GPT-5.6 Goes Public: Sol, Terra and Luna Change the Price Math

OpenAI is opening the doors on GPT-5.6 today, July 9, 2026, and it is doing so in a way that tells you a lot about where the model market is heading. Since June 26 the model had been sitting in a government-coordinated limited preview, open to only about twenty vetted partner organizations, and notably absent from ChatGPT itself. That kind of staged, supervised rollout used to be the exception. Now it is starting to look like the template, and today’s global preview expansion is the moment the rest of us finally get to touch the thing. What we get is not one model but three: Sol, Terra and Luna, a deliberate tiering that I think is the real story here, more than any single capability claim.

Let me lay out the numbers first, because the numbers are doing most of the talking. On the API, Sol costs $5 per million input tokens and $30 per million output tokens. Terra comes in at $2.50 and $15, and Luna at $1 and $6. Cached-input reads are roughly 90 percent cheaper, and cache writes are billed at 1.25 times the normal input rate, which matters more than it sounds if you are running agents that re-read the same long context over and over. For anyone building on these models daily, the caching arithmetic can quietly dominate the headline rates.

What strikes me most is how directly this is aimed at Anthropic. Sol, the strongest of the three and the one OpenAI is pitching hardest at coding work, lands at roughly half the cost of Anthropic’s Fable 5, which runs about $10 for input and $50 for output per million tokens. That is not a subtle undercut. That is a price posted on a billboard across the street. Whether Sol actually matches Fable 5 on the work that matters to you is a separate question, and I will get to my doubts in a moment, but the pressure this puts on Anthropic’s pricing is real regardless of how the quality comparison shakes out. When one vendor halves the price of the flagship tier, everyone else has to either justify their premium with visibly better output or start sharpening their own pencils.

The tiering is the point, and so is the benchmark problem

I want to be fair about what OpenAI is trying to do with the three-way split. The framing is that Sol is the strongest, Terra the sensible middle, and Luna the cheapest and fastest, and that you should match the model to the task instead of defaulting to the biggest one because it feels safer. This is genuinely good product thinking. Most of what people send to frontier models does not need a frontier model. Summarizing a meeting transcript, classifying support tickets, drafting a routine email, none of that needs the full weight of Sol, and paying Sol prices for Luna work is how API bills quietly triple. The industry has been saying this for two years, but pricing structures kept nudging people toward the top shelf anyway. A clean three-tier menu with a five-to-one price spread between top and bottom makes the right behavior the cheap behavior, and I respect that.

But I also want to be honest about the part that makes me squint. TechTimes, in one of the first hands-on reviews out of the limited preview period, praised Sol’s speed and its price and then flagged what they called a benchmark problem: headline numbers that are hard to reproduce, or that do not translate cleanly into real work when you sit down and actually use the model on your own codebase. This has become a depressingly familiar pattern across the industry, and I do not think OpenAI is uniquely guilty of it, but it deserves to be named. A model that scores three points higher on a coding benchmark and then fumbles the unglamorous parts of your actual repository, the weird build system, the legacy naming conventions, the test suite nobody loves, has not actually saved you anything. Benchmark theater is cheap to produce and expensive to believe. The only benchmark that matters is your own workload, run on your own data, ideally in the first week while your evaluation instincts are still sharp and before organizational inertia locks in a default.

So what would I actually do, starting today? If you write code for a living and speed of iteration is your bottleneck, Sol is the one to trial first, because that is where OpenAI has aimed it and where the price cut against Fable 5 bites hardest. Give it a real task from your backlog, not a toy prompt, and see whether the speed holds up when the problem gets ugly. Terra is where I would put the broad middle of knowledge work, drafting, analysis, longer reasoning over documents, anything where you care about quality but a wrong first draft costs you a minute rather than a production incident. Luna is for volume: classification, extraction, routing, summarization at scale, the plumbing tasks where a 20 percent quality gap is invisible but a five-fold price gap is not. And whichever tier you pick, structure your prompts to exploit that 90 percent caching discount, because at these prices the difference between a naive integration and a cache-aware one is the difference between a rounding error and a line item your finance team asks about.

The bigger picture is that model choice is turning into an actual decision rather than a default, and today’s launch accelerates that. The spec sheet says Sol is half the price of Fable 5. Your job, and mine, is to find out what that half-price buys on real work, with a healthy allergy to leaderboard numbers along the way.

Sources: This article draws on OpenAI’s announcement of the GPT-5.6 public launch and its published API pricing for Sol, Terra and Luna, the TechTimes hands-on review that raised the benchmark reproducibility concerns, and current AI pricing guides for the comparison with Anthropic’s Fable 5 rates.