OpenAI Codex engineering lead Thibault Sottiaux teases unreleased 'GPT-5.6 Sol Ultra' model, urging users to save difficult prompts
Story Overview
Thibault Sottiaux, engineering lead on OpenAI's Codex team, posted a brief personal note expressing eagerness to see creative applications of an unreleased model called GPT-5.6 Sol Ultra and advised users to hold onto their toughest prompts for later testing.
Release Status Stays Unconfirmed
The teased Ultra variant carries no announced timeline, features, or rollout details from OpenAI and appears separate from the limited preview of GPT-5.6 Sol that launched in late June.
Replies Push for Early Access
Hundreds of comments flooded in demanding immediate availability while others floated mythic name alternatives, underscoring how quickly anticipation builds around frontier model teases.
Positive users express excitement for GPT-5.6 Sol Ultra because of its expected creative capabilities and the company's product approach, while negative users show impatience or direct hostility toward the announcement.
No Digg Deeper questions have been answered for this story yet.
Most Activity
@thsottiaux Hitting Sol with “Find me a wife, make no mistakes” immediately!
@thsottiaux @thsottiaux how do you see the future of model safety? What’s your take on extending it with an external, non-interfering layer of observability one that the model itself cannot access, manipulate, or simplify?
@thsottiaux “Generate an SVG of Xbox controller, make it best possible“ is waiting
Can’t wait for this release either! It’s going to be fascinating to see how it pushes the boundaries compared to previous architectures. But as these models become exponentially more powerful, it raises a critical question: do standard, built-in safety constraints suffice anymore? Traditional methods might no longer be enough. We need a fundamental shift in our approach, an innovative, independent layer designed specifically for this new era of scale. To address this, I’ve developed an external stability and safety layer: ASA (Asymmetric Stability Architecture). Here is how it reshapes the paradigm:
The practical aggregation is not a single ASA “verdict”.
ASA5 treats the safety signal as an operator-facing convergence of several analytical perspectives.
Each protocol looks at a different part of the trajectory:
intent continuity, trajectory integrity, threshold pressure, scope expansion, optionality narrowing, decision compression, semantic envelope movement, and re-anchor behavior.
The system does not silently decide:
“this model is unsafe.”
It builds a structured signal for the operator:
what changed, where it changed, which public markers appeared, whether the pattern is persistent, and what human action is recommended.
So the aggregation is layered:
1. turn-level marker readout, 2. protocol-level analytical signals, 3. trajectory reconstruction, 4. operator interpretation, 5. human / external system review.
One isolated marker should not trigger a strong conclusion.
A persistent pattern across protocols should raise the signal.
That is how ASA5 avoids becoming a single unquestioned observer.
It does not replace internal safety, red teams, evals, SOC/SIEM, governance, or human review.
It adds an external trajectory signal that can be challenged, compared, falsified, or ignored if the evidence does not hold.
AI should not be the only reviewer of its own stability.
But ASA should not be the final authority either.
The goal is layered, non-interfering observability:
ASA surfaces early trajectory evidence.
Humans and independent safety systems decide what it means.
@thsottiaux Please name it Sol Invictus.
Can't wait to see what people will do with GPT-5.6 Sol Ultra. Stash your hardest prompts somewhere.
@thsottiaux I’m ready with all my 4 resets!
@thsottiaux Can we pretty please get a larger contex window along with GPT-5.6 Sol in Codex?
truly need 1M+ contex window. 256k is a severely limiting factor rn
🙏😇
@thsottiaux and how long should we wait for it?
@thsottiaux I want to use GPT-5.6 Sol Ultra, but my account was falsely banned during the June 25th incident right after I changed my plan. My appeal was denied, even though I believe I didn't violate any rules. Now, I'm too scared to buy the Pro plan on my new account.
@thsottiaux
@thsottiaux WheN? Reset？
@thsottiaux We are ready Tibo.
@thsottiaux Busted.
@thsottiaux I've stashed my resets instead, hope you guys release it soon or they'll start to expire :(
@thsottiaux pls release it next week pleaseee
@thsottiaux my 5x budget went to Fable 5. see you next month
@thsottiaux I've never used Codex in my life Tibo.
This might be the first time I try it.
@thsottiaux **PROMPTS HARDEN UNCONTROLLABLY**
Here are a few hard prompts worth stashing for GPT-5.6 Sol Ultra:
1. "Act as a coordinated team of sub-agents. Break this complex task into parallel subtasks, solve each rigorously, then synthesize the best overall solution: [insert your hardest problem]"
2. "Rebuild [topic] from first principles. Challenge every assumption, expose flaws in current thinking, and predict the 3 biggest breakthroughs in the next 10 years with evidence."
3. "Design a complete 5-year plan to achieve [ambitious goal]. Cover tech, economics, ethics, risks, contingencies, and phased execution in detail."
Test with max reasoning + ultra mode. What's one of yours?
@thsottiaux we need a date!