Subscription plans remain the primary way consumers and small businesses pay for AI, despite being heavily subsidized. These plans serve as powerful customer acquisition and marketing tools, as evidenced by OpenAI's generous resets boosting Codex adoption and forcing Anthropic to reverse subscription nerfs.
Subscription plans, even with a small revenue share, significantly impact margins and revenue per MW due to high subsidization. For Anthropic, subscriptions are 10% of revenue but consume over 40% of inference compute, lowering blended revenue per MW by ~$36M. For OpenAI, subscriptions are even more critical as they constitute a larger portion of their total revenue.
Understanding subscription limits is crucial for accurately modeling AI lab financials. Subscriptions grant "credits" for various (model, token type) combinations, with credit cost ratios differing from API price ratios, making the "value" of a plan dependent on the specific model and workload.
SemiAnalysis offers a Subscriptions Dashboard tracking limits for OpenAI, Anthropic, Meta, SpaceXAI, Cursor, Cognition, Z.ai, MiniMax, and Moonshot, with data updated daily to reflect changes in real-time.
The methodology involves running experiments to isolate token types and track meter movements to calculate the rate of each token type. Results are converted into tokens per window and per month, then priced to determine the API-equivalent value of each subscription plan.
OpenAI's recent changes include halving the API-equivalent value of their $200 plan and introducing a new $500 tier, which still offers lower value per dollar compared to Anthropic. Anthropic, on the other hand, offers consistent per-dollar value across all subscription tiers, currently providing a much better deal than OpenAI.
Both companies aim to reduce subscription subsidies, with Anthropic lowering the API-equivalent value for premium models as they release more powerful ones, while OpenAI cut limits across the board to Fable-level.
Chinese subscription plans offer varying levels of subsidization, with value per dollar averaging slightly less than OpenAI's ~12x. Third-party plans like Cursor and Cognition offer less generous limits and lower value per dollar compared to the first-party plans of OpenAI and Anthropic.
