How to use GPT-5.6
Ben's Bites ● Covered by 5 sources
OpenAI shipped GPT-5.6 to everyone, merging the ChatGPT and Codex apps into one, with a new Work mode. Three fresh models plus a wild Ultra setting can chew through your usage limits fast.
Based on reporting by Ben's Bites — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
OpenAI didn't just tweak the numbers this time. GPT-5.6 lands as three separate models — Luna, Terra and Sol — each running at five thinking levels: light, medium, high, xhigh and max. On top of that sits a new Ultra mode, which lets the models spin up subagents and, by most accounts, go a bit haywire. The catch is that the higher you crank the thinking level, the faster your usage limit disappears. One tester said Ultra mode was the first time they'd nearly run out of usage inside Codex entirely.
The app side got a shake-up too. ChatGPT's macOS app and the Codex app have merged into a single app, with a new mode called ChatGPT Work sitting alongside Codex. They look nearly identical, just tuned differently depending on whether you're coding or doing something else. There's also a new plugin, ChatGPT Sites, for spinning up hosted websites with an optional Login with ChatGPT feature — handy for sharing work, though one early user found it kept routing every query through the site and switched it off.
In practice, the three models seem to have distinct personalities. Sol handles UI work well, and noticeably better when given reference material to work from; at max thinking it also writes well and holds a decent conversation. Terra reads like a modest step up from GPT-5.5, sharper on UI and writing, and more steerable, which makes it a good candidate for skill-based workflows. Luna, meanwhile, has what one user called a mini-model smell — it can misread ambiguous prompts, though it rarely fails once a task is spelled out clearly. The recommended setup floating around is Sol medium for building and creative work, background agents for the tougher jobs, and Luna xhigh for everyday productivity tasks.
The rollout hasn't been entirely smooth. OpenAI reset usage limits four or five times over the weekend while patching bugs introduced by the app merger, and has temporarily lifted the five-hour usage cap. That sounds generous until you realize it also means someone could burn through an entire week's allowance in one sitting without the usual guardrail kicking in.
The more interesting demo might be Computer Use inside the Codex app, where these models can drive your cursor themselves — opening apps, clicking buttons, navigating by literally looking at the screen. Sol at medium or high thinking is apparently worth trying on a small task just to watch it work.
My take — AI-written commentary, not fact-checked reporting
Removing the five-hour usage cap while resetting limits four or five times in a weekend is not exactly a confidence-inspiring combo, and anyone who fires up Ultra mode without reading the fine print deserves the usage-limit heart attack that follows. The more telling signal here is that OpenAI shipped three differently-tuned models instead of one, which suggests even they don't think a single model can be good at UI, writing and ambiguous instruction-following all at once. That's a quiet admission worth paying more attention to than the flashy Ultra mode headline.
Read more about this at: Ben's Bites