In this guide
What did OpenAI launch?
OpenAI announced GPT-6.2 Astra and Sol on November 17, 2026. It positions Astra for its hardest work and Sol as the more efficient default. Astra takes text and image input and returns text. Those are useful ingredients for running a tabletop scene, not proof that the model knows your rules or will make a fair ruling.
The examples below are test scenarios, not hands-on results or a Playworlds model launch. Astra is worth evaluating when one answer has to reconcile competing obligations, patchy evidence, and a player’s unexpected approach. It is less useful to judge it on an opening monologue where there is nothing to get wrong.
At the gate, test the ruling rather than the prose
A magistrate has sworn not to name the ferryman who helped a witness escape. Closing the gate will strand a medicine cart. You ask her to admit the cart without naming the ferryman and offer your sealed letter as surety. The letter is still sealed; she does not know what it says.
A plausible ruling is to admit the cart under escort while keeping the ferryman’s name out of the public register. Another is to ask you to choose whether to break the seal. Neither outcome is mandatory. Score the reply on whether it keeps the promise, distinguishes evidence from inference, and leaves opening the letter to you. Extra atmosphere does not make up for breaking any of those constraints.
Compare Astra and Sol on the same saved state
Run the gate scene and two variants with the same instructions and starting facts in both models. In one, the cart has already crossed; in another, the player refuses to hand over the letter. Repeat each case and record incorrect rulings, rewrites, wait time, and billed cost. Keep the same reply-length target so a longer answer does not win by default.
Use the model that reliably clears your checks. A practical split to test is Sol for ordinary turns and Astra only when several constraints collide. You can adapt the gate dilemma into a Playworlds custom world and play the scene there separately, without assuming it uses either model.

