Tokenmaxxing guide urges routing AI tasks to cheaper models

Key points

• Meta staff used 60 trillion AI tokens in 30 days chasing an internal leaderboard nicknamed "Claudeonomics".
• Routing a 1,000-invoice job to a cheaper model cut its cost from $30 to $3.90, the article says.

A tokenmaxxing guide from Decrypt argues that most people waste money by sending every job to a flagship AI model when a cheaper one would do.

The article traces the waste to Silicon Valley leaderboards. At Amazon, staff admitted inflating their numbers with irrelevant tasks, and Uber reportedly spent its 2026 AI tools budget in about four months.

Decrypt's advice starts with switching to a smaller model for simple jobs and testing it against five recent tasks. It also recommends one clear prompt instead of several rounds of corrections, starting a fresh chat when the subject changes, capping output length and using free or local models where possible.

For developers, the guide covers agent costs: each loop resends the whole conversation, and orchestration across subagents multiplies the spend. Caching, batch processing at a 50% discount and hard budget caps all help. Decrypt notes that token prices have fallen more than 90% since 2023, yet spending on large language models has doubled since late 2025.

Read the full story on Decrypt →