Claude plans give about 5x the AI usage of ChatGPT's, SemiAnalysis finds

SemiAnalysis tested every subscription plan from nine companies: Anthropic, OpenAI, Meta, SpaceXAI, MiniMax, Moonshot, Z.ai, Cursor and Cognition. The report is by Andrew Megalaa, Max Kan and Dylan Patel. It came out on October 5.

The headline is the matchup between the two biggest US labs. SemiAnalysis compared the models each one markets as the daily driver for most users: Anthropic's Claude Opus 5.5 and OpenAI's GPT-6.1 Sol.

"Anthropic is an overwhelmingly better deal, offering ~5x the API-equivalent value across the board."

That is the firm's own summary. It adds a fair caveat: GPT-6.1 Sol is much cheaper per token than Opus 5.5. But it says the gap is still large even when you count raw tokens instead of dollars.

There are two ways to pay for AI. You can pay a flat monthly fee, like $20 for a chat plan. Or you can pay per use through the API, the way developers do, where every token has a price.

A token is a small chunk of text, roughly part of a word. API-equivalent value asks a simple question: if you used up your whole plan, what would those tokens cost at the pay-per-use price?

Think of it like an all-you-can-eat buffet. The plan is the entry fee. API-equivalent value is what the same plates would cost if you ordered them one by one from the menu.

Here is the tricky part. Neither company tells you how many tokens a plan includes. You only see a usage meter from 0 to 100%, over a 5-hour window and a 7-day window.

So SemiAnalysis worked backward from the meter:

• One token type at a time: each test used a prompt built to burn mostly one kind of token, such as fresh input, cached input or output.
• War and Peace as the text: some models refuse to answer large blocks of gibberish, so the firm used a portion of the novel instead.
• Counting in steps: a single request often doesn't move the meter, so it tracked how many tokens it took to move the meter by one unit.
• Tight margins: it kept adding steps until the estimate was within plus or minus 5%.
• Free cache reads: if 500 million cached tokens never moved the meter, it treated them as free.

Then it applied a typical agentic workload, the kind of long-running task where AI works through many steps, based on its own September usage. Finally, it priced the result at each company's API rates.

By the numbers: here is what SemiAnalysis's figures show for the everyday models, per month, on the agentic workload.

Counted in plain tokens instead of dollars, The Register reports that Claude Pro gives about 2.9 billion tokens a month, against 1 billion for ChatGPT's $20 plan.

At the very top, the picture is closer. SemiAnalysis says limits are "quite similar across the board" for OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1.

One wrinkle: Fable can only use 50% of a Claude plan's limit. So a $200 Claude plan still has half its limit left after $2,485 of Fable 5.1, while the matching OpenAI plan is used up after $2,897 of Astra. That spare half is what makes Opus 5.5 such a good deal on Claude plans.

In real life If you pay $20 a month for an AI chatbot and mostly use the default model, this study says your plan's real size depends a lot on which company you picked. Check your plan's usage page before a big project, not halfway through it.

What changed: last week, OpenAI halved the number of tokens its $200 plan gives for each model tier, and SemiAnalysis's tracking confirmed the 50% cut. OpenAI also launched a new $500 tier.

Here is what that looks like in practice:

• Old customers: $200 plans bought before the cut keep the old, higher limits until October 29. New buyers start with the lower limits right away.
• The $500 plan: it offers only 21% more GPT-6 Astra usage than the old $200 plan. Its headline feature is a 300 tokens-per-second Ultrafast mode, which SemiAnalysis is still testing.
• Pricing page: OpenAI used to advertise its plans as "5x more usage than Plus" and "20x more usage". That relative language is gone from its pricing page.
• Same value per dollar: after the change, the $100, $200 and $500 Pro plans all give the same tokens per dollar for every model.

Anthropic went the other way, a little. When it cut Opus 5.5's API prices (20% off input and output, 60% off cache reads), it raised Opus usage by about 20% on Max plans and about 50% on Pro. SemiAnalysis says that still didn't fully match the price cut. You can read more about the newest Claude lineup in our story on Claude Sonnet 5.5.

OpenAI does have one edge. None of its Pro plans has a 5-hour limit, only a weekly one, so it's easier to use your whole allowance. SemiAnalysis says it doesn't think that makes up for the gap.

The big picture: subscriptions are a small slice of revenue but a big slice of computing power. SemiAnalysis estimates that subscriptions are about 10% of Anthropic's revenue but can take up over 40% of its inference compute, the chips that answer your questions. The Register puts it at about 42%.

That pulls down how much money each chunk of data center capacity earns. SemiAnalysis estimates the hit at roughly $36 million in lower revenue per megawatt for Anthropic. It says subscriptions matter even more at OpenAI, where they are a bigger part of total revenue.

So why offer such generous plans? Dhruv Amin, co-founder of AI agent company Skydive, told The Register that the labs "subsidize their prosumer plans with subscription pricing massively" as a marketing funnel. People get comfortable with the tools at home, then bring them to work.

Yes, and the study caught it happening. While testing, SemiAnalysis found that one of three identical accounts had about 20% lower limits than the other two. It was also the oldest account.

The provider told the firm this was an "extremely tiny" A/B test on limits. It said it "didn't just decrease limits wholesale" and was testing "how to better balance when people hit limits". SemiAnalysis draws two lessons: providers can quietly change limits at any time, and its method is sensitive enough to notice.

No, and the report says so itself. It argues you can't say a plan is worth a set amount on its own: it depends on the plan, the model and the work you do.

Other yardsticks tell a different story. Artificial Analysis tracks the cost of finishing a specific task at API prices. By that measure, The Register notes, GPT-6.1 Sol ($0.72) is far cheaper than Claude Opus 5.5 ($5.98). SemiAnalysis also says the industry lacks reliable data on how efficiently each model uses tokens.

The catch for businesses: they often can't lean on subsidized plans. The Register cites a July McKinsey report that found 93% of enterprises went over their AI budgets. Amin said his company cut its own AI costs by more than 75% by routing work to cheaper open models such as GLM-5.3 Flash. More on the company behind Claude is on our Anthropic topic page.

The Register says OpenAI did not immediately respond to its request for comment.

• Pick by model, not just price: the same $20 can buy very different amounts of AI depending on which model you use most.
• Heavy ChatGPT $200 users: if you bought before the cut, your old limits run until October 29. Plan any big projects with that date in mind.
• Watch your meter: limits can change quietly, so a plan that felt roomy last month may not be the same this month.
• Businesses: subscription math doesn't carry over to the API, where cost per task can point the other way.