AI订阅大摸底：Claude补贴比国产Coding Plan还狠

动察 Beating AI News Flash: SemiAnalysis tested the AI subscription plans of OpenAI, Anthropic, Meta, SpaceXAI, MiniMax, Moonshot AI, Zhipu, Cursor, and Cognition one by one. It separately tested how much subscription quota is consumed by input, cache writes, cache reads, and output, then calculated based on its own September Agent usage structure, of which about 96.6% was cache reads. Finally, it applied each company's API list prices to see how much it would cost if the same batch of tokens were sent directly through the API. SemiAnalysis calls this number the "API-equivalent value."

Claude's Sonnet and Opus are the most heavily subsidized. For the same $200 per month Max 20x, using Sonnet 5.5 can yield up to about $12,529 in API usage, equivalent to 62.6 times the monthly fee; Opus 5.5 is about $11,726, equivalent to 58.6 times. Switching to the more expensive flagship model Fable 5.1, however, leaves only $2,485, or 12.4 times. The same plan can differ by nearly 5 times because different models and different types of tokens consume backend quota at different rates. Anthropic has left a large amount of quota for Sonnet and Opus, while clearly tightening it for Fable.

Domestic Coding Plans also heavily subsidize power users, but their intensity still cannot catch up with Claude. SemiAnalysis measured that the $132 MiniMax Token Ultra can be converted into about $3,088 in MiniMax-M3 API usage, equivalent to 23.4 times the monthly fee; the $168 Zhipu GLM Coding Max is about 11.6 times; the $199 Kimi Code Ultra is about 6.7 times. Even among them, MiniMax, the highest, has a converted value per dollar of only about 40% of Claude Sonnet and Opus.

OpenAI, by contrast, has recently begun actively tightening subsidies. The token quotas for all tiers of models in newly purchased Pro 200 have been directly halved, while previously purchased plans can temporarily retain their old quotas until October 29. Now the $200 ChatGPT Pro using GPT-6.1 Sol has a converted value of about $2,084, equivalent to 10.4 times the monthly fee; Claude Opus at the same price is 5.6 times that. Even setting aside the difference in API unit prices between the two models and counting only actual tokens, the two are still about 28.6 billion versus 10.2 billion, with Claude still 2.8 times higher.

But when it comes to the two most flagship models, the gap is actually quite small. The $200 plan using GPT-6 Astra is worth about $2,897; Claude Fable 5.1 is about $2,485. However, Fable has a separate quota limit, and can consume at most half of the total Claude plan quota. After using up that $2,485 worth of Fable, users still have the other half of the quota available for Opus, Sonnet, and other models.

This kind of heavy subsidy is hard to sustain over the long term. SemiAnalysis estimates that the subscription business contributes only about 10% of Anthropic's revenue, yet may consume more than 40% of inference compute. Vendors can also directly change the backend billing ratios corresponding to different models and tokens without changing the monthly plan fee. When SemiAnalysis tested three identical plans, it found that one account had about 20% less quota; the corresponding vendor confirmed that this account had entered a very small-scale quota A/B test.

AI subscriptions look like a fixed monthly fee, but how much model usage you actually get can be rewritten in the backend at any time.