🚨 codex, claude and grok users!! I have FINALLY tested this for YOU!!! the $20 subscriptions! Codex VS Claude VS Grok (Watch video for results) Claude Pro: - Opus 5.5 (max) - 10% Weekly limit = 115 Million tokens - 2 hr work - 5/10 Quality Codex Plus: -... GPT-6 Sol (max) - 9% Weekly limit = 29 Million tokens - 69 min work - 7/10 Quality SuperGrok: - Grok 4.7 (xhigh) - 23% Weekly limit = 19 Million tokens - 36 min work - 6/10 Quality All the models were used at MAX!! (grok 4.7 at xhigh) The surprising thing is that with my calculations is that I am getting $4,000 worth of API usage with Claude pro 🤯 AND ALSO!!! DO NOT BUY SUPERGROK!!! 23% of weekly limit for only 19 Million tokens ONLY!! Codex is nerfed too... you are still getting only $330 of API usage monthly My opinion: For $20, GO WITH CLAUDE!! You can always use those 4x extra tokens to improve again and again NOTE: Please forgive the video quality, this is my first video ever... and I don't know editing. Just sharing info about all these AI subs. Hope it helps 🙂show more

Here are the usage bars AFTER the test

Thanks for sharing your experience. I'm a ChatGPT plus subscriber but I think I'm about to buy a Claude Pro subscription, I have Opus 5.5 FOMO.

Hahaha... Definitly try Opus 5.5 ... its a good model (especially better than GPT 5.6 Sol) Let me know your experience after that.

Wow....Claude is wayy better than I thought. Opus 5.5 is pretty good at high and token efficient. Should get a lot of mileage that way.

Its not token efficient... Opus 5.5 is token hungry... but the limits are generous to cover up for that... I think for 1 Billion tokens per month for Opus 5.5 ... you'll be a able to do A LOT of work ngl

@RajanthaR I’m already at 2 billion in the past 5 days with deep seek Am I missing something? You can get anything done with 1 billion tokens?

Deepseek V4.1 Flash is really fast and it also uses so many tokens. In token count, my top model is also DSV4.1 followed by Luna followed by Grok. DSV4.1 is on par with Fable 5 and reliable. Doesn't come close to Fable 5.1 and more reliable than Opus 5. Opus 5.5 beats Fable 5.1 and it beats Astra in terms of consistency and reliability. Astra sometimes so really dumb things. Astra can also run around in circles. Only models that beat DeepSeek V4.1 Flash that is reliable is Fable 5.1 and Opus 5.5 and with the cost difference, they are definitely not worth it at API cost. Claude $20 Gives you $4000 ish of API spend. It's the most cost effective plan when you compare with Claude max plans. You can get 200+ TPS from Deepseek, 300+ TPS with CoreWeave and 400+ from Makora at 2x price. Can't go wrong with that.

@shownotover It’s wild how reliable DSV4.1 is it’s literally god tier imo best model since 4.6 when it came out. Now it’s the best imo aside from fable and astra and supposedly 5.5 but idk if I even need 5.5 tbh

Yeah. It's wild how Deepseek Flash leaped from preview to V0731 to V4.1 in a span of few months. Unless you're doing something new, I don't think you need to switch. The 20x increase in the max plan is only for the 5hr window btw. In a lawsuit it says that 20x weekly usage is 1.5x of the 10x plan. And 10x plan weekly is like 6.5x of $20 plan. Who knows how next Deepseek will be. Might catch up to Opus 5.5. You can just try the $20 plan and see how far the model can be pushed. Plus you get $100 of Cloud usage. That will probably burn at API rates so might be better spent with what the next Sonnet model will be.

So what are you saying—that Anthropic is somehow giving Max 20x users the equivalent of $80,000 in API usage, or that Claude Max subscribers are all idiots paying more money for usage that doesn’t even scale with the price?

No no 🙂‍↔️😭 I got this much usage somehow on a $20 plan 20x max plan is only 4-5x more usage than Pro plan anyways... Mostly you get $1100-1400 worth of usage on the $20 plan monthly Idk how I was able to juice out 115 million tokens out of Opus 5.5

Opus 5.5 is great—I’ve used it in Cursor—but I’m never subscribing to Claude again. I’ll never forget how a single conversation in Chat, not Code, burned through 36% of my five-hour usage limit when I was using Opus 4.7. I couldn’t even use the chat comfortably. Even though GPT-6 Sol was a disappointing release, I still really like its Chat mode. Even back when I was on the Plus plan, I never once ran into a usage limit there.

