLIs more context efficient. Disable what you don't use with
/mcp. - Measure it. You can't fix what you can't measure. I used ccusage, which showed me days of $82, $100 and $252. Today
/usagealso breaks down where your usage goes and recommends fixes. - Trim your CLAUDE.md. It loads every session. My rule is 50 to 80 lines. Anthropic's docs say under 200, with specialist instructions moved into skills.
- Default to Sonnet. Use Opus when you need Opus-level reasoning, and Haiku for simple subagents.
- Watch command output. Pulling thousands of log lines into a session bloats your context, even through a CLI.
- Combine your messages. Sending separate messages can 3x your cost. One prompt where you can.
- Know the cache timer. I said five minutes in the video. On a subscription within your plan, the main conversation now gets one hour. Usage credits, API keys and subagents get five minutes. Step away for longer than that and your next message reprocesses everything.
- Be smart with subagents. Each one runs its own context with its own system prompt, CLAUDE.md and tools, so spinning them up can roughly 7 to 10x your usage compared with a single session.
- Watch Claude work. If it heads the wrong way, stop it. Otherwise you find out after you've spent 80% of your tokens to produce zero value.
- Protect your context. Don't paste a whole docs site in. Only give Claude what it needs, when it needs it. Protect your context or you get a dumb Claude and hit your limits faster.
I dropped one tip from the video: scheduling your coding around peak hours. The peak-hours reduction is gone on Pro and Max, so that one no longer helps.
Three smaller tricks are worth knowing. /statusline puts your context, cost and usage on screen. claude -p runs one prompt with no conversation history, which is cheaper for quick questions. And the --max-turns flag stops Claude running away with tool calls.
Was Claude getting dumber?
I said in the video that sometimes Claude felt dumber, making mistakes I wasn't used to, and I asked if it was just in my head.
It wasn't. Anthropic's postmortem confirmed three separate changes that hurt quality, including the default reasoning effort being lowered from high to medium between March 4 and April 7. All three were resolved by April 20.
The bigger picture
I knew this was coming. We were watching the golden age of subsidized AI start to end. Every AI coding tool runs the same playbook: launch with absurdly generous limits, let users build their workflow around them, tighten when the compute bill comes due. Cursor did it. Windsurf did it. Then it was Claude Code's turn.
On the $200 plan, a power user could consume up to $5,000 a month in compute. That's a 25x subsidy, and that math doesn't work forever. My problem was never that Anthropic did it. It was how they did it, with no announcement and no transparency.
Since then the limits have gone back up, so this wasn't a straight line. But the lesson holds. If you run a business on one of these tools, build your workflow so a limit change doesn't stop you.
Who this is for
This is for anyone on Pro or Max who keeps hitting a wall, and for business owners who rely on Claude Code daily. If you're on the API, the billing is different but the context habits are the same.
For the rest of my Claude Code breakdowns, go to the Claude and Claude Code hub. If you want the operating system I've built around these tools, the AI Operator Kit has five systems for content, leads, outreach, sales and building your own tools. My free 19-page optimization guide from this video is in the free library.
In this video
- 0:00Anthropic Bans Third-Party Claude Code Tools
- 1:39The April 4th Ban Explained
- 3:05How to Claim Your Free Credit
- 4:28How Claude Code Limits Actually Work
- 6:10The 3-Part Crisis That Broke Everything
- 9:04The Caching Bug: 10-20x Overcharges
- 10:06Anthropic's Response
- 12:00OpenAI Resets Limits While Anthropic Cuts Them
- 12:4915 Fixes to Stop the Bleeding
- 17:00Trim Your CLAUDE.md & Optimize Settings
- 19:12Sub-Agents, Caching Timeouts & Output Bloat
- 21:47Is Claude Getting Dumber?
