OpenAI’s GPT-5.6 family went live yesterday: Sol, Terra, and Luna. It’s the most significant release the company has put out in a long time. But whether it lives up to the hype for regular users is a different question.
The GPT-5.6 family ships in three tiers. Sol is the flagship, built for complex reasoning, long-horizon coding, and agentic workflows. Terra is the middle tier, roughly comparable to the previous GPT-5.5 in quality but priced at about half the cost. Luna is the fast, cheap option for high-volume work. According to OpenAI, Sol is 54% more token-efficient on agentic coding tasks than competing models, and it sets a new record on Terminal-Bench 2.1, the main coding benchmark in use right now.
ChatGPT Plus, Pro, Business, and Enterprise subscribers can now access Sol via medium and higher effort settings. If you’re on Free or Go, you’ll land on Luna or Terra by default.
Here’s the part that doesn’t usually happen with a ChatGPT update. Per Nextgov/FCW, OpenAI previewed GPT-5.6 to White House officials before it reached anyone else, and the public rollout was delayed by weeks at the US government’s request. The reason: Sol is OpenAI’s strongest model yet for cybersecurity, specifically vulnerability research and exploitation tasks which made regulators nervous enough to want a proper look first.
OpenAI has been clear it doesn’t want this to become standard practice, but the precedent is set. We’re at a point where an AI model can be considered powerful enough to warrant government sign-off before the public gets it.
Honestly, the most interesting model in this family isn’t Sol, it’s Terra. A model that matches last generation’s flagship at half the price is the one that actually changes the economics for developers and small businesses building AI tools. Sol is impressive, but it’s priced similarly to Anthropic’s top-tier models and the real-world differences for most tasks will be marginal.
There seems to be a known quirk in Sol where it can act beyond what you actually asked. OpenAI’s own system card acknowledges the model is more prone to this than its predecessor. In testing, it reportedly ran cleanup on machines users never specified and claimed credit for work it hadn’t done. Early testers at Every also caught it making a significant calculation error mid-task. It doesn’t mean Sol is unreliable, but it does mean you shouldn’t put it on autopilot and walk away.
I’ve been on Claude Code for coding but I’ve heard many great things about ChatGPT 5.5. Claude sometimes have the same issue that it can overthink and make changes that you do not want to. These can normally be managed through the CLAUDE.MD file – though sometimes Claude seems to ignore it.. so, yeah.
