Software developer Calvin French-Owen has argued that a newer class of fast, inexpensive AI models could change the economics of consumer applications and routine business work. In an essay describing his own experiments, he points to GPT-5.6 Luna and GLM 5.3 as examples of smaller systems becoming useful enough for tasks that do not require the strongest available reasoning.
French-Owen reported seeing GPT-5.6 Luna produce about 100 tokens per second while working across code, email and a personal knowledge base. He said complex research runs generally cost tens of cents, including tasks that searched thousands of emails. These figures reflect his own usage rather than a controlled comparative study.
Cost is the core of his case. Consumer internet businesses have often attracted users before developing paid products, but AI features add an inference expense whenever a model processes a request. That creates a recurring marginal cost that search or social services built around conventional software did not face in the same way.
To illustrate the issue, French-Owen uses a personal evaluation that asks a model to research him, identify likely interests and build a daily news microsite using material from online communities. He estimates that a previous generation of models cost about $1 per run, making a daily product difficult to support with a $30 monthly subscription. With Luna, he says an adequate result averages roughly $0.10, reducing that illustrative monthly inference bill by about 90%.
The essay does not claim that smaller models have displaced frontier systems. French-Owen says he still chooses higher-cost models such as Fable 5 and GPT-5.6 Sol for coding work, and expects demand for top-tier capability to keep growing in engineering, science, model training and other areas that depend on novel problem-solving. He instead identifies a separate market for systems optimised around responsiveness, price and sufficient competence.
His business argument follows a similar distinction. Much executive and operational activity, he writes, consists of calls, reminders, coordination and other routine follow-through rather than rare technical breakthroughs. A fast model that handles those repeated tasks reliably may therefore create value even if it performs less well on the hardest evaluations.
Several obstacles remain before that idea can be broadly deployed. French-Owen highlights the need for better agent software, defences against prompt injection, and careful definitions of roles and permissions. Those controls become important when models can access email, internal knowledge or operational systems rather than simply answer isolated prompts.
The essay is ultimately a forecast, not evidence that a new consumer market has already emerged. Its contribution is a concrete economic hypothesis: as the price of acceptable inference falls, product ideas that were unattractive at dollar-level daily costs may become viable at cents per task. Whether those savings survive real-world reliability, safety and support requirements remains an open question.
