SemiAnalysis tested AI subscription plans from OpenAI, Anthropic, Meta, SpaceXAI, MiniMax, Moonshot AI, Zhipu, Cursor, and Cognition, measuring how plan quotas are consumed by input, cache writes, cache reads, and output. It then applied its own September agent usage mix, in which about 96.6% of activity came from cache reads.
From there, the firm priced the same batch of tokens using each company’s API rates and called the result the plan’s "API-equivalent value."
Claude Sonnet and Opus showed the largest subsidy
According to the comparison, Anthropic offered the most generous effective subsidy on Claude Sonnet and Opus.
On the same $200-per-month Max 20x plan, Sonnet 5.5 could reach about $12,529 in API usage, equal to 62.6x the monthly fee. Opus 5.5 came to about $11,726, or 58.6x.
The more expensive flagship model, Fable 5.1, dropped to about $2,485, or 12.4x. SemiAnalysis said the nearly 5x spread inside the same subscription tier comes from differences in how backend systems charge for different models and different token types. In its reading, Anthropic leaves substantial quota for Sonnet and Opus while applying much tighter limits to Fable.
Domestic coding plans also subsidize heavy users, but trail Claude
SemiAnalysis said domestic Coding Plan products also provide sizable subsidies for heavy users, though still below Claude’s level.
- MiniMax Token Ultra, priced at $132, translated to about $3,088 of MiniMax-M3 API usage, or 23.4x the monthly fee.
- Zhipu’s $168 GLM Coding Max came to about 11.6x.
- Kimi Code Ultra, priced at $199, came to about 6.7x.
Even MiniMax, the highest among those plans, delivered only about 40% of the per-dollar value seen in Claude Sonnet and Opus.
OpenAI has tightened quotas on newly purchased Pro 200 plans
SemiAnalysis said OpenAI has recently started pulling back on subsidies. Token quotas across model tiers for newly purchased Pro 200 plans were cut in half. Plans bought earlier can keep the old quotas for now, until Oct. 29.
Under the current setup, a $200 ChatGPT Pro subscription using GPT-6.1 Sol translates to about $2,084 in value, or 10.4x the monthly fee. A Claude Opus plan at the same price point is 5.6x that figure.
Even if API price differences between the two models are excluded and only actual token volume is compared, the gap remains wide at about 28.6 billion versus 10.2 billion, leaving Claude ahead by 2.8x.
The gap narrows at the flagship-model tier
At the top end of each company’s lineup, the difference was much smaller. A $200 plan using GPT-6 Astra translated to about $2,897 in value, while Claude Fable 5.1 came to about $2,485.
Fable, however, has a separate cap. It can consume no more than half of the total Claude plan quota. After using that $2,485 worth of Fable capacity, a subscriber still has the other half of the quota available for models such as Opus and Sonnet.
Fixed monthly pricing does not mean fixed usable capacity
SemiAnalysis said subsidies at this level are difficult to sustain over time. Its estimate showed that subscriptions contribute only about 10% of Anthropic’s revenue while potentially consuming more than 40% of inference compute.
The report also said providers can change backend charge ratios for different models and token types without changing the monthly subscription price. While testing three identical plans, SemiAnalysis found one account had about 20% less quota. The provider involved confirmed that the account had been placed into a very small-scale quota A/B test.
Its conclusion was that AI subscriptions may look like fixed-fee products, but the amount of model usage a customer actually gets can still be rewritten from the backend at any time.