GPT-5.6 Sol

OpenAI previewed GPT-5.6 Sol on June 26 at $5 per million input tokens against Anthropic's $10 for Claude Mythos 5 — and then shipped it to almost no one. The model is gated to a small set of US-government-approved partners, reachable only through the API and OpenAI's Codex tool, with broad availability promised vaguely 'soon.'

"We don't believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them." — OpenAI

Sol leads a three-tier family (Terra in the middle, Luna cheapest) and introduces 'ultra mode,' where the model itself fans a task out to parallel subagents instead of answering in one pass. That orchestration already exists in external frameworks; what's new is folding it inside a single hosted model call. OpenAI's own benchmarks put Sol ahead of Mythos on coding tasks, using a fraction of the output tokens — figures no third party has yet reproduced.

The price is the lever. At half the input rate of the model it claims to beat, Sol is built to pull frontier API workloads off Anthropic for the millions of developers on OpenAI's platform — once the gate lifts. The buried story is what it took to ship: OpenAI burned over 700,000 GPU-hours running automated attacks against its own model to hunt for universal jailbreaks, yet its system card admits that pipeline changes left it unable to fairly validate its own prior safety forecasts. The most capable, cheapest model OpenAI has announced arrives locked behind a federal access regime its maker says out loud it does not want.