- 23:32The Bigger Picture: Subsidized AI Is Ending
- 25:56Free 19-Page Optimization Guide
Full video transcript
5,618 words
Anthropic just banned every third-party claude code tool from your subscription. That includes Open Claw, Open Code, all of them cut off within less than 24 hours of notice on a Good Friday 2. And this is the latest in a series of moves that have been completely destroying the developer trust with Anthropic. $200 a month max subscribers getting locked out of their accounts in 19 minutes. limits quietly being reduced behind a fake promo. Cash bugs silently draining your tokens 10 to 20x
Faster. In this video, I'm going to break down a full story of what happened, show you everything anthropic isn't telling you and give you every fix that actually works. I put together a free resource doc right here. It's almost 20 pages long. And of course, this will be available in the description down below in our free StrideAI Academy. I've been covering clog code since day one on this channel. I read through about 12 GitHub issues filed about these rate limits. I
Went through different posts on Reddit, including ones with over 300 comments in just 6 days. I tracked down all the different tweets from all the different Anthropic employees to see who responded publicly. And I've been testing multiple different fixes to try to mitigate this usage drain that you've probably been experiencing lately. And let me tell you this, this is definitely not speculation. Everything that comes in this video comes from official sources, confirmed bugs, and real user data. And also too, this
Is 100% true. If you have been using Cloud Code, these limits are real, and you would have felt them in the last few days. All right, so [snorts] let's start off with some of the biggest news, which really just happened today, uh, which is from Boris. Well, it happened last night, depending on when you're watching this, and the creator of Claude Code basically tweeted out that starting April 4th, so now this would be today, uh, at 12 p.m. Pacific, your Claude
Code subscription will no longer cover third party tools. So, this is an issue for people who have been using things like Open Claw, Open Code, Nano Code, whatever the case is. really any AI harness that is a third party out there if you were using those with your Pro or Mac subscription that is officially done which it was sort of like a gray area to be honest. Um, you know, when we first saw OpenClaw drop, they were actually banning people for
Uh using their cloud subscription. And Boris says here in one of the threads that they don't plan on doing that. They plan on sending you an email if you are, I guess, caught, but you never really know with Anthropic. And you will have to pay extra usage, which is on a pay as you go to top up instead of your subscription. Now this is huge because it begs the question what does this mean for thirdparty AI harnesses in the future? You
Know of course right now Opus is really the best model out there but we are seeing some you know rapid development in the open source community as well as other tools. I mean we do have 5.4 codecs of course and now we're seeing Google's Gemma. Um so you know the competition is out there but I do know a lot of people including myself just prefer using clawed models. So to soften the blow, Anthropics basically offering a one-time extra credit usage which
Is equal to your subscription. So for example, if you're on the $20 plan, you get an extra $20, which you'll probably go through in like 5 minutes. Uh if you're on the $100 plan, you get $100 extra. If you're on the $200 plan, you get $200 extra. You have to claim it before April 17th, and it expires 90 days later. So, if you haven't done that already, go to your clawed desktop, go to your settings, go to your usage, and then
You should see a button where you can claim it. You can click on that. If it says it fails, just click it a couple times, and then I found that it usually loads and you'll see it on the bottom. You do have to enable that extra usage option at the bottom. So, just make sure that you do that. Now, they are also offering one refunds. So, if you do want a refund, I believe emails will be going out probably today. And
They are also offering like pre purchase extra credit uh discounts up to 30% if you buy for example $1,000 worth of usage. This is something that's been going on for a while and we're going to dive a bit into the actual timeline in just a second. As you can see, some anthropic employees like Lydia here was tweeting out about this on March 30th and people thought they were looking into it. People thought that this was going to be resolved. This was
