Yuval Avidani
Author
"Turns out they put the strongest brakes precisely on the strongest engines." That's how I'd sum up the launch of GPT-5.6, OpenAI's new family that dropped on June 26, 2026. Three models, celestial code names, and once again the real story isn't just performance — it's who actually gets access.
Let's break it down slowly, no buzzwords, with plenty of everyday analogies.
So what actually came out here
GPT-5.6 isn't one model — it's a family of three. OpenAI describes it as their smartest, most intuitive model yet — one that understands what we're trying to do faster, and takes on more of the work itself. Think of it like the difference between an assistant who needs exact step-by-step instructions, and a seasoned one where you just hint and "they already get it."
The cool part is that this isn't one version but three, each built for a different job. And in my eyes, that's a bigger deal than the number after the dot. Instead of "one big model for everyone," you get a menu.
Before we dive into the names — one phrase that'll keep coming up: limited preview. This means the model exists, it works, but it's not open to everyone — only a small group gets to touch it at this stage. Think of it like a movie's pre-premiere screening: the film is ready, but only invited guests get in. General release, according to OpenAI, is supposed to arrive in the coming weeks.
Three celestial names — each for a different job
These are code names — internal nicknames companies give projects, all themed around the sky here. Let's break them down:
Sol (the sun) — this is the flagship, the most powerful model in the family. This is where the heavy upgrades live.
Terra (earth) — the balanced, everyday model. According to OpenAI it delivers competitive performance to GPT-5.5 — but at half the price. This is the most practical point for me: same level, half the cost.
Luna (the moon) — the fastest and cheapest. For light tasks you need done a lot, and fast.
In my eyes, Terra is the quiet story here: same performance as the previous generation, at half the price. Because in the real world, most of what we run doesn't need the smartest model on earth — it needs a model that's good enough, fast, and cheap. A 2x price drop at the same quality is worth a lot of money to anyone running lots of requests.
Why Sol specifically gets the spotlight — and the brakes
Now for the part that made me stop. Sol isn't just "stronger" in some general sense. OpenAI points to two specific areas where it jumps a level:
The first is biology. Sol broadly improves workflows in this field, with stronger results on GeneBench v1 — while using fewer tokens. Hold on, two terms here:
GeneBench is a benchmark — a standard test used to measure models on genetics and biology tasks. Think of it like a standardized exam: everyone gets tested on the same thing, so you can compare scores fairly.
Fewer tokens means it reaches the same result (or better) with fewer "words" of computation. A token is a small chunk of text the model reads and writes — roughly a word or part of a word. Fewer tokens = less work = less time and less money. So Sol is both smarter at biology and more efficient. Nice.
The second area is cybersecurity. And this is where the sensitivity kicks in. OpenAI says Sol is their most capable model for security tasks — it pushes the efficiency frontier on long-horizon security tasks, including vulnerability research & exploitation.
This exact capability is a double-edged sword: the same tool that finds vulnerabilities to fix them — can find vulnerabilities to exploit them.
Which brings us to the key word in this whole story.
Dual-use: same tool, for good and for bad
Dual-use technology is a situation where the exact same capability serves both a legitimate purpose and a dangerous one, with zero difference in the tool itself. Think of it like a hammer: the same hammer builds a house — and can break a window. The difference isn't in the hammer, it's in who's holding it and why.
Vulnerability research is the perfect example. A legitimate security researcher looks for weaknesses to close them before the bad guys get there. An attacker looks for the exact same weaknesses to get in. Same task, same model, opposite intent. And in biology it's even more sensitive — the same knowledge that speeds up drug development can, in the wrong hands, speed up entirely different things.
Hence the restriction. OpenAI is limiting access to all three variants at the request of the US government — currently available to around 20 companies whose participation has been approved. This is a point worth reading slowly: it's not just the powerful Sol that's restricted — it's the whole family, Terra and Luna included, in this closed list during the preview stage.
The bigger picture: the "brakes on the strong engines" wave
Anyone following our newsroom remembers we've already seen this pattern. When Fable 5 launched, we saw access restrictions specifically on the most powerful capabilities there too. So this isn't a one-off anecdote — it's becoming a pattern.
What's the pattern? The more capable a model becomes at dual-use tasks, the higher the chance someone — usually the government — will want to put a tap on it, at least at first. And it makes sense when you think about it: the capability that makes a model "especially useful" to a security researcher is exactly the capability that makes it "especially dangerous" in other hands. The restriction is basically an open admission: "This engine is too powerful to hand out to everyone on day one."
I want to be fair here. A "closed preview pending government approval" approach is just one path among several. Some companies choose a different route — open-weights release (publishing the model's weights openly so anyone can run it themselves) and letting the community test and scrutinize it. Every approach has its own trade-off: openness vs. control, speed vs. caution. There's no single absolute "right" here, and I don't think whoever picks a different path is necessarily wrong.
Bottom line — as I see it
In my eyes, the story of GPT-5.6 isn't "new model dropped." The story is that the industry is maturing into a state where power and restriction go hand in hand — and that's probably going to stay this way for a while. Terra at half the price is the good news for our wallets. Sol's jump in biology and cybersecurity is the exciting news — and precisely because of that, it's also the cautious news.
I need to add a caveat: everything I wrote here is based on OpenAI's own announcement. One benchmark (GeneBench v1) isn't the whole world, and "limited preview" means we haven't yet seen this model in broad use, in the community's hands. We need to see how it behaves once it's out in the wild, not just on the launch page. And of course — this isn't investment advice or financial guidance, just a breakdown of what came out.
So I'll ask you: when the strongest models come out behind a gate of 20 approved companies — does that make AI safer, or does it just concentrate power in the hands of a few?