OpenAI has rolled out the GPT-5.6 series broadly, bundling it with significant product changes that reshape how developers and power users interact with ChatGPT and Codex.
Three Models, Six Thinking Levels
The GPT-5.6 lineup consists of three distinct models:
- Luna — a lighter, faster model. It handles clearly defined tasks well but can misfire on ambiguous prompts.
- Terra — positioned as an incremental upgrade over GPT-5.5, with modest improvements in UI generation and writing, and reportedly more steerable with explicit prompting.
- Sol — the most capable of the three, strong at UI work (especially with reference inputs), and notable for high-quality writing at Max thinking level.
Each model supports five named thinking levels: light, medium, high, xhigh, and max — plus a new Ultra mode that spins up subagents to tackle complex tasks in parallel. The tradeoff: higher thinking levels and Ultra mode chew through usage limits significantly faster. Ben Tossell (the newsletter's author) notes he came close to exhausting his Codex limits for the first time ever after running Ultra mode.
His current defaults: Sol at medium for building and creative work, background agents for harder tasks, and Luna at xhigh for day-to-day productivity.
Codex and ChatGPT macOS Are Now One App
The ChatGPT macOS app and Codex app have been merged into a single application. A new mode called ChatGPT Work sits alongside the Codex interface, with each mode tuned for its respective use case — coding vs. general productivity. Existing Codex users can update their app to get the combined experience.
During the rollout weekend, OpenAI reset usage limits four to five times while addressing bugs introduced by the app merge, and temporarily lifted the five-hour usage cap. That's a double-edged situation: more headroom in the short term, but easy to accidentally burn through your weekly allocation.
Computer Use Gets Real
One of the more compelling capabilities in the new Codex app is Computer Use — the model can self-drive your cursor, open applications, click buttons, and interact with on-screen content autonomously. Tossell recommends trying Sol at medium or high thinking for a small task as a starting point.
This puts GPT-5.6 more directly in competition with tools like Anthropic's Claude, which recently added an in-app browser to Claude Code for navigating docs and live applications. The Computer Use race is accelerating fast.
ChatGPT Sites: Build and Host From Chat
A new ChatGPT Sites plugin lets users generate and host web pages directly from within ChatGPT, with an optional "Login with ChatGPT" feature for gating content. It's a low-friction way to publish and share AI-generated content — though Tossell found it defaulted to publishing everything he built through it, which he quickly disabled.
For builders and marketers, this lowers the floor for spinning up quick landing pages or prototypes without touching a separate tool — though it's not a replacement for a proper SaaS website built with production-quality design and infrastructure.
What This Means for Builders
For startup founders and developers, the key takeaways:
- Model selection now matters — Sol, Terra, and Luna serve different use cases; defaulting to the most powerful setting will drain limits quickly.
- Computer Use is worth exploring — autonomous cursor control for small, well-defined tasks is now accessible without additional tooling.
- Usage limit management becomes a real workflow concern — especially if teams are running background agents or Ultra mode regularly.
- The Codex/ChatGPT merge simplifies the tool stack, but the combined app is still stabilizing post-launch.
With Anthropic extending Claude Code limits and Meta entering the API market with its Muse Spark 1.1 model (1M token context, competitive pricing), the competitive pressure on OpenAI to ship fast and iterate is clearly influencing the pace — and the roughness — of these rollouts.
