Note: I’m the owner of CULP : Claude Usage Limits Plugin (a SaaS). I’ll show what I do with CULP and so: how YOU could do it on your own without buying anything from me. Pinky promise.
Let’s first start:
How tf did we get to this point? Buying for something we can’t even use for a whole week before being asked to pay more : And we wish we could.
Claude is good
That’s not a battle. I personally don’t care about GPT 6.7 or Claude Mythos whatever.
But, if we come back to Claude (or at least, some, including myself), it’s because we find that Anthropic is giving us some sort of value.
For me, that’s the case.
KPIs
To reduce your Claude usage limits, it would be useful to know WHAT is the measurement.
For Claude’s subscriptions there are many, and that’s WHY everyone is hitting them SO fast.
(the more you add rules, the more difficult it is to follow them).
So here’s the list (From BIGGEST impacting factor to least):
1. Subscription Plan
2. The model you are using
3. Claude’s output (the most important if you want to skip sections)
4. Claude’s reasoning effort
5. Cache miss (Especially on the big conversations)
6. Long sessions
7. Claude’s input
8. Your location (No Anthropic Does not discriminate)
Subscription Plan
That’s clearly the most well-known one,
So if you’re living in a cavern (OR maybe you are an AI that doesn’t know this information in his cutoff knowledge?)
Here are Claude’s subscription plans:
Free
That’s basically a joke, just to let you know that there’s one, it gives you some amount of Sonnet in chat mode. Nothing more: no Claude Code, no Claude CoWork.
Pro
Claude’s Pro plan is $20 US.
That’s the baseline to “start with Claude”.
Important information is that: Claude’s plan doesn’t actually give any MEASURABLE KPIs. They’ll always use Pro plan as a “comparison” (5x, 20x).
Max
Claude’s Max plan is in fact, 2 plans : Max 5x and Max 20x (multiplier is usage from Claude Pro baseline)
Max 5x is at $100 US and 20x at $200 US
That’s where usually MOST users will get some fair amount of Claude’s usage.
Source: Claude plans
The more you pay, the more you get. That seems fair.
But the more you pay, the less per usage it costs. In this “usage multiplier” the math is easy:
$20 baseline.
$100 (YOU pay 5x the pro plan) -> You get 5x (Wow! What a DEAL!)
$200 (You pay 2x from that point) -> + 4x from that previous level. WOWW !!!!!!!!!!!!!
That’s in fact smart, if we think marketing-wise.
And that could be a reason why OpenAI followed this convention.
But the interesting part is that:
Is it real ?
Yes. Pretty much. Even : If you wipe out complaints about usage limits on the whole internet, some people find that it’s more like a 6x factor.
So first answer is: If you want more Claude usage : Pay that juicy $200 plan.
That doesn’t help me get more out of Claude with my current plan!
Yep, I know! Let’s dive in furthermore.
The model you are using
Anthropic has NOW 4 classes of models:
Fable (More “powerful” than Opus)
Opus
Sonnet
Haiku
Source: Claude models overview
If you’re hearing about “Mythos”. That’s in fact the same thing under the hood as Fable. Fable has more classifiers because “more capable”.
Source: Claude Fable 5 and Claude Mythos 5
OK. I will not tell you : “Don’t use Fable”. It’s a good model. I love it too. But I want you to realize something:
“In the trust me bro benchmarks”:
Opus 5 now mostly sits on top of Fable while being 2x cheaper.
| Bench | Opus 5 | Fable 5 |
|---|---|---|
| Frontier-Bench v0.1 | 43.3% | 33.7% |
| GDPval-AA v2 | 1861 | 1747 |
| ARC-AGI-3 | 30.2% | / |
| BrowseComp | 90.8% | 87.4% |
| Humanity’s Last Exam (no tools) | 56.3% | 56.5% |
| Humanity’s Last Exam (with tools) | 64.7% | 63.9% |
| OSWorld 2.0 | 70.6% | 66.1% |
| DeepSWE v1.1 | 68.8% | 69.7% |
| FrontierCode v1.1 Main | 53.4% | 53.5% |
| AutomationBench | 26.0% | 17.4% |
| Legal Agent Benchmark Held-out | 11.7% | 13.3% |
| HealthBench Professional | 59.8% | 66.0% |
| BioMysteryBench (hard) | 49.4% | 46.5% |
| BioMysteryBench (human solved) | 90.1% | 89.0% |
Source: Introducing Claude Opus 5
In fact: Even where Opus 5 loses against Fable 5, it doesn’t justify a 2x pricing.
