AI Agent Costs May Be Rising Exponentially, Challenging Economic Viability

Toby Ord's December 22, 2025 analysis questions whether AI agent costs are increasing exponentially alongside performance improvements. While METR data shows AI task completion times growing dramatically—from seconds with GPT-2 to hours with current models—the financial implications remain under-examined. Model parameters have expanded 4,000-fold over seven years, with token generation increasing 100,000 times. Ord introduces 'hourly cost' metrics comparing AI expenses to human software engineering rates, revealing that some models approach or exceed human pricing at peak performance. Analysis of METR's GPT-5 chart shows AI agents exhibit diminishing returns, with sweet-spot hourly costs ranging from $0.40 for Grok 4 to $40 for o3. At plateau performance, costs escalate significantly: GPT-5 reaches $120 per hour for two-hour tasks, while o3 costs $350 hourly for 1.5-hour horizons. The correlation between time horizon and cost suggests economic feasibility may lag behind technical capabilities. METR's benchmarking methodology, which prioritizes peak performance over cost efficiency, complicates direct cost analysis. Ord notes discrepancies between OpenAI's reasoning models and competitors' pricing, with potential data estimation issues affecting comparisons. The investigation reveals few researchers are examining cost trajectories, creating uncertainty about whether AI agents will maintain cost competitiveness with human labor as capabilities advance.