Voices
Calvin French-Owen: 'Small models have arrived', and the same personalized-news task went from ~$1 to ~$0.10
The Segment co-founder argues on calv.info (August 26, 703 points on Hacker News) that fast cheap models like gpt-5.6-luna at ~100 tps and GLM 5.3 now sit at the capability-per-dollar frontier, and that a task costing ~$1 on Sonnet-class models now runs at ~$0.10. He splits work into 'IQ 180' breakthrough tasks that still need frontier models and responsive operational work that is most of what businesses actually do, and puts the second category firmly in small-model territory. He names the missing infrastructure explicitly: prompt injection safety, roles, and permissions.
↳ Follow the thread