😂🤣 It was the same time I cancelled my Claude sub Opus 4.8 was launched and it ate my 5 hr limit in 20 mins... I canceled and moved to codex But now, idk how and why... its somehow fixed. But again, Good luck with Cursor... I have never used it... maybe in the future

Per your results Sol produced better quality, used less weekly allowance, took less time, but you still recommend Claude Code because of the token numbers? Looks like your recommendation focuses on the only number that isn't meaningful here

Why do I recommend Claude? Because: - Codex gives you only $300 API usage VS $1100 monthly - 115 Million tokens vs 29 Million is a difference The thing is, If I don't one-shot project and use with care, I can use Claude a LOT more than Codex But again, if you like Codex... please choose that... you will always curse me if anything goes wrong on Claude 🥹

I guess it depends where you consider the value lives, in raw API prices vs what it actually produced. It is hard (probably impossible) to tell whether the figures you observed translate across many use cases though. Thanks for your tests and reports

I got Claude Pro just to try it out, and I can comfortably run 4 mid-sized tasks within a single 5-hour window. My dad is on GPT Plus, while I’m on the 20x plan. And I can honestly say, even with him using Luna, he burns through his limits so easily that I genuinely feel bad for him.

🙂 Bro... you are on a 20x plan... what do you expect? That's like a $360 vs $8000 battle If you were both on $20 plans... I could have said that I feel bad for him too (even though I am still feeling bad for him) Get your dad a Claude Pro ... will work enough for him (And one opencode Go sub for Deepseek v4.1)

Nah, I wasn’t comparing my 20x plan to his Plus 😭 I meant that I can comfortably run around 4 mid tasks on Claude Pro’s $20 plan, while my dad burns through his GPT Plus limits surprisingly fast even though he mostly uses Luna. The 20x part was just context about what I personally use. And yeah, you’re right. I’m definitely considering getting him Claude Pro now 🙂

Oh... my bad 🥲 Good luck with coding bro... Also use these plugins... improve quality, token efficiency and saves usage limits: - Graphify - Ponytail - Obsidian - Codegraph Good luck!

@furkan_ceyran Hey, can you explain what the plugins do other than ponytail? Thanks.

@furkan_ceyran They save context, take notes, build memory and ofc save tokens instead of rereading everything

I told you so. Then you went and tested it, Great work, wonderful info for everyone!

Your welcome bro... anything else you want... I can cook it up too (for now) haha

You must not have believed the numbers I showed you so you went and proved it. I like that in people, keep it up!

Your conclusion to go with Claude is stupid. Your experiment proves that Codex is the superior choice: its faster, its cheaper, it uses less of your limit, its higher quality. Even if you were paying API usage, Codex is better. You are a Claude cuck.

Bro... its your choice... I gave my opinion... Choice is yours... sure go with Codex. Good luck

Great job. Opus 5.5 is doing great for me. I'm using chatgpt to help organize it though but opus is the coder

Cool bro... Good luck coding

No entiendo, me dices que gpt me entrega mejor calidad y gasta menos mi suscripción, pero me recomiendas Claude? Es absurdo, tu test solo demuestra lo ineficiente que es Claude. Has analizado 3 variables (calidad, tiempo y tokens) y usaste la peor para recomendar.

1. Claude usage is a LOT, you can revise your work 5 times rather than one-shotting on codex 2. Opus 5.5 at max overthinks and messed up at the end... Med Opus 5.5 >>> GPT-6 Sol 3. Because of stupid overthinking, it took so much time testing... But again, its your choice... You can still choose codex.

are you sure you used Opus 5.5? or why does it says "automode on" for claude

I definitly used Opus 5.5 because after all the tests... I asked Mimo v2.6 to make a report which can be seen at 11th sec of the video... it didn't give me "Oh and other models were used too" ... it simply said Opus 5.5

Thanks man. Let's hope Claude does not nerf it's limits.

I really hope too... But it will happen in future and I'll update all you guys

What jumps out: Claude used 115M tokens for a 5/10, Codex 29M for a 7/10. So the "$4,000 of API value" measures how many tokens Opus burns, not how much work gets done. How much of those 115M were cache reads? That changes the math a lot.

99.8% was cache read. Total cost was $90 Tbh, still feels the best value

good analysis bro. Sol did the job pretty well. right?

I mean, it all depends on how you look at it... For me, even though it messed up... usage is too much for me to overlook... Codex did a good job 7/10 but the limits are nerfed so... your choice

What the result actually shows is that COT-6 Sol and Grok xhigh are better than Opus 5.5 and that you don't need the max efford to get a similar resulte than 5.5 max. They did a better job quality and faster.

