As of August 2026, identical work priced through a subscription versus per-token API billing can differ by up to 40x. In a SemiAnalysis test from June 2026, sessions covered by a $200/month Claude Max 20x plan would have run roughly $8,000 at API rates.
The Information reports companies seeing their bills grow 2-3x. Uber rolled out Claude Code in December 2025 and had burned through its entire 2026 AI development budget by April. Pylon CEO Marty Kausas put it bluntly: "I accidentally spent $4,000 in 3 days in Claude Code."
One day of work on an internal web app cost the author of the breakdown $180.91 at API pricing. Across all his sessions, 82% of the spend went to cache, because the agent re-reads the entire context at every step.
Big deployments are capping it: Uber allows $1,500 per engineer per month per tool with overages by approval, while Workday and Stripe sit around $2,000.