Inside the claim that Anthropic subscriptions offer five times OpenAI’s value

How SemiAnalysis measured subscription value

AI subscriptions usually show customers a usage meter rather than a price for each input, output or cached token. SemiAnalysis tried to reverse-engineer those limits by repeatedly making calls designed to isolate one token type and watching how much each request moved the meter.

Because a single request may not move the meter at all, the researchers measured complete steps between meter changes. They discarded the partial steps at the beginning and end of a run, then kept collecting data until their estimated rate fell within a range of plus or minus 5%.

The team used those rates to estimate how many tokens fit inside each plan’s short-term and weekly limits. It then applied public API prices to the estimated token allowance for a chosen workload. The result is what the article calls “API-equivalent value,” or what the same mix of tokens would cost through an API.

For the agentic workload, SemiAnalysis used its own September mix of input, output, cache-write and cache-read tokens. That choice matters. A different workload can produce a different result because providers charge and meter token types differently.

Where the five-times comparison applies

SemiAnalysis found similar overall limits for GPT-6 Astra and Fable 5.1, but it said Fable could use only half of an Anthropic plan’s limit. In one $200 comparison, the publication estimated that an Anthropic subscriber could consume $2,485 of Fable 5.1 API-equivalent value and still have half the plan remaining, while the comparable OpenAI plan would be exhausted after $2,897 of Astra use.

The larger gap appeared in the mid-tier comparison. SemiAnalysis estimated that Anthropic’s Opus 5.5 offered about five times the API-equivalent value of OpenAI’s GPT 6.1 Sol across the subscription tiers it examined. It said the difference remained substantial when comparing raw token volume rather than API-equivalent dollars.

The article also reported that OpenAI had recently cut the API-equivalent value of its $200 plan by about half and introduced a $500 tier. SemiAnalysis said older $200 plans would retain their previous limits until Oct. 29, while newly purchased plans received the lower limits immediately.

Why the headline needs a qualifier

API-equivalent value is not the same as model quality, task success or user value. A model with a lower API price can look worse under this calculation even when its subscription includes a large number of tokens. Token efficiency can also change how much work a user gets from the same allowance.

The limits themselves are moving targets. SemiAnalysis said providers can change them through promotions, model launches or adjustments to internal credit costs. During testing, it found one account with limits about 20% below two otherwise similar subscriptions; the provider told the publication that the account was in a very small A/B test.

The practical takeaway is not that one provider always offers five times more. It is that subscription economics can vary sharply by the exact model and workload, and that a plan’s headline price reveals little about the quantity or kind of usage it actually buys.