OpenAI cut GPT-5.6 Luna pricing by 80% this week, down to $0.20 per million input tokens, the same week it crossed 1 billion weekly active users and 2 million business customers. Layer on this: AI agent usage among professionals is now at 80.8% daily-or-more, up from 47.3% a year ago.
This isn't a capability story anymore, it's a cost-curve story. When inference gets this cheap, use cases that were previously too expensive to justify start to pencil out:
- Always-on agents that poll and react continuously instead of running on a schedule
- Batch processing large document sets that used to get sampled instead of fully processed
- Embedding an LLM call into every step of a workflow instead of reserving it for the "hard" steps
Why this matters for what you're building
If your architecture still treats an LLM call as a scarce, rationed resource, gated behind caching layers and strict rate limits, it's worth revisiting that assumption this quarter. The unit economics that justified those guardrails six months ago may no longer hold.
The interesting engineering question isn't "can we afford to call the model here" anymore. It's "what breaks if we call it ten times more than we do today, and is that actually a good idea."
Curious what others are seeing: has this pricing shift already changed anything in your stack, or is the calculus still catching up to the sticker price?
Top comments (0)