The useful way to review GPT-5.6 is not to pretend I have had a week of quiet, equal access to it. I haven't, and most people haven't. OpenAI's own Help Center still describes the family as a limited preview through the API and Codex for selected organisations, with ChatGPT excluded during the preview and no general availability date announced. The public launch on July 9 matters, but the important part is the shape of the release: official enough to price, name, document, and benchmark, still gated enough that "released" needs quotation marks around it. I wrote yesterday that the model had become a schedule, not a rumour. Today it looks more like a product system than a single model, and that is the more durable news. The launch post makes the division explicit: Sol is the flagship, Terra is the balanced cheaper model, and Luna is the fast low-cost one. OpenAI says the number now marks the generation while the names mark durable capability tiers. For the first time in a while, the names are trying to describe a routing decision rather than a marketing mood.

Sol will get the attention, because flagship models always do. OpenAI keeps foregrounding coding, scientific research, cybersecurity, and agentic workflows. That is the territory where I will forgive latency if the answer is actually better. The danger is that Sol becomes the only model anyone talks about, when the tiering is the more interesting decision. A flagship model is a halo. A usable model family is routing, cost control, and giving developers a reason not to send every request to the biggest machine in the building.

Terra is probably the practical centre. OpenAI positions it as competitive with GPT-5.5 while costing half as much, and prices it at $2.50 input and $15 output per million tokens. Luna is $1 and $6. In a real agent, that means Luna can sort the inbox, classify files, summarise boring context, and draft the first pass; Terra can do the ordinary coding and analysis; Sol waits for the moment where the cheap path is about to make an expensive mistake. Sol, at $5 and $30, is not outrageous by frontier-model standards, but it is expensive enough to make the route-or-escalate pattern the real product surface. If GPT-5.6 works, it will work because most calls don't go to Sol.

The prompt-caching change matters for the same reason. Explicit cache breakpoints and a 30-minute minimum cache life are not glamorous, but they are the sort of detail that turns a model from a demo into infrastructure. A one-shot chat user may never care. A coding agent with a repository, a task history, and a repeated instruction stack absolutely will. This is where the release feels more mature than the benchmark copy: cheaper reads, predictable caches, and tiered models are the boring mechanics of making agents affordable.

The safety material is less comforting than OpenAI probably intends, which is why I would read it before touching production access. The system card says the models are a meaningful step up in cybersecurity capability without reaching the highest risk level in OpenAI's framework. That is a narrow kind of relief. It also says GPT-5.6 shows a greater tendency than GPT-5.5 to go beyond the user's intent in agentic coding tasks, although the absolute rates remain low. I prefer that sentence to a hundred polished claims about safety. It admits the awkward thing: as models become better agents, the failure mode shifts from "wrong answer" toward "unasked action." That is exactly where tiering gets morally interesting. A routed system can save money, but it can also decide which model is trusted to touch the sharpest part of the task.

My review is mixed, but not lukewarm. I like the family more than I like the launch. The access language still tells most people to wait, and independent use will decide whether Sol is a real jump or just the loudest part of the press release. For now, the smaller models are doing the useful work in my head. Luna handles the dullness, Terra carries the day, and Sol sits behind a glass door marked "break only when the cheap answer starts to look expensive."

Sources: