OpenAI's Sottiaux teases GPT-5.6 Sol Ultra for Codex users

A short post from OpenAI's Codex engineering lead is doing the rounds. Thibault Sottiaux told followers to "Stash your hardest prompts somewhere" and said he "Can't wait to see what people will do with GPT-5.6 Sol Ultra," flagging a tier of the recently-previewed Sol model that has not been formally announced. The Hacker News thread that pulled the tweet in is sitting on 155 points and 93 comments, which is where most of the interpretation is happening.

The interesting part isn't the branding, it is what Ultra is described as doing under the hood. Where Sol's existing Pro mode ran agents independently and only merged their work once they all finished, Ultra reportedly "goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work," with subagents trained to cooperate and allowed to communicate with each other during a task. Reported Terminal-Bench 2.1 numbers, and treat them as reported not settled, put Ultra at 91.9% against 88.8% for base Sol, edging Claude Mythos 5 at 88.0% and GPT-5.5 at 88.0%.

For anyone paying real money to run Codex, cost is the story to watch. Sol is already listed at $5 input and $30 output per million tokens, and one commenter in the thread with corporate access says internal guidance has flipped from encouraging token usage to asking people to conserve. Another commenter, postalcoder, points to The Information's reporting that OpenAI has found a way to cut inference costs by half, which is presumably part of how a compute-heavier tier can even land at a reasonable price.

The honest caveats are stacked. Sottiaux gave no launch date, no pricing, and no eval methodology; the 91.9% figure does not appear on OpenAI's own Sol preview page in what I could retrieve; and several thread participants point out you can already prompt Codex or Claude Code to spawn subagents today, so how much of Ultra is genuinely new architecture versus a behaviour trained into the same substrate is unresolved. What the reporting doesn't give you is the SKU, the quota structure, or when non-partner customers see it.

If it ships roughly as teased, the beneficiaries are the small circle of trusted partners already inside the GPT-5.6 preview, plus anyone whose Codex workflow keeps hitting a wall on the harder end of long-horizon tasks. The rest of us find out how much that class of work is actually worth to OpenAI.