DeskTheory is where founder-CEOs learn to run their companies on AI leverage.

What is GPT-5.6 Sol?

GPT-5.6 Sol is OpenAI's flagship frontier model, the top tier of the GPT-5.6 family alongside the cheaper Terra and the faster Luna. It shipped on July 9, 2026 after the first government-gated rollout of a frontier model, and in August it got 14x faster and up to a third cheaper.

Most model launches are a version number and a benchmark chart. Sol's two months on the market have been stranger and more instructive: it needed US government clearance to ship in June, and by late August OpenAI was discounting it. That arc, frontier restriction to commodity pricing in nine weeks, is the actual lesson of this model.

What it is (in plain English)

GPT-5.6 is OpenAI's current model family, and it comes in three sizes named like planets: Sol, the flagship for the hardest work; Terra, the capable mid-priced option; and Luna, the fastest and cheapest. When someone says "GPT Sol," they mean the top one.

Sol's release was a first. OpenAI previewed it on June 26 to a small group of cleared partners after running the release through the US government, then opened it up on July 9. A frontier model shipping through a government gate had never happened before; it tells you where regulators now put the top end of AI capability. In ChatGPT, the 5.6 family rolled in with Luna as the default on the free tiers, so the model your team gets for free and the model you'd pay to point at real work are not the same thing.

Then came August, when Sol stopped being precious:

• August 13: OpenAI previewed Ultrafast, a service tier that runs Sol at up to 750 tokens per second, roughly 14x standard speed, on hardware from the chipmaker Cerebras. It's a limited preview, expanding "as capacity grows."
• August 21: OpenAI cut Sol's price: input from $5 to $4 per million tokens (20%) and output from $30 to $20 (33%), through at least November 21, across the API, Codex credits, and eligible ChatGPT Work plans.

Why you should care as a CEO

Speed and price are the two dials that decide what a frontier model is actually for. Slow and expensive means "special occasions": the board memo, the gnarly analysis. Fast and cheap means "everything, always." OpenAI turned both dials in the same month, so every build-vs-buy call, vendor quote, and "AI is too expensive for that workflow" verdict priced before August is now stale.

Speed is not a spec here, it is a feel. At 750 tokens per second the answer lands before you've finished reading your own question, and the tool stops feeling like a slow oracle and starts feeling like a fast colleague. We watched this coming when OpenAI first put Sol on Cerebras silicon in June; Ultrafast is that bet reaching the product.

There's also a vocabulary payoff. When your team says "we use GPT," you now know the follow-up question: which one? Luna answering your employees' free-tier questions and Sol running your priced inference workloads are different animals, and vendors quoting you "GPT-5.6 powered" pricing know it too.

Where you'll see it

• In our news desk coverage: the June Cerebras preview and August's speed and price moves.
• In ChatGPT, where the 5.6 family is live and Luna covers the free tiers.
• In vendor pitches and API invoices, where the August cut should show up in your pass-through costs.
• In the Frontier Watch hub, alongside what is a frontier model? and our standing ChatGPT vs Claude vs Gemini comparison.

What you should do next

Rerun the cost math on anything you shelved this summer for being too slow or too expensive, and if a vendor passes AI costs through to you, ask when the Sol cut shows up in your bill.