SemiAnalysis Priced a Maxed-Out Month of ChatGPT Pro vs Claude Max. WILD

This is the video written out, with every figure in full, each linked to its source in the table below.

Another argument, and a measurement▶ 0:00

ChatGPT has Plus, and Pro at three prices: $100, $200 and, since 29 September, $500. Claude has Pro and two Max plans. There are five-hour limits, weekly limits, and limits that change when a new model comes out. It's a lot to keep up with, and there's always another argument about which plan is the better deal.

This time somebody measured it. SemiAnalysis tracked the usage limits on ChatGPT's and Claude's plans. On the models they say each lab pitches for everyday use, they put Claude's $200-a-month plan at more than five times the value of ChatGPT's $200 plan: $11,726 a month against $2,084.

Five times what, though?

Whose numbers these are▶ 0:41

SemiAnalysis publishes a model of AI labs' finances called the Tokenomics Model. The report, “Anthropic Subscriptions Offer 5x+ More Value Than OpenAI”, is by Andrew Megalaa, Max Kan and Dylan Patel, published at 20:01 UTC on 5 October 2026. We read the free part of it.

We didn't measure anything ourselves. Every figure on this page for what a plan's limits hold or are worth is SemiAnalysis's, as published, and where we divide or multiply their figures, the result is marked as our arithmetic. We read the report again on 21 October and on 26 October, and it hadn't changed: the free part's text, its dates and every chart image were the same.

To say which $200 plan gets you more, SemiAnalysis first had to work out how much a plan gives you at all, because the plans don't say. A plan shows you a usage meter, not a number of tokens.

How they read the usage meter▶ 1:11

The meter runs from zero to a hundred percent. There's one for the week and, on some plans, one for every five hours too, and sometimes a separate one for a model like Fable. Send a single request and the meter often doesn't move, because it moves in fixed steps, like whole percents.

So SemiAnalysis kept a running count of the tokens they sent and wrote the count down each time the meter went up a step. The tokens between two steps are what one step of that meter costs. They threw away the half-finished steps at the start and the end, and kept adding steps until the range of their estimate was within 5% either way. They read every meter a plan shows this way. If reading from the cache still hadn't moved the meter after 500M tokens, they treated cache reads as free.

They did this for one kind of token at a time, because each kind uses up the meter at its own rate:

• new text the model hasn't seen, with a random tag on every call so nothing gets cached;
• text being saved into the provider's cache, with the prompt marked for caching and a new tag each time;
• the same text read back from the cache, with a fixed tag, so the first call writes the cache and every repeat reads it;
• text the model writes itself, the most expensive kind.

For the first three they sent slices of War and Peace, since some models refuse to read gibberish. For the fourth they asked for long technical essays, up to 16K tokens a call, because models refuse a request like “repeat SemiAnalysis 100,000 times”.

From tokens to dollars▶ 2:22

A five-hour meter resets many times a week, so the weekly limit is what caps a month, and SemiAnalysis counts a month as 4.345 weeks. Each kind of token uses up the meter at a different rate, so how many tokens fit in a month depends on the mix. SemiAnalysis takes the mix from their own token usage in September, the pattern of a coding agent at work: 96.6% of the tokens are the model re-reading context it has already seen, 2.6% are saved into the cache, 0.4% are new input and 0.3% are output.

With that mix, the weekly limits on Claude Max 20x, the $200 plan, come to 28.6B tokens of Opus 5.5 a month. To turn that into dollars, SemiAnalysis prices the tokens at the API's list prices, what you'd pay per token without a subscription. At Opus 5.5's prices that mix costs $0.4099 per million tokens, about $0.41. So the same month of work that costs $200 on the plan would cost $11,726 through the API. SemiAnalysis calls that the plan's API-equivalent value.

It assumes you use up your whole weekly limit, every week of the month.

Opus 5.5 against GPT-6.1 Sol▶ 3:31

SemiAnalysis compares the models that, in their words, both labs market as “the intended daily driver for most users”: Claude Opus 5.5 and OpenAI's GPT-6.1 Sol. On the $200 plans, Opus comes to $11,726 a month and Sol to $2,084, so the Claude plan holds 5.6× as much. The $100 plans come to $5,725 and $1,055 (5.4×), and the $20 plans to $1,178 and $211 (5.6×). At every price, SemiAnalysis's 5 October figures put the Claude plan at more than five times the API value of the ChatGPT plan.

Counted in tokens▶ 4:05

SemiAnalysis raises an objection to their own comparison. Sol is much cheaper per token than Opus, so a comparison in dollars could be unfair to OpenAI. That's why they count tokens too. Claude Max 20x gives 28.6B Opus tokens a month and ChatGPT Pro 200 gives 10.2B Sol tokens, so the Claude plan gives 2.8× as many. At $100 the token gap is 2.7×, and at $20 it's 2.9×.

