When OpenAI launched the GPT-5.6 family on July 9, 2026, Luna — the family's fast, cost-efficient tier — was priced at $1 per million input tokens and $6 per million output tokens. That pricing lasted exactly three weeks. On July 30, OpenAI slashed Luna by 80%, landing at $0.20 input and $1.20 output — a $1.40 combined price that represents a staggering drop in what frontier-grade AI inference now costs.
The cuts didn't stop there. Terra, the mid-tier model, fell 20% from $2.50/$15 to $2/$12 per million tokens. Only Sol, the flagship, held its position at $5/$30 per million tokens.
OpenAI's official explanation points to "improvements in system efficiency." That's almost certainly true — hardware costs fall, inference kernels improve, and distillation techniques allow cheaper models to punch above their weight. But the competitive context is equally important.
The repricing came days after Anthropic launched Claude Opus 5 at $5/$25 per million tokens, matching Sol's capability range at a slightly lower price. And Google shipped Gemini 3.6 Flash at $1.50/$7.50 that same month, directly undercutting Luna's original pricing. The message from the market is clear: frontier model pricing has been in freefall since 2025, and no company can afford to anchor its cheaper tiers at yesterday's premium.
At $0.20/$1.20, tasks that were previously economically impractical — bulk document classification, large-scale data enrichment, high-volume customer support — now become viable with a model that still carries the GPT-5.6 architecture's reasoning headroom. Luna isn't a downgraded model; it's a 1.1M context window, frontier-built system now at budget pricing.
The rapid repricing also has a structural message for builders. The right response isn't to optimize hard for one model at a locked price point — it's to build with model-agnostic abstractions so switching costs stay low when (not if) the economics shift again. OpenAI cutting Luna this fast, this aggressively, is less a gift and more a signal: every lab is under pressure, and this won't be the last repricing.
For end users, the story is simpler: the cost of building serious AI-powered products just dropped another floor. That's good news any way you cut it.