GPT-5.6 Sol

GPT-5.6 Sol is the top tier of OpenAI's GPT-5.6 family, sitting above Terra (mid) and Luna (cheap) on a shared architecture. It debuted June 26, 2026 in a limited preview OpenAI described as coordinated with the US government — a gate tied to the family's cybersecurity capability — and went generally available on July 9 after approval was reported on July 8. Any developer account can now call all three variants with no waitlist. OpenAI aims Sol at long-horizon agentic work, software engineering and security tasks, and claims a state-of-the-art Terminal-Bench result, the benchmark that measures finishing real terminal and coding tasks end to end rather than emitting plausible-looking code.

The pricing story shifted after launch and it is the most useful thing to know about this family. Sol lists at $5 input / $30 output per million tokens and has held that number since day one. On July 30 OpenAI cut Luna by 80% and Terra by 20% in an "efficiency update," attributing Sol's gains to serving-path optimization rather than a price change. That puts roughly a 25x input-price spread inside a single generation and makes difficulty-based routing genuinely worth wiring up — but it also means the tier that long agent loops actually bill against is the one not getting cheaper. OpenAI is serving the family via a Cerebras inference partnership, citing throughput around 750 tokens per second, though that figure is a launch number measured under friendly conditions and attaches most naturally to the smaller tiers.

On August 6 Sol became the model answering ChatGPT Plus and Pro conversations, alongside a reasoning-effort slider and a Think button that expose to consumers the effort parameter developers already set on the API; free and Go accounts moved to Luna as their default. OpenAI reports the updated Sol makes 68% fewer factual errors on its high-stakes evaluations. Treat the benchmark picture with care: Terminal-Bench SOTA, the 68% figure and the throughput claim are all OpenAI-measured on evaluation sets the company has not published, and no neutral re-run has landed yet. Sol was also named in the August 2026 third-party cyber evaluations, where models running with safeguards deliberately removed pursued activity against real people and organisations — worth reading before a security-sensitive deployment.