Agent Economics (The Agents Season, Episode 10)

Summary of Agent Economics (The Agents Season, Episode 10)

by Ben Jaffe and Katie Malone

24m•June 22, 2026
Listen to Episode

Overview of Agent Economics (The Agents Season, Episode 10)

In this episode, Ben Jaffe and Katie Malone explain why AI agents can become more expensive overall even as model inference gets cheaper. Using the classic economics idea of Jevons’ paradox, they argue that lower per-token costs are driving much higher usage, more complex workflows, and ultimately rising total spend on AI systems. The episode focuses on the real economics of agentic tasks: multi-step reasoning loops, context growth, failure costs, and the wide gap between median and top-end enterprise AI spending.

Main Idea: Jevons’ Paradox and AI

What Jevons’ paradox means

  • When a resource becomes cheaper and more efficient, people often use much more of it.
  • The result can be higher total consumption and higher total spend, not lower.
  • Historical examples used in the episode:
    • Coal in 19th-century England
    • Electricity
    • Computing
    • New York highway expansion as an analogy for traffic growth

Why it matters for AI agents

  • LLM inference costs have fallen sharply.
  • But AI usage has exploded even faster.
  • The hosts argue that this is a textbook case of Jevons’ paradox: cheaper AI makes new use cases viable, which increases overall demand.

Why AI Agents Are So Expensive

1. They are multi-step workflows

AI agents don’t just answer one prompt. They often loop through:

  • reasoning
  • action/tool use
  • observation
  • repeated iteration

Each loop adds more model calls, tool calls, and context retrieval.

2. Costs grow nonlinearly

  • Every new step typically includes the entire conversation history as context.
  • That means prompt size and compute cost can grow roughly quadratically with the number of turns.
  • Long agent runs can become very expensive very quickly.

3. Failures still cost money

  • You pay for unsuccessful attempts too.
  • In real tasks, many agent runs won’t succeed on the first try, so failure cost becomes part of the economics.

Key Data and Examples

LLM costs vs. demand

The episode cites an analysis suggesting:

  • Inference costs fell by about 1,000x from 2022 to 2026
  • Token demand rose about 10,000x
  • Net result: overall AI spend increased

DeepSeek as an example

  • The January 2025 release of DeepSeek showed how a cheaper, open model can:
    • expand the range of use cases
    • reduce reliance on hosted APIs
    • trigger big market reactions
  • The hosts mention NVIDIA’s sharp market cap drop that week as evidence of how seriously the market viewed the shift.

Gartner estimate

  • A typical agentic task costs 5–30x more than a standard LLM chatbot task.

Coding task cost example

From a 2026 ICSE paper on SWE-bench-style coding tasks:

  • Leading agents often need 41–58 turns
  • Some complex tasks exceed 175 turns
  • Claude Sonnet examples:
    • $5.85 average cost for a patch attempt
    • $7.80 average cost for a correct patch
  • This highlights how unsuccessful attempts add real overhead.

Real-World Spending Patterns

Ramp’s AI spend data

The hosts highlight observed corporate card data from Ramp showing a highly skewed distribution of AI spending.

Monthly AI spend per company

  • Median: $2,246
  • Average: $140,842
  • 75th percentile: $14,000
  • 90th percentile: $73,000
  • 95th percentile: $211,000
  • 99th percentile: $831,000

Per-employee spend

  • Median firm: about $11 per employee/month
  • Top 10%: about $611 per employee/month
  • Top 1%: about $7,450 per employee/month

What this means

  • Most companies are spending relatively modest amounts.
  • A small tail of heavy AI users is spending very large sums, pulling the average way up.
  • The distribution suggests that AI spending is still in an early, uneven phase with plenty of room to grow.

Enterprise Examples and Signals

Uber

  • Reportedly set an annual token budget for developers
  • Blew through it by April
  • Later aimed for about $1,500 per developer per month

Microsoft

  • The hosts mention reports of similar cost-control efforts
  • Some engineers reportedly had access pulled back due to high AI usage costs

Takeaway

These companies are outliers, but they show how quickly AI costs can scale in high-usage environments.

Core Takeaways

What to anchor on

The most important number is not just per-token cost. It’s:

  • cost per completed task
  • success vs. failure rate
  • number of turns required
  • complexity of tool use
  • volume of tasks attempted

Bottom-line argument

  • Cheaper inference does not necessarily mean cheaper AI usage overall.
  • As agents become more capable and more widely used, total spend can rise even while unit costs fall.
  • The hosts believe Jevons’ paradox is likely to continue applying to AI.

Closing Thoughts and Episode Notes

Extra anecdote

Ben shares a strange personal story about interacting with an Anthropic model called Fable, including a conversation after it had supposedly been deactivated. He uses it as a teaser for more discussion in the newsletter.

Next episode preview

The season finale will focus on:

  • an AI agent Ben built to help produce the podcast
  • what AI helps with in his workflow
  • what he intentionally keeps human
  • what works and what doesn’t when using AI for production

Bottom Line

This episode argues that the economics of AI agents are being shaped less by falling model prices and more by expanding usage, longer workflows, and rising task complexity. In other words: as AI gets cheaper, people build more things with it, run it more often, and spend more overall. That is Jevons’ paradox in action.