GPT-5.6: A Very Serious Space Diner

Last Tuesday, my coffee maker beeped like a tiny robot judge, and my laptop refused to open a file named finaldraftfinaldraftFINAL. It was the kind of morning where even the toaster felt superior. Then I learned OpenAI was starting a limited preview of GPT-5.6, with three new flavors: Sol, Terra, and Luna. It sounded like a menu at a very serious space diner.
The lineup is cheeky but practical. Sol is the flagship, Terra is the all-rounder, and Luna is the quick, budget-friendly one. Terra promises GPT-5.5-ish performance at about half the price, while Luna tries to bring strong capability to the lowest cost. In other words, we finally got a model family that speaks the language of real bills: faster, cheaper, and not just one giant mystery box.
I have that relatable moment when I ask a coding assistant to fix one small bug, and it returns a novel, a new dependency, and a mild identity crisis. So I was intrigued by GPT-5.6 Sol’s max reasoning effort and ultra mode. Imagine giving the model extra time to think, then letting a team of subagents swarm the problem like very polite termites. For coding, Sol sets a new state of the art on Terminal-Bench 2.1, which means it handles command-line planning, iteration, and tool coordination with fewer eye rolls.
Beyond code, Sol shows improvement in biology and cybersecurity. On GeneBench v1, it does stronger long-horizon genomics work while using fewer tokens, which is like finishing a marathon while eating less. On cyber benchmarks, the models get better as reasoning increases. Sol is competitive on ExploitBench² using about one third of the output tokens compared with a preview model, and all three shine on ExploitGym. That is the kind of benchmark list that makes a Tuesday feel suspiciously fun.
Safety remains the unsexy hero. OpenAI says Sol launches with its strongest safeguards yet, using model-level refusals, real-time classifiers, account-level review, monitoring, and phased access. It is like installing a bouncer, a camera, and a very thorough clipboard at the club. The goal is to make harmful cyber assistance harder while still letting defenders debug, patch, teach, and test. During preview, users may hit blocks or delays, especially in dual-use areas where good security work can look a little too dramatic.
The release starts small because of a temporary government preview process, with trusted partners first and wider access planned soon. OpenAI says this is not a permanent default, but a step toward broader availability while cyber rules take shape. Pricing is straightforward: Sol at $5/$30 per million input/output tokens, Terra at $2.50/$15, and Luna at $1/$6. Later, Sol will run on Cerebras at up to 750 tokens per second, which is fast enough to make your old hardware send a polite letter of complaint. If the preview goes well, Sol, Terra, and Luna will soon be less like a secret tasting and more like the main course. Maybe your next draft will finally compile itself. It could be delicious. Very soon.