Right around the time of that entire Claude Code codebase leak. But of course, things turned out a little bit differently. Now, before I tell you what went wrong, you need to understand how Claude Code's limits actually work because Enthropic really doesn't make it that easy to understand. So, there's really two types of limits that actually matter. The 5hour session window, which you're probably experiencing most of the time, and that's the percentage bar you see on your dashboard, and the timer starts
When you send your first message. It's what drained that one developer's usage in 19 minutes instead of five hours. And there's a separate weekly limit as well. So this is the total usage across all sessions in a week. All right, this is across all models, etc. So even if your 5h hour window resets and you burn through your weekly tokens, then you're locked out. You're out of luck. And that's exactly where I am right now. As you can see here, I
Have used all my weekly cloud code uh tokens this week. And you know, I still have currently the 5h hour window open, but I can't even use those because of course my weekly usage overrides that. And you'll even see down here, I've even went through over $100 worth of usage of these extra, you know, credit tokens that Anthropic gave to us in the last couple hours. So, it's a huge difference because I've been on the $200 a month plan and I
Rarely ever would really deal with any limits when I was coding every single day nonstop all day from the time I woke up to the time that I went to bed. So, now looking at the pricing, we of course have the free plan. We have the pro plan at $20. We have the max five times plan at $100, the max 20x plan at $200. And you notice anything that's missing? Well, that would be the exact numbers. Enthropic doesn't publish the actual
Token limits. They use vague multipliers. So 5x pro limits, 5x of what, right? Nobody really knows. And that's really what one of the main community complaints is. All right. So here's what happened. At the end of March, three things went down simultaneously that pretty much created the worst token crisis limit in Cloud Code history. And if you've been an active user of Claude Code, I want you to ask yourself in October, November, December, January, February, did you ever really hit your
Claude Code limits if you were on the $200 a month plan or even the $100 a month plan? Probably not. Most people didn't. I rarely did. And I was really, really using Claude Code as a power user. So, the limits were genuinely generous. Everything was really subsidized. I knew this was coming, but just the way they did it wasn't the best way. So, we can see here Anthropic tweeted out on March 14th that they are doubling the usage, right? So, they
Say, "A small thank you to everyone using Claude. We're doubling the usage outside of peak hours for the next two weeks." And Claude also too, the last 3 months has pretty much went mainstream, right? With the surge of open claw and all these different things, content creators posting about it, we see that more and more people are have been switching and migrating from chatbt to claude, right? So, this sounded generous, but here's what actually happened. They reduced the base limit behind
The scenes. So, the 2x promo pretty much masked the downgrade. So, while everyone was really celebrating this double usage, the underlying limits were actually getting cut and the promo ended on March 28th when we were supposed to go back to the old normal, but going back to the old normal was actually the new lower limits that you've never experienced before. So, that's why it felt so sudden and it was really no and there was really no explanation from them or notification
On their end. So here on March 26th, Tyreek from Enthropic tweeted out saying, "To manage growing demand for Claude, we're adjusting our 5-hour session limits during peak hours." So translation is during weekdays from 5:00 a.m. to 11:00 a.m. Pacific, which is 8:00 a.m. to 2:00 p.m. Eastern, prime coding hours, your token cost will cost more. And it's the same 5hour window, but it's going to drain much faster. So of course, the community went absolutely ballistic on Twitter. So here's a tweet
Right here. And it's a satirical retweet where it says, "Tyreek, I need to give our customers less for the same price. How can I obiscate that? Claude, how about you'll move through your 5 hour session limits faster than before?" And that's really what we saw. Now, this went absolutely viral, and Anthropic claimed that this only affected 7% of users. 5 a.m. to 11:00 a.m. is when most US devs are going to be working. So, claiming it only affects 7% felt tonedeaf.
