# ChatGPT Plus vs Claude Pro: same $20, but one is worth 5x more.. **Channel:** KernelBrief **Source:** https://www.youtube.com/watch?v=CqFKw2dtGi0 **Transcript page:** https://www.withtranscript.ai/video/CqFKw2dtGi0 ## Chapters - 0:00 — Why $20 Plans Need Measuring - 1:13 — Comparing Value Across $20 Plans - 4:19 — Higher Tiers And OpenAI Cuts - 7:00 — Margins, Usage, And Profitability - 9:41 — Which Plan Is Worth Choosing ## Transcript **[0:00]:** ChatGPT plus and Claude Pro cost exactly the same $20 a month. This week, Simianlysis measured how much AI each one really gives you before it cuts you off. Priced at API rates, one is worth more than five times the other. So ChatGPT plus versus Claude Pro, which $20 buys more? We're going to price all seven of their plans from $20 up, one by one, and find out who's losing money on yours. **[0:28]:** Here's the problem. You neither company tells you how much you get. ChatGPT's Plus Card says expanded five times and never says how much. Claude's card says more usage with an asterisk. What you really buy is a meter. **[0:43]:** Every message fills it a little, it resets every five hours, and a weekly cap sits on top. So Semianalysis fed each plan pages of War and Peace over and over and watched how far the meter moved. Count the tokens it takes to move the meter one notch. Scale that up to a whole month. Then price every token at the API's list rate. **[1:04]:** That's the plan's value in dollars. And they measured it the way people use coding agents like Claude Code and Codex. Keep that in mind. It matters later. **[1:13]:** First, ChatGPT plus on GPT 6.1, Sol OpenAI's mid sized model, the one built for coding agents. **[1:21]:** Take a guess. If you used every last bit of plus for a month, what would those tokens cost on the API? $20, 200 or $2,211. That's about a billion tokens and ten and a half times what you pay on GPT6, Astra OpenAI's flagship. The same plan is worth $162, eight times what you pay. **[1:45]:** Not a bad deal. Now the same $20 at Anthropic, Claude Pro on on Opus 5.5, the model Anthropic calls its daily driver $1,178. That's 59 times what you pay and 5 1/2 times ChatGPT+. You might be thinking Opus costs twice as much per token, so of course its dollars add up faster. So count tokens instead. **[2:10]:** Claude Pro 2.9 billion a month plus 1 billion. Still almost three times as many. Even Sonnet 5.5, Anthropic's cheaper model is worth $1,241 on that plan. I read those numbers twice. So how does the same $20 buy three times the tokens? Here's how that happens. Remember, they measured coding agents. An agent works in steps, and every step sends the whole conversation back to the model, the task, the files. Every earlier step. Say the conversation has grown to a hundred thousand tokens. **[2:46]:** The next step sends all hundred thousand again, plus a few hundred new ones. So almost everything the model reads, it has read before. Those are cached tokens, and in Simianalysis mix, they're 96.6% of everything on the API. A cash token is cheap for Opus, $0.20 per million against $4 for a fresh one. But your plan doesn't charge dollars. **[3:09]:** It moves a meter, and each lab decides how far each kind of token moves it. On some plans, reading from the cache barely moves it at all. Those exchange rates are the whole game. The same $200 Claude plan is worth $2,500 on one model and $12,500 on another. And if you mostly chat instead of running agents, Opus on that plan drops from $11,700 to $9,200. **[3:35]:** But that's the everyday models at the top. It flips. Claude Pro doesn't include Fable 5.1, Anthropic's flagship at all. ChatGPT plus does include GPT 6 Astra. On the $200 plans, Astra is worth $2,897 a month, Fable 2,485. **[3:56]:** At $100, it's almost a tie. $1,322 of Astra, 1273 of Fable. And Fable only gets half your weekly limit. Use that up and you're done with Fable until the week resets. So if you want the biggest model, OpenAI's plans are the better deal. **[4:14]:** For the everyday one, Claude wins by a mile. That's two of seven. **[4:19]:** Plans three and four cost $100 a month. ChatGPT Pro 100 and the $100 Claude Max Pro 100 gets you $1,055 of SOL. Claude Max gets you $5,725 of OPUS. **[4:34]:** More than five times again. And per dollar, nothing changes. Each lab gives you almost exactly the rate it gave you at 20. Which raises a question. Why would anyone pay OpenAI more? **[4:45]:** Until last week, there was an answer. ChatGPT Pro200 used to be OpenAI's best deal per dollar, about twice what Pro100 gave you. September 29th, OpenAI cut it. Astra's value went from $5,734 a month to 2,897. Sol's fell from 5,904 to 2084. **[5:09]:** Half the tokens for the same dollar. 