A new SemiAnalysis analysis highlights a central problem for the AI industry: powerful models attract users, but they are expensive to operate at scale. According to the report, a fully used ChatGPT Pro subscription, which costs $200 a month, could cost OpenAI about $14,000 a month. A similarly priced Claude Max plan from Anthropic could reach roughly $8,000 in processing costs.
The company tested long coding jobs and more complex agentic AI workflows, which can consume far more tokens than a normal question. That matters because every extra token adds demand on data centers, GPU time, and energy use. The report says OpenAI’s margins are especially fragile, with some plans turning unprofitable above 11.4% usage and its most advanced plans reaching that point at 5.7%. Anthropic’s lower and mid-tier plans appear more resilient, but the same basic pressure remains.
The broader industry is already adapting. Microsoft, Meta, and Amazon have reduced some internal programs that encouraged unlimited AI use. Many firms now mix expensive models for harder work with cheaper open-source systems for simpler tasks, and that shift can cut costs sharply. At the same time, cheaper models such as DeepSeek are gaining attention, with Lindy CEO Flo Crivello saying his company moved from Anthropic models to DeepSeek V4 and saved millions of dollars. Even so, the strongest frontier models may still be too costly for simple flat-rate pricing, which is why some observers expect them to move toward usage-based API billing instead of standard subscriptions.