But even still, 7% when you think of the amount of anthropic users is a huge amount. All right, so this is something most people don't really know about and it's one of the worst things. Claude Code uses prompt caching. So when you have a long conversation, parts that haven't changed get cached, so it's 90% cheaper. But there were bugs in the caching system that cause cash misses on almost every request. You can see working version here on the left side. So
Turn one costs 15 cents and then turn four costs 2 cents. the cash reads grow, cost drops, and then the broken version on the left where cash reads basically stay flat and costs go from 4 cents to 40 cents per turn. So, users were essentially paying like 10 to 20x more than they should have been. So, there were additional bugs as well. And a lot of people thought that, you know, oh, these are just the bugs. Uh, things will go back
To normal soon. But when the base limits were lowered, peak hours kicked in, and Cash Bugs inflated your cost 10 to 20x. So all in all, it wasn't just like one or two things. It was kind of like a few different things and it really formed the perfect storm of pretty much all your limits just being drained. All right, so on March 30th, you can see here Lydia tweeted out and she's a product lead at Claude Code, of course, basically saying,
"We're aware of people hitting usage limits way faster than expected, actively investigating." And you can see here, 4 million views, 13,000 likes. This thing went absolutely viral. And you can see here she also had some updates within this thread here and people were responding and whatnot. We also had Bora's tweet here basically saying, "Working as hard as we can to make this better. It's not easy growing at this rate." And basically said, "A number of significant improvements came out. Make sure
You're on the latest cloud version." And then on April 2nd, Lydia also tweeted out, "Thank you to everyone who spent time sending us feedback and reports. We investigated and we're sorry that this has been a bad experience. Here's what we found. Peak hour limits are tighter and the 1 million contact uh sessions got bigger. That's most of what you're feeling. We fixed a few bugs along the way, but none were overcharging you. We also rolled out efficiency fixes and added popups
And products to help avoid large prompt cash misses. Digging into reports, most of the fastest burns came down to a few tokenheavy patterns. Some tips on 4.6 is the better default on pro. Opus burns roughly twice as fast. Switch at session start. Lower the effort or turn off extended thinking when you don't need deep reasoning. Switch at session start. Start fresh instead of resuming large sessions that have been idle over 1 hour. Tap your context window. Long sessions cost more. So
Claude code autoco compact equals, you know, 200,000. I'm rolling out more efficiency improvements. Make sure you're on the latest version. If a small session is still eating a huge chunk of your limit in a way that seems unreasonable, run for/feedback and we'll investigate. And you can see here, people thought that this was tonedeaf. You can see here, let me get this straight. You increase the context window to 1 mil and make it default. And then it's on us to autoco compact
Back to 200k. And then you give us opus, but literally no one, even on the max 200 plans, can use it longer than 15 minutes. Is this a joke? Tell me this is April Fools. And this is rightfully so, just due to the fact of how Enthropic went about this. And while Enthropic was doing this, OpenAI was, you know, doing things like resetting limits for codecs for everyone multiple different times, doubling usage windows. So, a lot of people just switched from
OpenAI to Claude. And now a lot of people are switching from Claude back to OpenAI. And you can see here, these tweets are getting a lot of traction as well. So, 9,000 likes, uh, a million views for basically both of these tweets right here. So, you can see the sentiment shifting in real time as well. And the ironic thing is opening eye just signed a Pentagon contract and uninstalls were spiking. You know, you probably saw a lot of news about this.
