What's this about?
OrangeAI relayed that a relay service has delisted the Luna model because serving it loses money every day — the price is so low that the margin no longer covers the surrounding cloud costs. His conclusion: this is approximately the floor for model pricing, and below it the only remaining direction is on-device.
What exactly is the cost that isn't the model?
When people compare inference prices they compare the per-token rate, but a service around a model has costs the model doesn't: bandwidth, storage, load balancing, monitoring, support, payment processing, fraud handling. Those are roughly fixed per request regardless of how cheap the tokens are. So there's a floor where the token price approaches zero while the surrounding cost doesn't move, and reselling stops making sense at any volume. That's the point this describes reaching.
Why is this worth noticing?
- It puts a floor under the price war that has run all month — GPT-5.6 Sol cut 50%, Luna cut 80%, Gemini 3.7 Flash at introductory pricing. Someone in the chain stops being able to absorb it first, and it's the resellers.
- "Below this, the only place left is on-device" connects directly to today's FreeToken paper, which is about making a personal machine a viable serving platform — the same economics from the other end.
- It complicates the cheap-model story. A model can be too cheap to intermediate, which means low headline prices don't automatically translate into a healthy ecosystem around them.