Calvin French-Owen: the same task went from about $1 to about $0.10

The Segment co-founder published "Small models have arrived" on August 26, and it took 703 points on Hacker News (calv.info). His measurement: a personalized-news task that cost about a dollar on Sonnet-class models now runs at about a dime. Ten times cheaper, doing the job well enough that he stopped reaching for the frontier model.

His frame is a split between "IQ 180" work and operational work. Breakthrough tasks, the ones where you need a model to find something genuinely hard, still want the frontier. Responsive operational work, which he argues is most of what businesses actually do, sits comfortably in small-model territory now. He names gpt-5.6-luna at around 100 tokens per second and GLM 5.3 as sitting on the capability-per-dollar frontier.

I'd push back on the split a little. It's clean but it assumes you know in advance which bucket a task falls in, and in my experience the classification is the hard part. A task that looks operational turns out to need one genuinely hard inference in the middle, and a cheap model fails it in a way that's expensive to detect. The routing problem French-Owen's frame implies is real work, not a free lunch.

What makes his post more than one person's anecdote is how much of this week's material lines up behind it independently. Zhipu released GLM-5.3-Flash under MIT: 320B total, 18B active, natively multimodal, with Z.ai claiming it beats GLM-5.2 across benchmarks at a tenth of the price and approaches Claude Opus 4.8 on coding and agentic work (transformers v5.16.1). Databricks separately claims 10% better quality than GLM-5.2 at a tenth of the cost, though that's vendor-supplied and unverified. Kwindla Hultman Kramer open-weighted PhoneLLM, a 30B/3.5B-active voice-agent MoE claiming GPT 5.6 Terra parity at 94% lower cost with 1,300ms faster P95 time-to-first-token, at about $0.00025 per agent-minute (Hugging Face). And DeepSeek now bills off-peak rates at exactly 50% of peak for about 79% of the week (paddo.dev).

French-Owen also names what's missing, and it's the same list this newsletter keeps writing about: prompt injection safety, roles, permissions. A dime-per-task model that will execute anything a webpage tells it to isn't cheaper. It's cheaper per attempt and unbounded per incident.

Practical move for this week. Take one non-interactive job in your stack, the nightly summarizer or the classifier or whatever runs on a cron, and swap the model to GLM-5.3-Flash or an equivalent small model. Keep the old output for a week and diff. If quality holds, you just cut that line item by 90%. If your job runs on DeepSeek at all, move the cron from 09:00 UTC to 11:00 UTC and the bill halves with identical latency. That's a config change, not an engineering project.