There was like literally a whole quit GPT movement. Plus with the whole open claw claw code thing, anthropic traffic was basically jumping skyrocketing about 30%. So their GPU infrastructure really just couldn't keep up. But now people are really getting pissed off with these limits. So with that being said, how do we actually mitigate these, you know, issues, right? So, I put together this full document, this guide that you'll be able to get for free, 100% along with all the other resources
That we cover on this channel, and you can just grab this in our StrideAI Academy. It's 100% free. So, just join the school community down below. Grab this document and you can start implementing right away. So, on a high level, these are the fixes that we're going to cover. I'm going to cover some of the most important ones that you can start implementing right now to start mitigating this usage problem. So, the first one's pretty simple. Just make sure you update
Your cloud code. I know we saw that with Anthropic. It's Anthropics on Vice, but just run cloud update. Hashbugs were partially fixed in recent versions. So, if you're on an older version, you're probably paying more than what you need to. Next is use for/clear between different tasks. So, don't let conversations grow forever. Keep in mind, message 50 is going to cost dramatically more than your first message. So, think of it this way. One task for one session. Another thing is stop
Resuming sessions because this triggers a full context reprocessing and bypasses the cache. Always start a fresh session. Next is add aclaude ignore file into your actual project route. You could see potentially up to a 40 to 60% token reduction just from telling claude to skip node modules, you know, disc uh build lock files, etc. It really depends, of course, what you're doing, what task it is, but try to use this ignore file in your project if possible. All right, another thing
Similar to what Lydia mentioned is capping thinking tokens. So, set max thinking tokens to 10,000 in your settings.local.json and extended thinking defaults to 32,000 which is overkill for most tasks. So, here's the config. You could just copy and paste this into your settings file or just give this document or these certain areas to Claude and ask Claude to help you do this and optimize your setup. All right, next is try to replace MCP servers with tools like CLIs or just remove
MCP servers in general if they're not needed because one MCP server alone could add 15 16 20,000 tokens to every single message even if you never use it. So disconnect the unused ones. You can just do for/mcp or like I said replace them entirely with CLI. So a lot of different tools superbase or cell whatever the case is they have their own CLIs. GWS CLI. Now, CLI will typically give you a 70% token reduction. You can also check out tools like
CLI anything here, which is a GitHub repo where you could turn pretty much any tool if you have the code to a CLI and be able to give it to Claw Code to, you know, to control the specific app. All right. Next is to install CC usage. So, this is a free tool. It has over 12,000 GitHub stars. And this essentially shows you exactly where your tokens are going by session, by model, by day. And I always like to say you
Can't fix what you can't measure. So you got to always be tracking things. So you can see here I ran CC usage right here. So we can see like all my usage for these days right here. We can see how much I'm using a day. We can see like yesterday I used $82 then $100, $252. Uh you can see here $3,400. So you're going to want to be tracking all this. There's also a few other repos. So there's CLI continue. So
This is a repo for when you actually hit your limits. If you do hit your limits mid task, instead of waiting for those limits to reset, you can actually use this tool which will automatically switch to whatever other you know CLI tool that you want to use whether that be Gemini CLI, Codeex, Open Code, whatever the case is. You can just basically import that context over without reexplaining what you're doing. Now to scale right here is another CLI tool for tracking
Uh open code, claude code, open claw, really any CLI. As you can see here, it's another pretty well uh put together repo that you know all this like I said will be linked in that resource. So you can check this one out as well. And we also have this repo which is claude token efficient. So this is essentially one file drop it in your project. It's the claw.md file and it keeps responses tur and can reduce total tokens on output heavy
Workflows. You can check this out. You may want to of course tweak it specifically for your use case. Now speaking of claw.mds, you're going to want to trim your claw.md. So this basically loads into every single message. Every token in it gets charged every time you send a prompt. So keep it under 50 to 80 lines. And you can see our claw.md guide for full detailed breakdown. It's over 20 pages long. It's very in-depth and I did a complete video guide
Covering this uh yesterday. So, go check that out if you missed it. And this alone will help you out a bunch. So, defaulting to sonnet will definitely help you out. And only use opus if you need opus level reasoning. And you can also use haiku for your sub agents. You're also going to want to try to schedule your coding times around these peak hours. So, we cover these peak hours. Just make sure that you know ideally if you're doing it after
