SemiAnalysis maxed out every Anthropic and OpenAI subscription tier and found a $200 per mon...
SemiAnalysis ran an experiment that put a real number on something the AI industry has been hinting at for months. They maxed out every paid tier from Anthropic and OpenAI and measured the actual compute draw against the subscription price. The result is the clearest evidence yet that the consumer AI business is being run as a land grab, not a margin business.
A $200 per month Claude Max subscriber can cost Anthropic up to $8,000 in monthly compute. A ChatGPT Pro subscriber on the same plan tier can cost OpenAI up to $14,000. That is a loss of 40 to 70 times the subscription price on the heaviest users. The labs know this and are absorbing it on purpose. Power users are the ones writing benchmarks, building viral demos, and shaping the public read on which model is at the frontier, so keeping them locked in is worth the burn.
On the same day, the Wall Street Journal reported that OpenAI is considering significant token price cuts on its API. The timing reads as a preemptive move against an expected Anthropic price drop, and it lands right as OpenAI approaches its IPO window. Consumer tiers are already underwater, and now API pricing, which has been the one place these companies could plausibly defend gross margin, is heading down too.
For anyone building on top of these APIs, the math shifts in two directions. If your product was priced against current API rates with a thin gross margin, a 30 to 50 percent token reduction is a tailwind worth planning for now rather than reacting to later. Capacity you treated as a premium tier, things like long context windows, high reasoning effort modes, or top tier multimodal calls, will get cheaper faster than most roadmaps assume. The complex routing logic many teams built to downshift to smaller cheaper models for cost reasons may stop being worth the engineering overhead it costs to maintain. Caching layers that exist purely to dodge token spend fall into the same category.
The quieter signal underneath all of this is that frontier labs are pricing for market share, not unit economics, and they will keep doing so for as long as the capital markets let them. The era where inference cost was the binding constraint on what you could ship is closing out. What replaces it is harder to predict, but it will probably look like competition on latency, reliability, tool use quality, and the surrounding product surface rather than on raw token price. Worth watching whether Anthropic responds to the rumored OpenAI cut quickly, because that will tell us how synchronized this race to the bottom actually is.