GPT-6 Sol & Luna: Half the Price, Half the Mistakes – What It Means for How We Build Software
OpenAI halves API prices with GPT-6 Sol and Luna while promising significantly fewer errors. Why this matters more for small studios than a mere price cut.
On September 22, OpenAI introduced the GPT-6 generation — and it is far less spectacular and far more significant than the usual model launches: instead of new records, it is about the two factors that decide whether AI becomes standard in everyday development or remains a novelty. Price and reliability.
The numbers
GPT-6 Sol (the workhorse) costs $2 per million input and $10 per million output tokens, GPT-6 Luna (the small, fast variant) $0.10 and $0.50. That is at least 50 percent below the GPT-5.6 generation’s prices — combined with improved factuality: OpenAI reports significantly fewer user-flagged errors in coding scenarios.
The relationship between the two models is what’s interesting: Luna at its maximum tier beats the predecessor GPT-5.6 Sol at its medium tier in benchmarks — at a tenth of the cost. And Sol beats Claude Opus 5 on AutomationBench at roughly 9 percent of what Opus costs per task.
What “half the mistakes” means in practice
The real lever is not the price — it’s the error rate. AI-assisted development rarely fails today because of the models’ capabilities. It fails at the moment you stop trusting them. Every hallucinated API, every silently sabotaged refactor costs more review time than the generation saved. Halving the error rate doesn’t just change the cost math — it changes the trust question: tasks that used to be “AI drafts, human corrects” become “AI does, human spot-checks”.
Concretely, that means routine work — migrations, writing tests, catching up on documentation, boilerplate features — increasingly becomes a background task. Scarce human attention migrates to where it actually belongs: architecture, product decisions, edge cases.
The cost math for small studios
For studios running API-based products or agentic workflows, the token price simply halves. What used to be a budget argument against AI features in products (“it costs us cents per user, and those add up”) becomes a footnote for most use cases at Luna prices of $0.10/$0.50. API costs are hardly the limiting factor anymore — the engineering effort to integrate them well is.
Our perspective
At kofel, we use AI-assisted development every day — for kofel Studio, for client projects, and if you’re reading this blog, you may have read the results. What price points like these change: we can be far more generous about deploying agents for routine tasks that previously didn’t pay off economically. And products with built-in AI can let that feature drift from premium add-on to assumed standard.
The era of expensive tokens is drawing to a close. What remains is the real challenge: knowing what to fill the cheap tokens with.