SemiAnalysis said tests of subscription usage meters at Anthropic and OpenAI show a large gap in API-equivalent value for the companies’ mid-tier models, even as their flagship tiers come out close. In results published on Oct. 5 U.S. time, the firm said a full month of Claude Max 20x using Claude Opus 5.5 translated to about $11,726 at official API list prices, while ChatGPT Pro 200 using GPT-6.1 Sol translated to about $2,084. That puts Claude’s result at roughly 5.6 times the OpenAI figure under that setup.
How SemiAnalysis measured subscription value
The firm treated each subscription as a pool of credits and then estimated how quickly different models and token categories consumed that pool. It converted the monthly allowance into what it called an API-equivalent value, or the amount a user would pay if the same usage were billed through the official API instead of a subscription. SemiAnalysis said that figure is not the customer’s actual bill and not the company’s actual cost.
Its testing method isolated one token category at a time: fresh input, cache write, cache read, and output. It repeatedly called the models with prompts designed for measurement and recorded how many tokens had been consumed whenever the official usage meter moved by one unit. It reported numbers only after reducing the error range to within ±5%.
Mid-tier models produced the widest gap
The workload used in the comparison was an agentic workflow, where AI systems execute multi-step tasks on their own. In that setup, most tokens came from cached input, meaning content that had already been processed in the conversation and could be reused at the lowest price tier. SemiAnalysis used its own September usage mix, with cached input accounting for 96.6%.
At the flagship level, the two companies were close, with OpenAI slightly ahead. GPT-6 Astra on ChatGPT Pro 200 produced an API-equivalent value of $2,897, while Claude Fable 5.1 on Max 20x came in at $2,485. SemiAnalysis added one qualification: Fable can consume at most half of the plan’s total allowance. After that $2,485 worth of usage is exhausted, the Claude subscription still retains half its allowance for other models, while the ChatGPT plan is fully spent.
The larger divergence appeared in the mid-tier products each company positions for regular use. On the $200 plans, Opus 5.5 versus GPT-6.1 Sol came out at $11,726 versus $2,084. The same pattern appeared in other tiers: Max 5x at $5,725 versus ChatGPT Pro 100 at $1,055, and Claude Pro at $1,178 versus ChatGPT Plus at $211.
Because Sol already carries a lower official list price than Opus, SemiAnalysis also compared the plans in token terms rather than dollar terms. It said the monthly totals were about 28.6 billion tokens versus 10.2 billion tokens, leaving Claude ahead by about 2.8 times even on that basis.
Three conditions behind the “5x” claim
SemiAnalysis said the headline multiple should not be detached from its assumptions. If a model’s list price is unusually high or low, API-equivalent value can become distorted. It also said the industry still lacks reliable data on how many tokens the same task really consumes across platforms. For that reason, the roughly fivefold result rests on three conditions: mid-tier model comparisons, API-equivalent pricing as the yardstick, and an agentic workload.
OpenAI’s $200 plan was cut on Sept. 29
The report said OpenAI had often been described by independent developers as generous on limits, but argued that description no longer holds. On Sept. 29, OpenAI cut the allowance on its $200 plan in half. Thibault Sottiaux, who leads Codex, explained the move publicly on X, and SemiAnalysis said its own tracking measured the same scale of reduction.
Astra’s API-equivalent value fell from $5,734 to $2,897, a 49% drop, while monthly token capacity declined from 3.8 billion to 1.9 billion. Sol fell more sharply, from $5,904 to $2,084, a 65% decline. SemiAnalysis said token capacity there was down 48%, and GPT-6.1 Sol also saw a cut in cached-input pricing without a corresponding restoration in allowance.
The newly introduced $500 plan charges 150% more per month, yet Astra usage is only 21% higher than the old $200 plan, according to the report. SemiAnalysis also said Pro 200 previously delivered about twice the value per dollar of Pro 100, but after the limit cut, Pro 100, 200, and 500 now offer the same number of tokens per dollar.
One point in OpenAI’s favor, it said, is that Pro plans do not have a five-hour usage cap, making it easier for subscribers to use their full monthly allowance. Even so, SemiAnalysis concluded that this did not offset what it described as roughly a 4x API-equivalent advantage for Opus 5.5.
What the subsidy says about compute and margins
SemiAnalysis estimated that subscriptions account for about 10% of Anthropic’s revenue but consume 41.7% of its inference compute. On a revenue-per-megawatt basis, which it used as a measure of how much revenue each unit of power and compute can generate, subscriptions produce about $16 million compared with about $102.9 million for API usage. By that estimate, subscriptions reduce overall revenue per MW by about $36 million.
The margin picture looked worse under the report’s assumptions. Using a 92% API gross margin assumption, SemiAnalysis said a Max 20x subscription fully used on Opus 5.5 would imply a gross margin of negative 369%, while a plan fully used on Fable 5.1 would imply 1%. At a 20% average utilization rate, those figures would be 6% and 80%, respectively. The report said these were internal estimates, not financial figures disclosed by Anthropic.
Anthropic and OpenAI are tightening subsidies in different ways
SemiAnalysis said both companies appear to be adjusting subscription subsidies, but not with the same method. Anthropic’s approach is less direct: the more expensive the model, the lower the API-equivalent value. On Max 20x, Fable 5.1 works out to 12.4 times the monthly fee, while Opus 5.5 reaches 58.6 times. OpenAI, by contrast, was described as cutting its allowance in one move to a level close to Fable.
The report ended by saying the current fivefold gap is a snapshot, not a promise. SemiAnalysis expects that a model tier above Fable would likely come with lower subscription allowances, and might not be included in subscriptions at all. It also said Chinese vendors still subsidize their plans, with average API-equivalent value per dollar slightly below OpenAI’s roughly 12x level.