Videx (@_Videx) header
Videx (@_Videx) profile picture
Joined May 2022
Photo by Videx (@_Videx)Photo by Videx (@_Videx)Photo by Videx (@_Videx)Photo by Videx (@_Videx)Photo by Videx (@_Videx)Photo by Videx (@_Videx)Photo by Videx (@_Videx)
GPT gives you 40% more tokens Or how internet analysts bullshit people On October 5, a fucking awesome study came out, proving that Claude gives you 5 times the usage allowance OpenAI does, across every subscription tier. And okay, I’d already given Opus 5.5 the benefit of the doubt on graphics, 3D, design, and video editing. Half of that turned out to be complete bullshit in my own tests. So I was still willing to accept that Claude gives you way more usage than GPT. Sure, fair enough. But not fucking 5 times more! Especially since my experience using 6.1 SOL and Opus 5.5 on $200 and $100 subscriptions doesn’t show anything like that difference in token allowance. So I went digging. Let’s start with the fact that the entire study extrapolates usage into API costs. The headline tells us you get 5 times more usage. They also used OpenAI’s announced pricing, not the pricing actually in effect. What I mean is that their study factors in a 50% token allowance cut that won’t take effect until October 29. But that multiplier is in dollars. In tokens, you get twice as much. To be precise: 5.63x in dollars and 2.80x in tokens. The counting method matters, though, because those multipliers and values describe one specific metric: token consumption. They gave the models random tasks without evaluating the results, along with huge chunks of War and Peace to measure token consumption, caching, and all that on the same basis. Then they compared how many tokens went in and out, and what percentage of each subscription’s allowance got used up. After that, they applied API pricing to work out how many $$$ worth of tokens you get inside the subscription. For the record, they mainly compared 6.1 SOL against Opus 5.5, and that’s a very important detail. Watch closely. I have access to Opus 5.5 on a $100 subscription, and I burned through around 75% of its allowance in about 5 days. Roughly the same happened with GPT using 6.1 SOL + Astra, with a lot of Astra. And I wasn’t the only one using Claude. The account owner was constantly vibe coding stuff on it. GPT is exclusively mine, and I use it 24/7. Meanwhile, I’m on high-xhigh, and he’s blasting ULTRA. In theory, that suggests Opus gives you a shitload of usage, since you can burn through it on Ultra and still last more than 3 hours. According to their formula, Opus gives you 37% more tokens, since GPT’s plan doesn’t change until October 29. The $100 subscription gives you 37% more than the $200 subscription. But in practice, didn’t both subscriptions run out at the same time? Exactly. Which means that 37% token difference, who could’ve guessed, got eaten up by Claude’s massive consumption. And if that extra consumption doesn’t make it more effective at completing tasks, their bigger allowance is just hot air. On top of that, I don’t have the 5-hour limits I hit twice on Claude. I also have 3 resets sitting there, they added another one yesterday, plus 48k credits left out of 52k ($2,500), which they gave out when they introduced the $500 plan and cut the current Pro allowance. But that’s not all! Earlier, I said they didn’t evaluate the cost of completing tasks. It didn’t matter whether the model actually succeeded. They just burned tokens and checked which subscription had more. And here’s where the math gets interesting. I spent 2 days working nonstop with 6.1 SOL on 30% of my weekly allowance, vibe coding a full rework of ShojiWM and building my own DE for it from scratch in QuickShell. I wrote different animation and shader variants. Around 1.4 billion tokens over two days using SemiAnalysis’s methodology, while Opus 5.5 should’ve burned through 950 million using that same methodology. So, again. We can see the difference in tokens. We can see the difference in price. But on my actual tasks, the subscriptions run out at roughly the same time. Meanwhile, you can see on my channel that GPT did better at 3D, that I use it for website design just like Claude, and that its video editing is around the same level too. Which means GPT gets tasks done at Opus’s level, like for like on subscription allowance, for the same money, but you also get resets. So you literally get even more tokens and can finish even more projects. Check my GitHub. What the fuck is this? Here’s what the fuck it is: marketing and consumer-brain rot won. Claude’s overall vibe, its tendency to make pretty MVPs, and people’s desire to feel smarter than everyone else. They’ve created a class of people who’ll call it holy dew no matter how much you fucking piss in their eyes. And they’ll smugly present these “proofs” as evidence that their subscription is the better deal.
11
4
4
52
2,471
How Google feels after releasing weak Gemini 4 Argon
2
51