And more closely : Why would you want a “bio capable” model ? You would be blocked against those classifiers. And we’re talking about coding.
And Sonnet?
Of course this is a choice ! But, it depends on what tasks. Sonnet 5 is good. But, it’s not like 3.5 that was more capable than Opus 3.
| Bench | Opus 5 | Sonnet 5 |
|---|---|---|
| SWE-bench Verified | 96.0% | 85.2% |
| SWE-bench Pro | 79.2% | 63.2% |
| FrontierCode 1.1 Main | 53.4% | 42.7% |
| CursorBench 3.2 Max | 70.0% | 61.5% |
| DeepSWE v1.1 | 74% ±4% | 54% ±4% |
Sources: Introducing Claude Opus 5, CursorBench 3.2, DeepSWE v1.1
And on DeepSWE 1.1, Sonnet 5 costs $26.40 per task against $11.84 for Opus 5 (AT API PRICING, that doesn’t mean it reflects subscription usage)
Source: DeepSWE v1.1
It’s behind on every benchmark and this is strongly reflected in tasks.
But: Those are mostly Software Engineering tasks. And I want to shift focus to this:
Do you use Opus in chat mode ?
I would not let you say yes !
Exceptions for people who need Claude for certain professions. (Like marketing)
Do you use Opus for doing websites?
Even GPT 3.5 turbo was able to do HTML! What are you doing! And it’s so verbose, it outputs so many tokens..
Quick check for what model you could need:
| What you need | Sonnet | Opus | Fable | Justification |
|---|---|---|---|---|
| Lite chats | ✅ | ❌ | ❌ | Sonnet + Web is pretty good. |
| Frontend Design | ✅ | 🟡 | ❌ | That’s a big trap in here : If you don’t know on your own what you want, you would be tempted to use Opus, In fact, Opus could do a DESIGN.md (from Google Labs standard) and Sonnet implements that. |
| Mobile Apps | 🟡 | ✅ | 🟡 | Mobile apps move fast, and building an app needs some sort of software engineering knowledge. But just to set it up. After this, with strong architecture, Sonnet can do the job. |
| Marketing | 🟡 | ✅ | ✅ | In fact, marketing is a challenge that the model can’t have seen in its training (every business is different). And having the ability to understand challenges and clients is a more difficult thing than doing tasks. |
| Brainstorm | ✅ | ❌ | ❌ | Need a model that asks you questions ? Even a local model could do it. Even: that could be a good idea : If you need to explain something to someone dumb, that sharpens your brainstorming skills by a lot. |
| Legal | ❌ | ❌ | ❌ | Use Claude for Legal. |
| Making a game | ✅ | ✅ | ✅ | Not every model could do the same tasks, they do have all their strengths and weaknesses, and combining them is a good idea. |
| Building desktop applications | 🟡 | ✅ | ✅ | For Sonnet, I would say it depends on what you wish to make. But it’s limited to some electron-like apps. |
| Making your own compiler | ❌ | ❌ | ❌ | Don’t do that. Or you’ll go to jail. |
| Something else | 🟡 | ✅ | ✅ | Plan with Opus. Ask Opus to check on web for benchmarks for your specific needs, LLMs love trust me bro benchmarks and will point out exact needs. |
Sources : Trust me bro
So: Use the right model for the right task, and if you are too lazy, Opus 5.
Claude’s output
Okay, that’s where my SaaS actually infer at the most.