The two gaps come from the same sum. 28.6B Opus tokens at $0.4099 per million make $11,726. For Sol, in the same mix, 10.2B tokens at about $0.20 per million make $2,084. (SemiAnalysis prints Sol's price rounded: $2,084 over 10.2B tokens is $0.204 per million by our arithmetic.) At these prices each Opus token is worth 2.05× as much as a Sol token (our arithmetic), so about 2.8× the tokens at about twice the price per token gives the gap of more than five times.

SemiAnalysis's answer to their own objection is that the gap is still massive in tokens. They call API-equivalent value “the single best number for capturing the ‘value’ of a subscription plan”, and say raw tokens can be the better measure when a model is a particularly bad or good deal at API prices. They don't say whether Sol is one of those.

What neither number counts▶ 5:20

Neither the dollar figure nor the token count says how many tokens each model needs to finish the same job. SemiAnalysis calls that token efficiency “an extremely relevant factor”, and says the industry lacks reliable data on it. Many people cite a chart from Artificial Analysis for this, they note, but in their words, “we do not believe the benchmark tasks in the AA Intelligence Index are at all representative of real work people do with LLMs.”

About six hours after the report went up, at 02:24 UTC on 6 October, Tibo, who works on Codex and ChatGPT at OpenAI (his profile reads “Codex & ChatGPT @OpenAI”), posted: “The real story is this one.” He was sharing a post by Meet Pandya that said “Counting only equivalent API cost would show only half the picture IMO.”, next to an Artificial Analysis chart in which an Intelligence Index task costs far less on GPT-6.1 Sol than on Opus 5.5. The video doesn't give that chart's figures, and this page doesn't either.

The flagships: GPT-6 Astra and Fable 5.1▶ 6:09

On each lab's flagship model, the two plans come out much closer, with ChatGPT a little ahead. GPT-6 Astra on ChatGPT Pro 200 comes to $2,897 a month, and Claude Fable 5.1 on Claude Max 20x to $2,485. The $100 plans are almost level: $1,322 and $1,273. On ChatGPT Plus, Astra comes to $162. Fable isn't part of Claude Pro's usage limits at all. Anthropic's Help Center says Pro users can use it with usage credits, paid for separately.

Now use each $200 plan's flagship model as far as the plan allows. Astra can use the whole ChatGPT plan, and it's used up at $2,897. On Claude's Max plans, Fable may use up to 50% of the weekly limits at no extra cost, so on Max 20x Fable's share runs out at $2,485. In SemiAnalysis's words, the Claude plan “would still have 50% left after consuming $2,485 worth of Fable 5.1”, and that half is there for Claude's other models, Opus 5.5 and Sonnet 5.5. The part of the report we read doesn't put a dollar figure on that half.

Potion, our podcast tool▶ 7:21

Potion is our podcast tool. If you watch this channel, you probably follow a few topics more closely than any podcast does. With Potion, you ask your AI, like ChatGPT or Claude, for a podcast episode on exactly what you want to understand. Potion narrates it and puts it in your podcast app, so you can listen on the go.

It works with ChatGPT and Claude, even the free versions, and it's free to start. Try Potion if you'd like your own podcast feed, managed by your AI.

Why the labs land so far apart▶ 8:08

Why are the two labs so far apart on the everyday models, and so close on the flagships? SemiAnalysis starts from money. They estimate that subscriptions bring Anthropic about 10% of its revenue but use about 42% of its inference compute (10.0% and 41.7%), and they call those rough estimates. They say subscriptions matter even more to OpenAI, where they make up a larger share of revenue.

Both labs, SemiAnalysis says, want to subsidize subscribers less, and they've picked two different ways to do it. Anthropic does it by making each plan worth less on its more expensive models. Its Help Center says Fable models use a plan's weekly limits faster than Claude's other models, so far fewer Fable tokens fit: 3.2B a month in the half of Max 20x's limit that Fable may use, against 28.6B of Opus in the whole limit. Fable's blended price in this mix is $0.78 per million tokens, about 1.9× Opus's, but that doesn't make up the difference. Claude Max 20x comes to $11,726 a month in API value on Opus 5.5, and $2,485 on Fable's half of the limit.

SemiAnalysis also prices Sonnet 5.5 on the same plan. On 7 October, after the report, Anthropic halved Sonnet 5.5's price for cache reads, from $0.20 to $0.10 per million tokens, and cache reads are 96.6% of this mix, so SemiAnalysis's Sonnet figure no longer matches today's list prices. We leave it out, as the video does. Opus 5.5's and Fable 5.1's prices hadn't changed when we checked on 21 October.

SemiAnalysis then estimates, using their own assumptions, how much of the fee Anthropic keeps after paying to serve a Max 20x subscriber. If the subscriber uses 20% of what the plan allows on their model, a figure SemiAnalysis calls a more realistic average, the gross margin is about 6% on Opus and about 80% on Fable. If they use all of it, it's −369% on Opus and +1% on Fable. These assume the API itself earns a 92% gross margin.

