AI Agent Costs Rising Exponentially Despite Capability Growth
Original: Are the costs of AI agents also rising exponentially? (2025)
Why This Matters
Highlights potential disconnect between AI capability progress and economic viability
Researcher Toby Ord questions whether AI agent costs are growing exponentially alongside their capabilities. While METR data shows AI can handle longer tasks over 7 years, rising computational costs may offset productivity gains, potentially making AI less cost-competitive with humans despite technical progress.
Toby Ord raises concerns about missing cost analysis in AI development trends. METR data shows AI agents' task duration capabilities grew exponentially over 7 years - from GPT-2 handling seconds-long tasks to current models completing hours-long tasks 50% of the time. However, model sizes increased 4,000x and token generation grew 100,000x during this period. Ord proposes measuring 'hourly cost' - the financial cost of completing tasks at a model's 50% time horizon divided by human completion time. For example, Claude 4.1 Opus handles 2-hour human tasks, so its cost should be divided by 2 for hourly rate comparison. If costs grow faster than capabilities, AI systems become less economical despite technical advances, resembling 'Formula 1 of AI performance' - impressive but impractical.