Anthropic never did state it for subscriptions but, if we connect the dots:
At API pricing:
| Model | Base Input Tokens | 5m Cache Writes | 1h Cache Writes | Cache Hits & Refreshes | Output Tokens |
|---|---|---|---|---|---|
| Claude Fable 5 | $10 / MTok | $12.50 / MTok | $20 / MTok | $1 / MTok | $50 / MTok |
| Claude Mythos 5 (limited availability) | $10 / MTok | $12.50 / MTok | $20 / MTok | $1 / MTok | $50 / MTok |
| Claude Opus 5 | $5 / MTok | $6.25 / MTok | $10 / MTok | $0.50 / MTok | $25 / MTok |
| Claude Opus 4.8 | $5 / MTok | $6.25 / MTok | $10 / MTok | $0.50 / MTok | $25 / MTok |
| Claude Sonnet 5\ | |||||
| through August 31, 2026 | $2 / MTok | $2.50 / MTok | $4 / MTok | $0.20 / MTok | $10 / MTok |
| Claude Sonnet 5\ | |||||
| starting September 1, 2026 | $3 / MTok | $3.75 / MTok | $6 / MTok | $0.30 / MTok | $15 / MTok |
| Claude Sonnet 4.6 | $3 / MTok | $3.75 / MTok | $6 / MTok | $0.30 / MTok | $15 / MTok |
Source: Claude API pricing
The Claude’s output pricing is 1 for 5 against output.
Even if Anthropic didn’t state it, that means : If you reduce what Claude outputs, it reduces usage limits ?
Yep!
I was asking myself: “How to reduce that freaking output!”
And actually, here’s a clever solution from old dark times (before 2022) :
Emmet
Emmet lets lazy programmers write HTML in abbreviations:
main>section.s$*20>(header>h1{SECTION $}+nav>ul>li*10>a{Link $})+div.grid>article.card*20>(h2{Amazing Card $}+img[src="https://picsum.photos/400/200?random=$"]+p*5{Lorem ipsum dolor sit amet consectetur adipisicing elit.}+ul>li*10{Feature $}+button{Discover})
That alone, outputs:
- more than 1500 HTML nodes
- 265 475 characters
That means I could turn 259 chars into 265 475 characters ????
Yep. Pretty cool isn’t it?
Source: Emmet abbreviation syntax
Does Emmet exist for all programming languages ?
Not as a standard. Emmet was made for HTML and CSS.
But, that doesn’t mean that:
LLMs don’t understand those patterns
And if you add a skill on top of that ! It learns it in one shot !
Can you add it to Claude Code ?
Of course ! MCP ! MCP Guys ! You can actually add all the tools you want.
Disclaimer: MY plugin that I sell 9.99$/month doesn’t actually use “MCP” in this strict term. What I actually did is a Claude Code Plugin, because adding more tools to Claude kills what we want : Reduce usage. So my plugin intercepts
PreToolUsehook so Claude does it in its native tools. But it’s less complicated to add it as a tool, if you want to do it quick and dirty.
Source: Claude Code hooks
What does it look like?
Simple example for Python:
fn[demo]>p("WOW $")*10000
def demo():
print("WOW 1")
print("WOW 2")
print("WOW 3")
...
print("WOW 9999")
print("WOW 10000")
// Output 218 905 characters
Does that reduce quality?
No, because it’s even more clever for LLMs that don’t like to output everything (the famous //…)
Any other techniques ?
I got another for you guys!
Did you know that Claude in its edit tool needs to write:
- file_path,
- old_string,
- new_string,
- replace_all,
I want your focus on old_string.
That string alone is your second enemy. How Claude needs to use it:
- Read a file
- If it needs to replace a part, it needs to output it the exact same way.
I want you to understand:
If Claude needs to change 100 lines of a script into 3:
Claude needs to output : 103 lines!
That alone has a big cost!
How to fix ?
In fact, that part is not new, it’s actually been challenged among others for years.
But possible solutions:
old_string : line_start:line_end -> do a script that captures it and replaces that with the desired part programmatically.
OR
REGEX : Claude writes regex patterns in that old_string, same thing as previous technique, search replace to output that in the tool.
Does that reduce quality ?
This one could. Yes. In fact, if today this is the most adopted solution, it’s because it’s more reliable, LLMs can’t miss what it wanted to replace, because it writes what it wants to replace.
But with some safeguards + a tool undo_last_action: That’s working.
Those 2 previous techniques alone, give 90% of what my own plugin claims to save. Think about it carefully.
Claude’s reasoning effort
Of course that too helps:
Claude supports those reasoning effort levels:
Low
Medium
High
xHigh
Max
Source: Claude reasoning effort
The fact alone is that:
The way Claude “reasons” / “thinks”, is by throwing away your money. It is in fact a bit true, but it helps quality very much, it’s helping to understand what it has in its context.
But, like for models, depends on what you need. And Anthropic themselves publish benchmarks of their models according to the reasoning effort, and we see a big change on it.
