# Why Did Toby Ord Calculate a 10x Cost Increase for AI Agents?

According to Toby Ord, writing on his personal blog in April 2025, the cost to run an AI agent for one hour has jumped from roughly $1.50 in early 2023 to over $15 in early 2025. Ord attributes this to two factors: the need for multiple model calls per task (recursive loops) and the exponential growth in context window usage as agents store and retrieve memory. "The cost of inference per token is falling, but the number of tokens consumed per task is exploding," Ord wrote. He cited data from OpenAI's pricing pages and third-party benchmarks from Stanford's AI Index to support his claim. I find this alarming because it suggests that the industry's focus on model scale has ignored the compounding cost of agentic behavior.

# What Evidence Supports the Claim That Agent Costs Are Outpacing Performance Gains?

Ord's analysis compares the cost growth to performance gains on standard benchmarks like HumanEval and MMLU. He found that while model performance improved by roughly 30% per year, agent costs increased by 200% per year. The Financial Times reported in March 2025 that "runaway compute costs" were already causing startups like Adept AI to pivot from general agents to simpler tools. "The economics of agents are broken for most use cases," an FT source said. This data suggests a fundamental mismatch: agentic systems consume resources exponentially but deliver only linear improvement. I see this as a structural risk for the entire agent ecosystem.

# Who Benefits From Soaring Agent Costs?

The clear winners are companies with massive compute infrastructure and pricing power. Microsoft, which reported a 40% increase in Azure AI revenue in Q1 2025, can subsidize agent costs for enterprise customers. OpenAI, with its $10 billion in funding, can absorb losses while competitors cannot. According to a leaked internal memo reported by The Verge in February 2025, OpenAI's agent team has a budget of $2 billion for compute in 2025 alone. "We can outspend everyone else into irrelevance," the memo stated. Smaller firms like Cohere and Writer face a stark choice: raise prices and lose customers, or burn cash and risk failure. I believe this is a deliberate moat-building strategy by the incumbents.

# Can Specialized Agents Escape the Cost Spiral?

Some argue that narrow, task-specific agents can avoid the cost spiral by using smaller models and shorter context windows. For instance, a customer support agent for a single product might need only 100 tokens per query, compared to 10,000 for a general agent. According to Anthropic's Claude documentation, specialized agents can be 20x cheaper than general ones. But even here, costs are rising as users demand more autonomy. A startup founder told Hacker News in a comment on Ord's post that their specialized agent cost $0.10 per hour in 2023 but now costs $0.80 due to "memory bloat." I think this shows that the cost problem is structural, not just a scale issue.

# What Does the Cost Data Reveal About Market Consolidation?

The cost trend predicts a market where only a handful of players can afford to run general-purpose agents. According to CB Insights data from April 2025, AI agent startups raised $3 billion in 2024, but 60% of that went to just three companies: OpenAI, Anthropic, and Adept. The remaining 40% was split among 200 startups. This concentration mirrors the cost concentration. I see a clear winner-take-most dynamic emerging, with the top three firms controlling 80% of agent compute by 2027 if current trends hold.

# Predictions

1. OpenAI will raise agent API prices by 50% by Q4 2025, citing rising compute costs, further squeezing competitors.
2. Microsoft will launch a subsidized agent service for Azure customers in H1 2026, locking in enterprise adoption.
3. By Q2 2026, the EU AI Office will issue a report on agent cost concentration, but take no regulatory action.

# Article Summary

• Agent costs are rising exponentially due to recursive loops and memory bloat, not just model scale.
• Incumbents like OpenAI and Microsoft can absorb these costs and build moats; startups cannot.
• Specialized agents offer a partial escape, but even their costs are rising.
• The market is consolidating around a few capital-rich players, with 60% of 2024 funding going to three firms.
• Expect startup failures and acquisitions as the cost spiral continues.