The bottom line: Fully utilized flat-rate subscriptions for ChatGPT and Claude generate costs that exceed monthly fees by a multiple, forcing providers to rely on low average utilization rates while enterprise customers increasingly switch to cheaper alternatives or in-house systems.
According to an analysis by SemiAnalysis, fully utilized ChatGPT Pro subscriptions can incur costs of up to $14,000 per month for OpenAI, despite only receiving $200 in fees. The profit margins of AI providers therefore depend heavily on usage remaining well below the limits.
According to a SemiAnalysis calculation based on standard API rates, OpenAI faces a massive disparity between subscription fees and actual compute costs. The $200/month ChatGPT Pro subscription could incur token-based API costs of up to $14,000 per month if usage limits are fully exhausted. Anthropic’s comparable offering Claude Max 20x, also priced at $200 per month, reaches a theoretical cost ceiling of approximately $8,000.
Profitability thresholds fall well short of maximum utilization. OpenAI already records losses at average utilization rates of just 11.4 percent on standard tiers ChatGPT Plus and ChatGPT Pro 5x. For Anthropic, the profitability threshold for Claude Pro and Claude Max 5x sits at approximately 20 percent utilization. At the most expensive premium tiers, the situation deteriorates sharply: OpenAI enters loss territory at 5.7 percent usage rate, Anthropic already at 10 percent. A growing cost driver is autonomous AI agents, which consume up to 1,000 times more tokens per request than standard inputs.
Large technology companies have already responded to these cost dynamics. Microsoft, Meta, and Amazon have scaled back or restructured internal deployment of intensive AI applications. In one documented case, unrestricted use of Claude models within a single organization incurred monthly costs of $500 million. As a countermeasure, many organizations are implementing dynamic load balancing systems: complex queries are deliberately routed to expensive high-end models, while routine tasks are handled by cheaper alternatives. According to reports in the Wall Street Journal, operating costs can be reduced by up to 95 percent through this selective steering.
Some technology companies have already completely overhauled their infrastructure. The AI assistant startup Lindy, for example, shifted all its traffic from Anthropic models to the DeepSeek V4 model, which, according to CEO Flo Crivello, maintained performance while drastically reducing costs and saving the company millions of dollars. Other organizations are investing in developing their own systems based on open-source models trained on internal company data to ensure cost control.
Experts expect that operating costs for mid-tier performance models could eventually decline to around $20 per month through infrastructure expansion. However, the most advanced flagship models will remain extremely costly to operate.
Source: www.it-daily.net · Published June 16, 2026
Lumi AI News — AI-assisted curation per Article 50 EU AI Act. Paraphrase and classification by Lumi News Pipeline v1.7.1.