Peak hours on off peak or we're doing it on weekends you're going to save a lot of tokens. Another one is watching for your command output bloat. So say if you know you're running something to check the versel logs well if there's thousands of logs that you're pulling in for example then that's going to bloat your context. So when you're doing tool calls to different things even with the CLIs or whatever the case is just be conscious of that. Another one
That's easy to mess up on is sending a bunch of different separate messages because this can 3x your cost. So try to combine your messages into one prompt ideally if possible. And then keep in mind Claude uses prompt caching to avoid reprocessing unchanged context. But it has a 5minute timeout. So if you're in a session and it's been 5 minutes of inactivity, your next message will reprocess everything from scratch at full cost. So, if you're stepping away, consider either compacting the
Session, clearing a session, or just starting a new one. I know it kind of sucks because a lot of the time you do want to maintain all that context that you've been working on, but this is where we're at right now. This is what we have to do if you want to mitigate your usage limits. And be smart about using sub agents. you know, the more sub agents you spin up, that's going to roughly 7 to 10x your token usage opposed
To a single standard session because each agent runs its own full context window in separate instances. So that's going to have all the system prompts, cloud.mmds, tool calls, skills, whatever that you're loading into it already, additionally as context. And then also too, one thing that I do with the 5hour timeline, if I'm about to reach that limit, let's say I'm about to go to bed, it's like 2:00 a.m. and I see, okay, my 5 hour limit is, you know, in the
Next hour I get another 5 hours. Um, but I haven't really, you know, used too much tokens. Well, what I'll do is I'll just start doing a bunch of stuff to try to maximize and push the limits so that when I wake up, I have a fresh limit, but I got a bunch of stuff done right now. You do want to be conscious because I have done this to the point where as you can see for this week I'm already at
My weekly limit now right so you do want to keep those few things in mind now another thing you can do in claude is you can run forward slash status line and this is going to set up claude code status line UI so this is kind of like CC usage or those other repos I was showing you so you can see here mine has opus 4.6 six 1 million contacts. You can see we have our context percentage right here. The cost,
The number of lines we've uh created and removed, our 5hour usage right here, 0% because right now I'm using 100% of usage, which I've used 100% of the seven days. So now I'm on the actual pay per usage. And you can pretty much set this up with whatever metrics you want. There's a bunch of different ones. These are the ones I'm using. But once you actually run that forward slash status, you can just basically talk to Claude, ask it what options
You have and just basically tell it to set it up for you. And you can see here ours is even colorcoded and it changes based on usage goes from yellow, green, red, etc. Okay, one thing you could do, and this is probably not something you want to be doing all the time, but you can use print mode for quick tasks. Um, so you can just run this right here, uh, Claude-P, and this will pretty much just prompt Claude one time. You're
Not going to have conversation history or context build for this, but you're just going to get the answer. Going to be way cheaper than an interactive mode for questions, but it's probably not something you want to be using all the time. You can also limit a gentic loops with the flag-max turns to prevent Claude from running away with tool calls. And with that being said, you're also going to want to watch Claude as it's doing its work because I don't know
About you guys, maybe it's just in my head, I've also I don't know if it's recently or maybe just cuz I'm bloating my context with some things and you know, whatever the case is, but sometimes Claude is dumb. All right? Or sometimes I've noticed it getting a little bit dumber, right? Making mistakes that I'm like, why are you making those mistakes? I don't know if they're nerfing it at all or maybe it's just in my head. Let me know in the
Comments down below if you've been seeing anything like that. It's still amazingly smart, but of course, we're used to these very high standards and if anything changes, you kind of you're already accustomed to doing it that way, right? But yeah, watchclude as it works because it can make mistakes. If it does make mistakes, you're going to want to catch them opposed from it just going and going and then you figure out, oh damn, it did the wrong thing and I've wasted
Like 80% of my tokens to produce zero value. And then when you're prompting Claw, just in general, be mindful about what you're pasting into it. you know, don't just grab like the entire, you know, onepage docs of like some text stack and paste it all in there. You want to be smart with what you're actually giving it as context. We really dove into this in yesterday's video talking about progressive disclosure. So, I really want you to take that same mindset and
Philosophy and just paradigm when it comes to using cloud code in general. This is like the same way skills work. This is the same way your claude.md should work. Whenever you're working on a codebase, you really want to only give Claude what it needs when it needs it because you want to protect your context. If you don't protect your context, you're going to get a dumb Claude. If you don't protect your context, you're going to reach those limits way faster and