Source: Introducing Claude Opus 5
That alone, I couldn’t give you a premade answer, but for Opus / Fable : Medium – High is the best range for “Pareto” (80% of results for those 20% costs).
Cache miss
It is important that you understand that, I see this problem very often:
When you talk to Claude, the whole conversation has a “cache” on Anthropic server, so they don’t need to decode all the information again.
But with all the active sessions at the same time, Anthropic can’t store that indefinitely.
For Anthropic’s subscriptions the cache expiration is : 1 hour.
Source: How Claude Code uses prompt caching
That means : If you left a conversation for more than 1 hour. Talking to Claude again will cost you every piece of character you’ve written in the past.
So: Hit that /clear on Claude Code (or + on Claude Desktop)
Long session
That has a big effect on your usage limits too.
Let’s bring back that sweet table from API pricing:
| Model | Base Input Tokens | 5m Cache Writes | 1h Cache Writes | Cache Hits & Refreshes | Output Tokens |
|---|---|---|---|---|---|
| Claude Fable 5 | $10 / MTok | $12.50 / MTok | $20 / MTok | $1 / MTok | $50 / MTok |
| Claude Mythos 5 (limited availability) | $10 / MTok | $12.50 / MTok | $20 / MTok | $1 / MTok | $50 / MTok |
| Claude Opus 5 | $5 / MTok | $6.25 / MTok | $10 / MTok | $0.50 / MTok | $25 / MTok |
| Claude Opus 4.8 | $5 / MTok | $6.25 / MTok | $10 / MTok | $0.50 / MTok | $25 / MTok |
| Claude Sonnet 5\ | |||||
| through August 31, 2026 | $2 / MTok | $2.50 / MTok | $4 / MTok | $0.20 / MTok | $10 / MTok |
| Claude Sonnet 5\ | |||||
| starting September 1, 2026 | $3 / MTok | $3.75 / MTok | $6 / MTok | $0.30 / MTok | $15 / MTok |
| Claude Sonnet 4.6 | $3 / MTok | $3.75 / MTok | $6 / MTok | $0.30 / MTok | $15 / MTok |
Source: Claude API pricing
Do you see “Cache writes, cache reads”.
What does that mean ?
Told you in previous section that every character from the past will cost you money again.
In fact, even with cache, that’s the same case:
But price is lower.
Let’s do an example:
User > Hi Claude, please make for me a whole WordPress website! Price: (10 words at 5$ per 1M, it’s input for Claude)
Claude > Let me read current project Price: (5 words at 25$ per 1M, it’s output for Claude)
Claude > Read_file(README.md) Price: (2 words at 25$ Claude output function tool call) + 5 words at 0.50$ (Last turn from Claude becomes cache input) + 10 words at 0.50 (from user first sentence).
So, it’s stacking !
And if LLMs counted in words, that would be nice! But they count in “tokens” (roughly approx. 4 characters).
And now Claude models do have a 1M context window. That alone, could mean with “cache”, it could have inputted the same inputs/outputs thousands of times, at 1M, you will have multi-million cache hits from turn to turn.
Sources: Claude models overview, Claude API pricing
So: Slow down again a bit and hit that /clear again!
Claude’s input
As I already stated from the previous example. It’s the price from Claude reading things. The more it reads, the more Claude costs money.
Your location
No, your location alone will not cost you more. But it’s more subtle:
What is your primary language ? Not everyone speaks French.
Me, I’m from Quebec, Canada (French).
Do you speak Spanish ? Or Mandarin ?
Because, this has an impact:
Claude’s way to understand words relies on a tokenizer, where it understands characters. And not every character has the same weight in UTF. That means :
Your main language COULD cost you more money.
Source: Claude API pricing and token usage
Solution: IF you’re that hard into getting more Claude: talk to Claude in English.
Conclusion
I tried to explain those complicated things with “over-simplification”. In reality when I’m talking about Claude’s processes, it’s more complicated, and all that information is publicly stated by Anthropic, but that information alone could help you get through your whole week of usage limits.
If you want to get everything I stated in this article (And much more!), you can use my plugin CULP. It is 9.99$/month. But with what it gets you back from saving, I think it has a real value for you.
(AND of course, if you don’t want to, just copy that article, paste it into Claude Code, and Claude will figure out a way to do it)
Can’t wait to get feedback from you! I’m reading everyone, if you’ve got any other questions feel free to ask!