200 new subscribers got the smaller plan that day. If you already had one. You keep the old limits until October 29th. Here's how OpenAI's help page explains a lower allowance to reflect our increasingly efficient models. **[5:25]:** The idea is that newer models get more done per token. Simianalysis says nobody has reliable data on that yet the cut they measured. Meanwhile, the $200 Claude Max is worth $11,726 of Opus. One Plan 7 is new Pro 500 at $500 a month. So what do the extra $300 buy against the old Pro 200? **[5:49]:** It costs 150% more and gives you 21% more Astra. What the extra money really buys is speed Ultra fast up to 300 tokens a second in Codex, and Ultra Fast drains your allowance at eight times the normal rate. That's all seven plans priced. So which plan gives you the most per dollar? Put all seven side by side. **[6:12]:** Every ChatGPT plan now lands at about 10 times its price on SOL. Every Claude plan about 58 times on Opus before the cut, OpenAI paid you to climb on Astra Pro 200 gave about 3 1/2 times what plus did per dollar. Now Pro 100, 200 and 500 all give the same so the scoreboard at 2100 and $200. Claude wins on the everyday model. On the flagship OpenAI wins at every price one point in OpenAI's favor. **[6:46]:** Its Pro plans have no 5 hour window, so it's easier to use all of what you pay for. Semianalysis doesn't think that closes a gap this big. And a bigger OpenAI plan no longer buys a better rate, just more of the same. **[7:00]:** So how can anyone sell $1,178 of AI for 20? Because the average user doesn't come close? **[7:09]:** Semianalysis calls 20% a realistic average. And plans aren't a side business by Semianalysis. Rough numbers subscriptions bring Anthropic 10% of its revenue, but use 42% of its compute. And for OpenAI, Semianalysis says plans are an even bigger slice of revenue. A megawatt of chips earns about six times more serving the API than serving plans. **[7:34]:** Then they worked out the margins. Assuming the API itself runs at a 92% gross margin on Claude Pro with Opus Anthropic breaks even if you use 21% of your limit, use all of it and its margin is minus 371%. At that realistic 20%, its margin on the $200 plan is about 6%. On Fable, the model it rations, it's 80 chatgpt plus on Sol. Even at 100% OpenAI keeps a 16% margin. **[8:06]:** It doesn't lose money on you at all. So quick question for the how much of your plan do you actually use? Under a fifth? About half? Or do you hit the limit every week and that's why your plan keeps changing under you? **[8:19]:** Look at the last two weeks. OpenAI halved Pro200, and on Friday Google moves everyone without a paid plan to flash light its smallest Gemini model. Even its cheapest paid plan AI loses pro ChatGPT's free plan, for the record, still has unlimited chats on GPT 5.6 Luna, an older, smaller model. And Gemini's limits already work like a meter based on compute refilling every five hours up to a weekly cap. Anthropic squeezes More quietly when Opus 5.5 came out cheaper on the API, Claud Pro got 51% more tokens in API dollars. **[8:58]:** It's still worth 18% less. OpenAI did the same with Plus 19% more tokens, 19% less value and semi analysis caught one provider quietly testing 20% lower limits on a few accounts. They only learned it was a test by asking the price on the page doesn't move the meter behind it does so ChatGPT plus versus Claude Pro the same $20 plus gets you $211 of SOL, about a billion tokens, and OpenAI makes money on it, even if you use every last one. Claude Pro gets you $1,178 of Opus, almost 3 billion tokens, and anthropic starts losing money on you past about a fifth of it. **[9:41]:** So is ChatGPT+ worth it at 10 times its price? **[9:45]:** Yes, and it's probably the plan least likely to shrink because OpenAI loses nothing on it. But if you run a coding agent all day, Claude Pro is the better deal right now by more than five times. Just remember why it's so good for a heavy user. Anthropic is paying for most of it, and deals like that get cut. My rule? **[10:03]:** Pick the plan for the model you use most and watch the meter, not the price. I'm Nibel and this is kernel brief bite sized AI one layer at a time. If seeing behind the meter helps you choose, subscribe. We'll recheck these numbers the next time the limits move on a plan. Claude gives you 5 times more on the API it flips on 10 independent tests. Opus 5.5 cost 8 times more per task than GPT 6.1 SOL. That video is on your screen right now.