You're just going to get way less value. Definitely make sure to check that video out. All right, so I want to take a step back and zoom out for a bit. What's happening here with Claude Code? It really isn't unique. And I 100% knew this was going to happen. I was literally just thinking about it last week. I was like, damn, we're currently in the golden age of AI right now. And of course, subsidized AI, we're really watching the golden age
Of subsidized AI come to somewhat of an end. It's still there, but it's not going to be what it actually was. Uh, every AI coding tool really follows the same playbook. You know, you launch with an absurdly generous limit funded by VC money. you know, users build their workflow around it. Then you tighten your limits when cost becomes unsustainable. You know, you have more GPU costs, etc. And then backlash happens, but most users basically stay with it because they're locked. And
It's essentially like how drug dealers get their clients hooked on the drug. And then once you're addicted to the drug, it's like, "Hey, if you're a business owner and you built your entire business around Claude Code," and then they're like, "Okay, instead of $200 a month, you're going to have to pay $1,000 a month." There's going to be some that actually still do it. Even though yes, you can switch to OpenAI, Codeex, etc. At the end of the day, these providers
All eventually will do this because they're all extremely subsidized. Like we saw this with Cursor. I don't know if you guys remember, but Curser had the $500 fast request a month at $20. And you used to be able to use like a lot of credits on Cursor and I would rarely hit the limits and I was on like the 200 or 300, whatever it is, dollar a month plan. But but then of course things changed and there was backlash when cursor
Did this and wind surf did the same thing if you remember and now it's claw code's turn. The really big issue I have and the problem was just really the way they went about it. They didn't even announce it. They kind of tried to like finesse their way out of it in a way it seems. Uh and on the max 20x plan you're pretty much paying or you were paying $200 a month and if you were a power user you could
Pretty much consume up to $5,000 a month in compute. So that's like a 25x subsidization. At full usage, the API equivalent would be like 3500 a month. Uh, and that math just really doesn't work forever. And of course, it can kind of work because most people actually won't be like a huge power user. But if you think of the huge surge of users switching over with all these different things going on, I can see why they had to do it. But
Like I said, just the way they did it wasn't really as transparent with the dev community and didn't really build that trust. Like the playbook in it of itself isn't really them being greedy. It's essentially like the AI SAS growth playbook with like compute costs. Like you get the users hooked, capitalize on the market share, and then correct when the bill comes due. Anyways guys, that's pretty much it for some of the tips that you can start applying right away to
Try to mitigate this huge, you know, usage drain that you've probably been experiencing. There's a lot of other uh tips as well as you know just resources and whatnot that I linked to in this actual document. So it's about a 19page document and like I said once again this whole document as well as all the documents on this channel will be available all the free resources in our stride academy. So make sure to join the Stride AI Academy if you want
To connect with like-minded AI entrepreneurs enthusiasts as well as myself as we're growing this community. I really have a lot of cool stuff planned on the way for you guys for content. I'm, you know, really trying to pour a lot of time and effort and energy into this. And honestly, my mindset lately is just like give as much away for free as possible. You know, in the past, I've always been like, do I want to share this? Cuz I also run
Like a full marketing AI business, right? So, it's like, oh, do I need to hoard these secrets? Honestly, I really just want to share as much as possible with you guys and not hold anything back. So, if you want to stay up to date with all that and see the videos that I planned on the way that I'm going to share with you guys, then make sure to subscribe to this channel, like the video, comment down below what your thoughts are,
Uh, if you have any suggestions or improvements that to that I can make, let me know cuz I'm always trying to improve the quality of these videos. And I appreciate your time here watching these videos. It really means a lot to me. But other than that, guys, make sure to join the Stratai Academy so we can connect as well as you can get all these free resources. But I hope you enjoy this video. I will see you in the next video,
Guys. Do you have any crazy stories that you've been experiencing with Claude or Enthropic or CLI's Claude code with these usage limits or anything like that? Or do you prefer codecs? Do you prefer Gemini CLI? Let me know in the comments down below. Or if you have any other strategies that I personally have missed that could help others out with this whole fiasco, drop it down below. But I will see you in the next video, guys. Keep hustling, keep grinding, and
Of course, guys, accelerate your stride. Take care.