What about the 5x tokens you got on Claude? You can do 5 times over if you DO NOT one-shot things... I am stupid, but you aren't and you'll get a lot more usage ngl

It is simple. Open AI gives less tokens because their models use less and Anthropicss more cuz their model uses more. Usage apparently is proportional to token consumsion by every model.

@thsottiaux Where are those “great limits”? Or do you still see Plus users as casual users?

@thsottiaux He sees all the plus users as "targets" and somehow make them get Pro 5x plan 🫠🫠

@thsottiaux I already paid for the 5x Pro plan. It actually lasted a decent amount of time, but still not even 3 days. Brutal, especially for me as a Brazilian. It’s extremely expensive here.

@thsottiaux Codex never worked for me in Sep (was working good till Aug) same with SuperGrok (I was able to use it for 3 days straight 24/7 in hermes agent but now can't even do 10 hrs) And... Claude... that's surprising to me

Since users in China cannot access Claude, which subscription do you recommend?

That'll need workaround... Buy Codex (I know its nerfed but best you got) and then connect with Hermes agent and use Deepseek API as "implementor" So Codex thinks and Deepseek does.

Thanks for the reply. If possible, I’d love for you to review Grok 4.7 in Cursor; I believe it offers a higher usage limit than SuperGrok.

Sure, I will but currently out of budget... I will test it next month... have to cancel SuperGrok and Codex

However, cladue is limited, so codex is still relatively friendly.

Go with Deepseek API $2 minimum... That'll give you at least 300 Million tokens... No need to pay for Codex $20 for Luna

Honestly speaking codex has always felt nerfed to me since I started using it idk after astra tho, for me claude code is the perfect use getting the max out of it rather than just getting something bland

I agree with you... Before Opus 4.8 which was 1.6 months ago... Codex was good... now its nerfed af

Glad I didnt follow the hype and didnt get chatgpt plus when Astra came out stuck to Opus 5 knew they are gonna release something fast to take over🤣

Claude has been on top for 2 years now... Total dominance

Exactly feel a bit behind as I actually started using claude code since July before that I was on codex for a bit and antigravity

wait that's not what I experieced at that time when grok build just launch , Elon cut that hard?

Yeah!!! Don't buy grok for now

How much did claude pay you?

I wish they had paid me ☹️

5/10 quality on opus 5.5?

yeah... check the video from the end

@thsottiaux is this right?

@thsottiaux He won't answer haha

Can you test again 10 bucks Opencode go?

Opencode Go is very transparent ngl... we don't need that do we?

Great bro 😅 even though limits are nerfed i think codex is still good coz it uses less tokens and has lower pricing so api pricing wise codex still wins

Maybe... I think Claude is giving a LOT of tokens here... But whatever you choose, it's best for you

首富的api 居然是最贵最垃圾的

What do you mean by that ? 😭

true, the only thing is Claude’s quality is 10/10

well, it wasn't able to 1-shot the project with the same prompt so... its kinda bad

Thanks for sharing,

No probs... Let me know if I can do it some other way if it helps

What about using only mobile app?

No... its on Linux I have recorded the screen

Качество клода 5/10, аххаха, вот клоун, а.

I am saying that the quality isn't as good as others but there are so many tokens and so much work that you can repeat the process one or two times and you will double the quality, tbh It's not like you can one-shot an entire project. 🥲

Does claude have 5 hour limits?

Thank you for this

No problem

Also, you should mention that the groc subscription also has grok bot so there’s a lot more usage than just grok build…

Grok bot washed the limits in 4 prompts... those seprate limits are a joke (SuperGrok $30 plan)

Oh on superfrok plus it’s not bad

Yeah but if the budget is $20-30 you can't do anything can you?

Nope lol also, I don’t know who is expecting in 2026 to pay $20 a month and expect to get anything done. If it’s truly for work and making you money, then $100 a month is nothing.

I agree with you bro. SuperGrok Plus is not something to buy right now because the limits are so much nerfed. If I had $100 then I would definitely get Claude Max.

Lol I felt this too, That's why I'm wondering why 5.6 SOL is taking more usage than I used to work with Claude. So rn Claude ain't expensive it's GPT

Exactly!

20刀的claude周订阅有10亿？

YES! 1 Billion tokens is what I got I have also conducted two recent tests in which I tested Claude Code and in each test I got at least 50 million tokens for 10-13% weekly usage... You can expect around 500 million to 900 million tokens per week

谢谢，非常有用的数据

Thanks, i will buy Claude Today

Good luck bro! Lemme know if you need anything... I have plenty of info

I'll do that (: I'll be working a lot on my Unity game with Opus 5.