$/task beats token price: outputs are <5% of tokens in agentic coding
xeophon · x · 2026-09-23
xeophon argues token price is the wrong metric: for many use cases (notably agentic coding), output tokens are <5% of total tokens, while >90% are cached inputs—which also dominate what people actually spend. Token efficiency matters too, so $/task is the best yardstick, in his view. A substantive addition to the debate over Chinese models' rising prices.
Related event: Cost per task, not token price, is the real AI metric(2 posts)→
More from coding & agent
- Firecrawl raises $75M for Alexandria, an agent knowledge library that pays publishers — emmanuelvivier · 2026-09-23
- Claude Opus 5.5 adds time budget: let the model decide how long to work on a task — JeremyNguyenPhD · 2026-09-23
- Rabbit Launches OS3, a Cloud AI Agent That Controls Windows, Mac and Linux Remotely — emmanuelvivier · 2026-09-23
- AI coding reignites testing debate: bigger test units are fundamentals, not dead — mattpocockuk · 2026-09-23
- Firecrawl Raises $75M for Alexandria, a Knowledge Library for Agents That Pays Publishers — emmanuelvivier · 2026-09-23
- Why is dormant Aider (no commits in 4 months) still the top open-source harness LLMs recommend? — marlene_zw · 2026-09-23