 |
| GPT-5.6 Features: Sol, Terra & Luna Explained |
The Dawn of GPT-5.6: How Sol, Terra, and Luna Are Changing What AI Can Actually Do
For the past few years, "using AI" has mostly meant typing something into a chat box and reading what comes back. GPT-5.6 — OpenAI's new three-model lineup made up of Sol, Terra, and Luna — is the clearest sign yet that this is changing. These models aren't just answering questions anymore. They're finishing jobs: wiring hardware, building out business systems, and chasing down math problems that have sat unsolved for years.
What's striking isn't any single trick the models can do. It's that they follow through. Older models were good at giving you a starting point — a list of steps, a rough plan, a first draft. GPT-5.6 tends to stay with a task until it's actually done, which turns out to matter a lot more than it sounds like on paper.
Here's what that looks like in practice, through three very different stories.
A Farmer in Hokkaido Automates His Greenhouse
Heroki runs a farm in Hokkaido, Japan, and like a lot of farmers, he was stuck doing something by hand that had no business still being manual: opening and closing his greenhouse doors, day after day, regardless of the weather or his schedule.
He turned to GPT-5.6 to fix it — not by asking for general advice, but by working through the whole problem with the model as a kind of on-call engineer. It told him which parts he actually needed to buy. It walked him through wiring a Raspberry Pi, which is not a trivial thing to do if you've never touched one before. And it talked him through mounting a motor that could physically swing the doors open and shut.
None of that is "AI advice" in the usual sense. It's a machine helping someone build a working piece of hardware, one wire and one bracket at a time. That's the part worth sitting with: this didn't stay on a screen. A farmer with no electronics background ended up with an automated greenhouse, and the AI was the one who knew how to get him there.
Turning a Five-Minute Brain Dump Into a Real Business Plan
Not every use case involves a soldering iron. Jake, who works with the cereal brand Three Wishes, uses GPT-5.6 for something almost the opposite: taking the messiest, most scattered thinking he can produce and turning it into something a business can actually run on.
He describes it as handing over his most disorganized thought streams and getting back something polished on the other side. A five-minute ramble becomes a structured plan. Loose ideas get paired with the company's actual historical data and turned into dashboards that reflect what's really happening in the business. Presentations and spreadsheets come out branded correctly, using the company's real assets and current packaging, instead of generic placeholder templates.
This is the kind of thing that used to require a small internal team — someone to organize the thinking, someone to build the deck, someone to make sure it matched the brand. For a company like Three Wishes, that team is now largely one person and a model that's willing to work through a task end to end instead of stopping halfway.
Disproving a Conjecture That Resisted Three Years of Work
Then there's Bartosh, a mathematician in Poland who'd spent three years trying to prove a particular conjecture and getting nowhere. When he switched to Codex 5.6, the outcome flipped entirely — he ended up disproving it.
What made the difference wasn't raw speed. It was how the model broke the problem apart. Instead of grinding through one long chain of reasoning, Codex 5.6 split the work into parallel streams and ran multiple agents on different pieces of the problem at once. Somewhere in that process, it didn't just check Bartosh's existing approach — it produced a genuinely new idea, one he hadn't tried in three years of working the problem himself.
That's a meaningfully different kind of help than "faster calculation." It's closer to having a research partner who approaches a stuck problem from an angle you hadn't considered.
Why This Particular Release Feels Different
Across all three stories, the common thread isn't intelligence in the abstract — it's persistence. A farmer needed someone to see a hardware project through to a working motor. A brand needed someone to turn scattered notes into something usable, without a dozen back-and-forth check-ins. A mathematician needed something that wouldn't just retread the same reasoning he'd already exhausted.
GPT-5.6's Sol, Terra, and Luna tiers give people different sizes of that same capability — Sol for the heaviest, most demanding work, Terra as a solid everyday option, and Luna for quick, cheap tasks that don't need much depth. But the underlying shift is the same across all three: these models are built to stay on a problem until it's actually resolved, whether that resolution is a working piece of farm equipment, a business dashboard, or a disproven conjecture.
That's a small phrase — "sees it through" — but it's doing a lot of work here. It's the difference between a tool you have to supervise closely and one you can actually hand a problem to.
0 Comments