OpenAI took what SemiAnalysis calls “the nuclear option”, cutting every model at once to about the value per dollar Anthropic gives Fable.

OpenAI's cut on 29 September▶ 9:33

OpenAI did that on 29 September. It cut the tokens on its $200 plan in half, for every model. Tibo announced it at the time, saying it would come to half the API value of the old plan, and SemiAnalysis's tracking confirmed it. For GPT-6 Astra the value roughly halved along with the tokens, from $5,734 a month to $2,897 (−49%), and the tokens from 3.8B to 1.9B.

OpenAI's everyday model on that plan also changed, from GPT-6 Sol to GPT-6.1 Sol, and its value fell further, from $5,904 to $2,084 (−65%), while its tokens fell 48%. It's the same sum as before, tokens times price, and both parts shrank: the tokens fell by about half, and GPT-6.1 Sol has a cheaper price for cached input, which is nearly all of this workload. SemiAnalysis's blended price for Sol in this mix went from $0.30 to $0.20 per million tokens.

OpenAI also added a $500 plan, ChatGPT Pro 500. It gives 21% more Astra than the old $200 plan did, $6,955 a month against $5,734, for two and a half times the price. On Sol it gives more tokens than the old plan but less API value, because of GPT-6.1 Sol's lower price.

Subscribers who already had the $200 plan when the cut came keep the old limits for a while. OpenAI's Help Center has the exact rule. A Pro 200 subscription that was active at any point from 22 September through 10 a.m. Pacific Time on 29 September keeps the previous allowance through 29 October 2026, as long as it stays active.

In SemiAnalysis's words, OpenAI “used to be lauded by indie developers for their generous subscription limits compared to Anthropic, but this is simply no longer true today.” They do give OpenAI one point. None of ChatGPT's Pro plans has a five-hour limit, so it's easier to use more of your month. SemiAnalysis doesn't think that makes up the difference.

How far to trust it▶ 11:04

While refining the method, SemiAnalysis found that one of three identical subscriptions they were testing with one provider had about 20% lower limits than the other two. They asked the provider, which SemiAnalysis doesn't name, and were told the account was part of an “extremely tiny” A/B test, a test where some users get different limits. SemiAnalysis takes two things from that: providers can silently change limits at any time, and their method is sensitive enough to notice.

So these numbers are a snapshot, as published on 5 October, from SemiAnalysis's own accounts, priced for their own coding workload. They also assume the whole weekly limit gets used every week. If you use less, you get less of that value. And the comparison is for this coding mix only: the free part of the report doesn't compare the two labs' plans on a chat workload.

Update, 26 October: on 22 October, SemiAnalysis posted on X that Anthropic appears to be A/B testing Claude Max 20x limits. They measured two Max 20x accounts on the same day, both on Opus 5.5. One came to $11,726 a month, the figure above, and the other to $9,385, with a five-hour window 19% larger but a week 20% smaller. Against ChatGPT Pro 200's $2,084 from the report, $9,385 is 4.5× as much (our arithmetic), where $11,726 is 5.6×. SemiAnalysis doesn't say whether this is the A/B test they described in the report, and Anthropic hadn't confirmed it when we checked on 26 October.

Which plan, then▶ 11:51

If you mostly use the everyday models, SemiAnalysis's 5 October numbers put Claude's plans at more than five times the API value of ChatGPT's, and about 2.8× the tokens, at every price. If you want each lab's flagship model, ChatGPT's $200 plan is a little ahead, because on Claude's Max plans Fable can use only half of the weekly limit, which leaves the other half for Claude's other models. Fable isn't in Claude Pro's usage limits at all. If you had ChatGPT's $200 plan when the cut came on 29 September, you keep the old limits through 29 October, as long as you stay subscribed.

Since the report came out, Anthropic has added monthly API credits to its Max plans, for running your own apps and agents on the Claude Platform: $100 a month on Max 5x and $200 on Max 20x, announced on 7 October. They cover the Claude API, Claude Managed Agents, the Claude Agent SDK and the playground, not Claude Code. SemiAnalysis's numbers don't include them.

And all of it can change.

SemiAnalysis's Subscriptions Dashboard▶ 12:49

SemiAnalysis tracks those changes in a Subscriptions Dashboard that comes with their Tokenomics Model. Their report describes it as re-checking each plan's limits daily, and it covers more providers than this video does. The free part of the report ends where it turns to OpenAI's and Anthropic's models on third-party plans like Cursor and Devin. The rest is for their paid subscribers.

This isn't sponsored, and we have no relationship with SemiAnalysis.

The date still ahead▶ 13:14

The one date still ahead in SemiAnalysis's report is 29 October, when the old limits run out for anyone who had the $200 plan when the cut came. Which plan are you on, and does this match what you actually get out of it? Tell us in the comments under the video.