AI model costs vary 100-fold; companies overspend by defaulting to flagship models
Decrypt reports that top AI models cost 100 times more per token than smaller ones, yet many users default to expensive flagships, with Meta burning 60 trillion tokens in 30 days and Uber exhausting its 2026 AI budget in four months.
Why it matters
The story highlights significant cost inefficiencies in enterprise AI adoption, with companies potentially overspending by orders of magnitude by not routing tasks to appropriately sized models. This has direct implications for AI infrastructure budgets and token-based pricing models.
Key facts
Anthropic's and OpenAI's top models cost 100 times more per token than their smallest models
100 timesReported by Decrypt
Meta employees burned through 60 trillion AI tokens in 30 days earlier this year
60,000,000,000,000 tokensReported by Decrypt
Gergely Orosz estimated Meta's token usage at roughly $900 million at Anthropic's list prices
900,000,000 USDReported by Decrypt
Uber burned through its 2026 AI tools budget in about four months
Reported by Decrypt
Jue Wang of Bain & Company told the Associated Press that many companies default to Opus for everything including generating emails
Reported by Decrypt
What happened
Decrypt reports that Anthropic's and OpenAI's highest-end models cost roughly 100 times more per token than their smallest offerings, yet many users and companies default to the flagship models for all tasks. Meta employees consumed 60 trillion AI tokens in a 30-day period earlier this year while chasing spots on an internal leaderboard nicknamed "Claudeonomics," which tech columnist Gergely Orosz estimated at about $900 million at Anthropic's list prices, though Meta likely pays less. Uber reportedly burned through its 2026 AI tools budget in approximately four months. Bain & Company consultant Jue Wang told the Associated Press this summer that she sees similar waste across many companies, with users defaulting to Anthropic's Opus model even for simple tasks like generating emails. The article explains that every word entered into ChatGPT or Claude and every word returned is metered in tokens, which average about three-quarters of a word, determining usage limits on subscriptions and bills on API access.
How the story developed
- First report
First report by Decrypt
TokenMaxxing: How to Get the Most Out of Your AI Models
- Update
Decrypt publishes article on AI model cost optimization
Decrypt publishes 'TokenMaxxing: How to Get the Most Out of Your AI Models' detailing cost disparities between AI models and enterprise overspending
- Update
Meta employees consume 60 trillion tokens in 30-day period
Internal leaderboard 'Claudeonomics' drives massive token usage at Meta earlier in 2026
- Update
Uber exhausts 2026 AI tools budget in four months
Usage leaderboards push AI bills up at Uber
- Update
Jue Wang comments to Associated Press on widespread AI model overspending
Bain consultant notes many companies default to expensive Opus model for simple tasks
Affected entities
Original sources
| Publisher | Report | Role | Published |
|---|---|---|---|
| Decrypt Crypto media | TokenMaxxing: How to Get the Most Out of Your AI Models | Original |