Tokenflation: When "Hi" triggers 33 tool calls
5 hours ago
- Tokenflation describes the phenomenon of coding agents using excessive tokens (context, reasoning, tool calls) for simple tasks like responding to 'Hi', which can trigger up to 33 tool calls and significant waiting time.
- The real cost of tokenflation is not just API token fees (often < $0.01) but the waiting time, valued at about $0.016 per second based on a $120k salary, making total costs like $0.84 for a 'Hi' response on models like Sonnet.
- In a benchmark of 14 models, ambiguous prompts ('WTF' or 'Hi') caused more tool calls, timeouts, and failures than clear tasks ('commit'), with waiting time dwarfing API costs (e.g., 99% of bill for some models).
- Labs focus on leaderboard metrics like SWE-bench, ignoring efficiency on everyday tasks, and vendors' per-token billing creates a perverse incentive for more exploration, leading to tokenflation in real-world use.