Calvin French-Owen argues cheap, fast models make consumer and business AI economics viable
- Calvin French-Owen says gpt-5.6-luna runs at about 100 tokens per second in his testing and can search thousands of emails for API costs in the tens of cents.
- His personalized daily-news experiment cost about $1 with Sonnet-class models, but averages about $0.10 with Luna, changing the viability of a consumer subscription product.
- French-Owen argues that most business work is "token spewer" work: responsiveness, follow-ups, and moving many tasks forward, rather than rare "IQ 180" technical breakthroughs.
- He expects frontier-model demand to continue in engineering, hard science, and model training, while demand grows for fast, cheap, good-enough models used in routine business work.
- The author says business deployment still needs harnesses, prompt-injection defenses, roles, and permissions before small models can handle more organizational work.
Hacker News 의견들
I have had good results with Luna, but I want comparisons that put end-to-end response time, cost, and performance at different settings side by side. TPS alone does not answer whether Luna Max beats Terra Medium on value.
People chasing frontier models are late to this. Small models have been good enough for many tasks for a while, especially if you do not have a Fable-sized budget.
Composer 2 and 2.5 were already cheap, fast, and very capable while everyone focused on GPT 5.5. A 32B model running locally on consumer hardware is good enough for some workflows too.
For my recent work, Sol trips up about as often as Luna by over-assuming or over-engineering simple API exploration scripts. I would rather course-correct Luna on small tasks because the wasted tokens cost less.
I expect more products like Replit, which is already leading with free Luna usage.
I do not see big models as products by themselves. Inference is a new form of compute, and most products only need reliable tool calling, accurate recall, and instruction following. Small models will likely cover most of that, while hardware demand stays high even if hosted inference demand falls.
I had this realization in early 2024 with a 7B local model and Guidance. I gave it test pseudocode, had it write tests for approval, then write code until they passed. Even if models stopped improving, there would still be years of useful system-building to explore.
Running small models at the core of a business mostly shifts external frontier-model API calls into internal APIs for locally run models. Cloud providers need to make local-model deployment much cheaper and easier, though DigitalOcean already has an option.
The two work categories resemble Paul Graham's maker and manager schedules. I am interested in which roles models take over and how they let people spend more time in the mode they prefer.
Large models bundle world knowledge, language ability, and reasoning primitives. In applications that need little world knowledge or language nuance, intelligently deployed small models can beat naively deployed large ones.
Small, cheap hosted models are already accessible. Luna at $0.20 input and $1.20 output pricing is more than adequate for summarization and simple tool calls.