The math doesn’t work. Not at full utilization.
AI research firm SemiAnalysis published empirical findings from stress-testing every major subscription tier available from OpenAI and Anthropic. The methodology: buy the plans, run extended coding sessions and agentic workloads, and calculate what the consumed tokens would have cost at published API rates. The result for a fully-utilized $200/month ChatGPT Pro plan was up to $14,000 in API-equivalent token value. Claude Max at the same price point hit up to $8,000.
That’s a 40-to-70x subsidy gap. Not 10x, which has been the rough rule of thumb for AI subscription pricing relative to API access, but 40 to 70x, at the high end of utilization.
The important framing caveat: the $14,000 and $8,000 figures represent API-equivalent token cost, not the labs’ actual compute cost. API pricing includes margin. The true cost to OpenAI or Anthropic to serve a maximally active subscriber is likely lower than the API-equivalent figure, but directionally, the economic exposure is the same. Heavy users are expensive.
Analysis
The $14K and $8K figures represent API-equivalent token cost, not the labs' actual compute cost to the server. API pricing includes margin. The true cost to serve a maximally active subscriber is likely lower, but the directional exposure is the same: heavy agentic users are expensive regardless of which benchmark you use.
The real story is the distribution problem. Not every $200 subscriber runs extended agentic sessions all month. Most don’t come close to the utilization ceiling. The labs are betting that enough subscribers are light users who subsidize the heavy ones, keeping blended unit economics acceptable. SemiAnalysis’s research surfaces the risk in that bet: as agentic workflows become mainstream, the light-user buffer erodes. Coding agents, research agents, and document-processing pipelines run continuously in ways that consumer chatting never did.
Secondary coverage of the SemiAnalysis findings corroborates the headline figures. The specific utilization thresholds at which each plan enters negative gross margin territory are in the paywalled primary SemiAnalysis report and weren’t reproduced in accessible secondary sources, so those percentages aren’t included here.
This connects directly to what enterprise teams are already experiencing. Two weeks into token billing, enterprise AI teams are hitting costs they didn’t budget for, that’s the buyer’s side of the same dynamic. SemiAnalysis’s research is the lab’s side: the subscriptions funding those budget overruns are themselves being subsidized at rates that can’t survive aggressive agentic adoption.
SemiAnalysis infers that economic pressure may push labs toward restricting their most compute-intensive future models to API access only. No lab has announced that move. It’s an analyst projection, not a reported plan.
What to Watch
What to watch
any changes to the $200 tier structure from either OpenAI or Anthropic, rate limits, model access restrictions, or the introduction of separate agentic usage pricing. Those signals would confirm that the subsidy math has become untenable for high-utilization users. A flat-fee subscription model built for chatting doesn’t naturally fit agentic workloads billed by the task.
TJS synthesis: SemiAnalysis has quantified what the enterprise billing data was already suggesting: the $200 flat-fee model is a customer acquisition instrument running at a loss for the heaviest users, and agentic workloads are turning more subscribers into heavy users. Watch for a structural change to how the $200 tier works before the end of 2026, either a usage tier above flat-fee, model-gating, or a hybrid that prices agentic tasks separately. The labs can’t publish the SemiAnalysis math and leave the pricing unchanged.