A Hacker News thread exploding right now is about one thing: GPT-5.6 Sol Ultra is heading into Codex, and the community is anything but calm.

It started when Thibault Sottiaux — engineering lead for Codex at OpenAI — posted a tweet that essentially said: "stash your hardest prompts, you'll need them soon." That's the kind of cryptic teasing that sets the underground alight, and it did exactly that.

So what actually is Sol Ultra? It's an extended variant of the GPT-5.6 Sol family, and what sets it apart from everything else in OpenAI's lineup is the architecture: instead of a single agent working its way through a task, Sol Ultra coordinates multiple subagents that communicate and collaborate along the way. The idea is that complex coding tasks — the ones requiring planning, iteration, and coordination across dependencies — get solved faster and better when you have a small internal "team" of agents on the job.

On Terminal-Bench 2.1, a benchmark that tests command-line workflows with real-world complexity, Ultra is reported to sit at around 91.9%. That's above base Sol (88.8%) and above what is being reported for Claude Mythos 5. These numbers are still unofficial and come from community sources — take them with a grain of salt, but the direction is clear.

When OpenAI's own engineering lead says "stash your hardest prompts" — it's worth listening.
GPT-5.6 Sol Ultra is coming to Codex — and it smells like war - Bilde 1

What makes this particularly interesting is the context: Codex is already positioning itself as a serious alternative to GitHub Copilot for autonomous coding tasks. Sol Ultra entering Codex means OpenAI is going after the same segment that Microsoft/GitHub has dominated — and with a very different approach than autocomplete and chat.

Access is still limited to select partners via API, and the pricing of $5 in / $30 out per million tokens places this firmly in enterprise territory. This is not a model for hobbyists — it's a model for teams that are genuinely willing to pay for AI to tackle the really hard problems.

Worth noting: these are still early signals based on community leaks and a single tweet from an insider perspective. No complete technical documentation is out yet. But when the HN underground reacts this collectively, it's usually because something is actually happening.

Keep an eye on OpenAI's Codex page over the coming days.