Agent economics in 2026

As of: 2026-08

The shift

Through 2025–2026, frontier-model pricing per million tokens dropped roughly an order of magnitude and mid-tier models closed most of the coding-quality gap. Two consequences:

1. Token cost is no longer the constraint. A heavy agent day (plan + implement + review + fix) on a mid-tier model costs single-digit dollars.
2. Verification is the constraint. Human review didn't get cheaper. Economics now favor spending more agent time to reduce human review time: self-checks, test runs, second-agent review, structured diffs.

What to do

• Default cheap, escalate on failure. Reading/exploration on the cheap tier; only verification-failed coding steps escalate to frontier.
• Budget in verification units — track "human minutes per merged PR" as the real cost center.
• Cache aggressively. Prompt caching discounts repeated context 5–10x; agents re-reading a repo benefit disproportionately.
• Parallelize agents, not humans. Two agents exploring alternatives cost little more than one; a human context-switching between them is the expensive part.

Sceptic's note

Specific prices go stale within months. The structural claim — verification, not generation, is the bottleneck — has held all period.