For most of the last month, the most interesting thing about OpenAI's GPT-5.6 was that almost nobody could use it. The models existed, they had names, and they were locked behind a US government preview. On Wednesday, OpenAI said that ends this Thursday.
What OpenAI announced
In a short post on X, OpenAI said its GPT-5.6 family, Sol, Terra, and Luna, will launch publicly this Thursday, and that preview access is expanding globally right now. Sam Altman added his own four-word version: "GPT-5.6 sol launches thursday!" The company's own post has been the clearest sourcing so far, echoed by outlets like Cointelegraph and Crypto Briefing, which pegged the public date to July 9.
Why the gate mattered
The release matters partly because of how restricted it was. GPT-5.6's preview began in late June, limited to a small set of US government-approved organizations that could reach it through developer tools. That was the same federal sign-off that pulled Anthropic's Fable 5 and Mythos 5 offline for weeks earlier in the summer. For a stretch, two of the biggest US labs both needed government clearance before a new model could ship, and there was open speculation that GPT-5.6's wider release might end up US-only. A public, worldwide rollout is the clearest sign yet that this particular gate is lifting.
Three models, three jobs
GPT-5.6 isn't one model. It ships as three, each aimed at a different kind of work:
- Sol is the flagship, built for hard, multi-step tasks that run for a while. Early chatter also points to a "Sol Ultra" mode that runs several sub-agents in parallel.
- Terra is the everyday model, positioned at roughly half the cost of GPT-5.5 for similar performance.
- Luna is the fast, cheap option for high-volume, routine work.
That structure maps closely onto Anthropic's Opus, Sonnet, and Haiku tiers, and the pitch underneath it is blunt: comparable quality to the top Claude models, for less money.
The reviews don't agree yet
Almost everything anyone has said about how good GPT-5.6 actually is comes from OpenAI itself or a handful of early testers, and they don't agree. Some testers say Sol is roughly as capable as the best Claude models at a lower price. At least one trusted tester said it feels weaker than Fable 5. An NVIDIA engineer's early account claimed Sol finished in 30 hours what Opus 4.8 took 64 to do, while also noting it "constantly tries to implement complex things and makes mistakes." Independent, public benchmarking is thin, because until Thursday most people can't run it.
Finally, Fable 5 has some competition, and that's exciting for a few reasons.
Here's how I plan on using it. My bet is GPT-5.6 edges out Opus 4.8 on architecture and heavy builds, so I'll run Fable 5 as the orchestrator and fan those jobs out to GPT-5.6. I'll keep Opus on design work, because I'd wager it still has more taste than GPT-5.6. TBD on that.
I've got a video coming that stress-tests the whole setup, so keep an eye out.
Really exciting times. It feels like we're crossing another chasm in raw AI intelligence, and hopefully it pushes Anthropic to make Fable 5 more widely available too.