What OpenAI put behind the preview wall
OpenAI has opened a limited preview of GPT-5.6 in three sizes. Sol is the flagship at 5 dollars per million input tokens and 30 dollars per million output tokens. Terra sits in the middle at 2.50 and 15 dollars, and OpenAI says it matches the capability of last year's GPT-5.5 flagship while costing about half as much. Luna is the fast, cheap tier at 1 dollar in and 6 dollars out, aimed at high-volume work where latency and price matter more than peak reasoning.
The catch is access. During the preview the three models run only through the OpenAI API and the Codex coding tool, and only for a small group of trusted partners and organisations that have an account representative. This is not a self-service launch. OpenAI says general availability will follow in the coming weeks, and the release ships with what the company calls its most robust safety stack yet, with tighter controls on higher-risk and cyber requests and on repeated misuse.
The price of a fixed capability keeps halving
The number that matters for a budget is not the flagship price but the price of last year's flagship this year. Terra delivers GPT-5.5-class capability at roughly half the token cost, which means a workload that was economic to run on 5.5 is now half as expensive to run at the same quality. That pattern has held across recent generations, and it is the single most useful assumption a European builder can carry into a three-year plan: the price of a fixed level of AI capability keeps falling by roughly half each cycle.
The practical move is to design around the trajectory rather than today's list price. A feature that looks marginal at Sol's rate may be comfortably profitable at Terra's, and clearly profitable at Luna's once volume grows. Teams that lock their architecture to a single model tier miss this, because the cheapest sensible way to run a given task keeps shifting down the ladder. The cost curve, not the launch headline, is the thing to plan against.
Why the newest model is not on sale yet
The second signal is about access, not price. The most capable new models now arrive partner-first, gated behind an account relationship rather than a public endpoint, and only reach general availability weeks later. That is a two-tier market forming in plain sight: well-connected organisations test and build on frontier capability before anyone can buy it, and everyone else waits for the general release. For a smaller European firm, that gap is a real competitive variable, not a footnote.
The discipline this calls for is simple. Treat a preview-only model as a signal of where the cost curve is going, not as a dependency you can ship on, because you cannot buy it at scale and its terms can change before general availability. Build production systems on models that are generally available and priced, and use the preview tier to plan the next move rather than to run the current one. The capability is real, but the thing you can put in a contract is what counts.
Read next: Microsoft Bets 2.5 Billion That Tools Are Not Enough | OpenAI Offers Washington a Stake
