🚨 I tested Claude Pro ($20) AGAIN!!! I used Opus 5 at MAX settings This is CRAZY!!!! 30 minutes? that must be it, right? NO!!! here is what I actually got: - 1 Hr and 24 minutes of usage!!! - 13% Weekly limit consumed - $37 API usage -... 40 Million tokens!!! Meaning!!!! You'll get ~$1130 worth of API usage for $20!!! (on max) And a TOTAL of 43 HOURS of WORK!!! (in a month) Its MUCH BETTER than Codex!! I'd recommend using Med (cause it gives more tokens) Codex!! Good bye, I don't think I'll ever come back!show more

Those numbers look really good for $20. One of Codex’s biggest advantages has been the occasional resets, but with the capacity issues they’re facing right now, I doubt we’ll see them as often as we used to.

That's why I switched to Claude Code and tested it out for you

I also noticed that I burn the Codex limits in about 10-15mins with Sol High on the 20usd a month plan. CC with Opus 5 on med/high offers significantly more.

That is why I tested out Claude Code for all of you guys. I was so fed up with Codex.

Your evaluation highlights the problem really well. I've also compared Claude and Codex, and I noticed the exact same issue.

Is the issue? I don't think there is an issue

put your two runs next to each other: 44 min was 7% of the week, 84 min was 13%. that's 6.3 and 6.5 minutes per 1% so the weekly meter is counting time, not effort and not tokens. medium gave you 727k tokens a minute, max gave 476k, the week cost the same either way

Yeah and that's why I would advise everybody to use Opus 5 at medium settings because that is how you get the maximum amount of work done.

right, medium buys you about 1.5x the tokens per minute of the week the open bit is whether those extra tokens are worth as much. same task on both, count the retries, that's what decides it

this is what a week is supposed to look like

I think so Maybe next week I'll become a fan of Codex 😅

yeah codex usage is worse than claude

It was a lot better before September but then it got worse.

Yeah, back then I used to code for upto to 2hrs but now it lasts 20mins

Lasts only 20 minutes for Codex, not Claude Code, to be honest

True

Depends heavily on what you're doing tho. Long context sessions burn the weekly limit way faster than a bunch of short ones.

But I did use the longer context Opus 5

Fair enough lol. I meant the agents running all day kind of usage, normal sessions never got me either.

The more people use it, the less quota each will get. I think they (both Claude n ChatGpt) devide their total capacity by total number of active users. That's why it's always changing.

Ohhh... That might be the case ngl

👀 nice

Thank you 😊 🙏

40 million tokens for $37. My coffee budget is judging the math.

Stop drinking coffee and start buying Claude.

Coffee stays. It's the only employee that never files a bug report.

A few things I’ve also observed that you might have noticed too, but missed in the post. Speed has increased a lot as well. Earlier, it felt so slow, but now it feels much faster. For the first time, I’m saying this, it takes me around 1 or 2 hours to hit the 5-hour session limit. Earlier, I used to hit it in just 30–40 minutes. The Claude model also feels production-trusted now. It doesn’t make as many foolish mistakes. If I give it a command and something is wrong, it usually corrects me instead of blindly following it. Some issues were still there around 20% of the time, but yeah, overall it feels much better.

I agree with you bro. I kind of hated Claude just two weeks ago but I think it's now fair to say that I am biased towards Claude over Codex.

Bro, if they introduce the Fable model in the $20 plan, it would probably become the best subscription for intelligence access we could have right now. Just hoping that happens.

Won't do that because they want scarcity

Today it's much much better. I use Astra xhigh on 5 sessions since yesterday, only consume 22% weekly limit. Normally it wouldn't last a day. I'm really happy that people leaving codex so they have more compute for me

Bro, I do not know how this happened but if it did happen let me know what plan you're using and I will see if I can get that plan and test it out, codex, again.

It's been 24 hours with Astra xhigh 5 sessions, now only consume 27% weekly limit. I still have 3 bank resets too btw. Rate limit much much better. Just don't use subagent

Come on bro... I have never even tasted a $100 codex plan so I would never know 😶 nor I have $100 to test anything, $200 is so far haha If you say that the $200 plan has so many limits, then I guess it must have. When I'll have money, I'll try it out too!

If your budget is tight, I recommend stick with Codex plus or Claude Code pro plan. Then pay for DeepSeek v4.1 flash API, use it in OpenCode. 3$ cover 500M tokens (with 300 token/s speed). It's my life saver this week before normal reset (I just don't use bank reset)

Claude subscriptions are worth for their subscriptions currently

Yeah... Codex is nerfed ngl

The useful number is cost per finished task, not tokens packed into a $20 seat.

