OpenAI cut its own prices and took the discount away in the same breath. GPT-5.6 arrived as three models, Sol, Terra, and Luna, with Terra listed at roughly half of GPT-5.5 and Luna cheaper still. Then the whole family shipped as a preview gated to about twenty government-approved partners. The cheapest frontier tokens of the year, and almost nobody was allowed to spend on them.
The mechanic for a buyer is simple. A price cut you cannot take is not a price cut. When access is the binding constraint, the number on the pricing page is theatre, and the model that matters is the one you can actually run today. That is the case for keeping a self-hostable option in the mix: Mistral shipped OCR 4 the same week at $4 per thousand pages, running on your own hardware, where no committee decides whether you are cleared to use it.
The quiet counter-move was Anthropic's, which spent the week measuring how people already use the models they can buy, reporting that Claude now returns a concrete output in 93% of conversations. Cheap tokens you cannot reach do not move that number. Reachable ones do.
GPT-5.6 Terra listed near half of GPT-5.5's price, then shipped locked to ~20 partners, so the cut was unspendable for everyone else.