And I think it's amazing ngl

$37 of free usage" is such a charming way to say "I found a bug and I'm milking it before it gets patched." Hope you enjoy your 40 million tokens, king.

I enjoyed it to the fullest

Thats wassup

Nice bro, please try with High 😁

Okay sir! I believe it's going to be relatively similar. For example if Max ran for 1 hour and 24 minutes, Hi can run for 1 hour and 40 minutes or something. To be honest the quality difference isn't great and is almost similar. But if you insist then yes, I will do it again. Not now, maybe 1 or 2 days later, because X algorithm has started to deboost posts like these. (My account might get shadow banned or something haha)

Can you try this with $100 max too?

I must have done that way earlier but I don't have $100

Thank you for this. I went from Cursor to Codex, and while I like the Codex experience, the limits just make Cursor way better. But with your tests, I think I might consider Claude Code instead and see where it takes me. 😁

Sure. Because I got monetized today, in the future I will create posts where you can find out the exact limits and comparison between the three.

Anthropic is back

It is SO BACK! I LOVE IT!!!

The surprising part is the gap between wall-clock time, weekly percentage, and token/API burn. Tracking all three together gives a much clearer picture than any single counter—and makes Opus vs Codex comparisons fairer.

Yeah

Claude Pro is better than GPT Plus for normal usage but GPT plans get multiple usage resets, making GPT Plus more worth it

lately we haven't got any reset at all for 8 days straight so I don't think it's going to be worth it because they are lacking compute.

toilet paper shortage type riots coming.

@grok I am not good in English... What does this guy means?

@Jasper_Black He means your post shows Claude Pro is such a great deal that many people will rush to sign up, just like the panic-buying and fights over toilet paper during the 2020 COVID shortages. So similar chaotic demand is coming.

If you use Astra Low on the Plus plan ($20), the 5-hour limit lasts about 15 minutes, but it deducts around 10% from your weekly quota rather than 20%.

No bro, I have tested out the Codex plan for the last week and to be honest it consumes 15-20% of your weekly limit each time your 5-hour limit is drained And on September 9th when I tested out Codex on Astra, I used GPT-6 Astra on maximum settings and it worked for 40 minutes but then the limits got nerfed so I'm just leaving Codex

I have Codex 5x and wish i could get 20x. I’ll probably try Claude $20 and see how it goes. Codex lately really sucks, it’s not just people complaining anymore

Currently I would advise you to not get a $200/month subscription for Codex because the usage limits are so generous. Maybe try some other subscriptions, for example, Super Grok Heavy or anything similar. I think they're running an offer for Super Grok Heavy for $160 for 2 months or something. (Keep in mind, you get $1500 Usage on API rates for Grok 4.6 on a $30 plan)

ok this is a bit intresting, but probs will wait for their new model, as for codex, it doesnt let me go 20x otherwise it wouldve made sense to me but 5x doesnt last a day, now im only using luna🥲and its dump lately

感觉是不是弄错了，你可以用ccusage统计一下试试呢？我一个多月前用Claude的时候，5小时窗口基本上在8美元左右

I agree with you that a month ago Claude usage limits were nerfed but right now Codex is doing exactly what you are telling me was happening with Claude. Right now Claude is giving more usage and to be honest I'm very happy with that.

😁

In 17 minutes - 5h usage and 20% of weekly gone in CODEX (Astra Medium).. I think OpenAI is joking with it's customers.. Rally think to switch

I think so. They have a shortage of compute. That's why the limits are so low.

But we are paying for a service right? So why AT LEAST we get what we paying for?! Good Bye Codex.. moving to Claude Code

Yeah that is the whole point of me leaving Codex right now hahaha

Yea I've noticed that I get quite a lot of Opus 5 bang for my $20/month buck these days 👍

You are welcome

Do you Get more compute from Claude 20x?

No you absolutely will get a lot less compute on a $200 plan on Claude Just get 10 subscriptions of $20 and you will get at least $12k of compute

That is less than Med but worth it ngl

Should def use Opus at Med. Good results and more usage

You think that’s how it translates ? Lmao

Enlighten me then how it does?

In case anybody wants to use Both Codex and Claude on one project. I recently read about something that was supposed to combine them somehow but I cant remember. Anybody out ther with a Clou?

I don't know, man. It's probably better to use Claude and Codex separately because of the cache hit problem.

yes claude usage is better than codex

For now Claude usage is better than Codex but who knows? The future might change. 😅

Working daily with Opus, i never hit the session limit on Max 5, switching to Fable only for complex tasks. For me i hit session limits when using code-review or subagents.

For reviews I think you should be using Jev... Otherwise your structure is amazing

Easy to understand when 67% of the user base left

What bloody os is that btw