OpenAI's next Codex update breaks hard coding jobs across multiple AI models
OpenAI's new 'ultra mode' runs multiple AI models in parallel for hard coding work, and the US government is among the first testers.
OpenAI's new 'ultra mode' runs multiple AI models in parallel for hard coding work, and the US government is among the first testers.
OpenAI's next version of its coding assistant will not run a single model. It will run several in parallel, an architectural shift the company is making as it previews its GPT-5.6 family and readies it for general release.
The shift arrives inside a new "ultra mode" that OpenAI introduced in its GPT-5.6 preview blog on Friday. The mode, the company wrote, "goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work." An OpenAI product description names subagent orchestration, not a larger model, as the source of the capability jump.
What that means in practice: rather than asking one large language model to plan, write, and debug a long task, the mode breaks the job into pieces and dispatches them to focused subagents that report back. Each subagent handles a slice; the main model or an orchestrator stitches the result. The technique has been a research staple in the field for a year, and competitors have shown pieces of it in their own products, but putting it inside a shipped developer tool is a different kind of commitment. It implies OpenAI believes the coordination overhead is worth the speed and quality gain for hard coding work, and that the orchestration software has caught up to the model side.
That choice is now official for Codex, OpenAI's AI coding assistant. Thibault Sottiaux, the company's engineering lead for Codex, confirmed on X that "Ultra will be in codex." The tweet landed the same day the preview went live and is the clearest public signal that subagent orchestration is not just a blog claim but a product surface in development. Independent writeups of the preview, including an in-depth review at eesel, note that the mode is gated to trusted testers in the partner program and that no public Codex interface yet exposes it.
OpenAI is gating GPT-5.6, which includes the flagship Sol tier, a balanced tier called Terra, and a fast cheap tier called Luna, behind a small set of trusted partners whose usage is shared with the US government. The framing is part of a broader engagement around a federal cyber executive order, an unusual structural detail for a model launch. It signals that OpenAI is positioning the family as load-bearing infrastructure for sensitive workflows, and that federal trust is a precondition for early access rather than a marketing afterthought. The partner-only window also keeps the model out of the public eval scene until OpenAI decides otherwise, which means the first independent benchmarks will not appear until either the public preview or general availability.
That positioning raises an open product question the company has not answered. Discussion on Hacker News frames Ultra as distinct from the existing Pro tier, leaving developers to guess whether Ultra replaces, layers on top of, or competes with what they already pay for. The preview blog does not mention pricing for Ultra specifically, and OpenAI has not released a system card section that describes its benchmarks, latency, or cost profile. The model surfaces in the preview are a new maximum reasoning effort setting for Sol and the subagent mode; everything else, including Luna's plan tier and price, is unannounced.
A separate thread of coverage, including a Seeking Alpha report and a TechTimes writeup, frames the launch as a software-only cost cut that lets OpenAI halve inference spend without new chips. That story traces to a paywalled Information newsletter and has not been confirmed in the GPT-5.6 system card material. It is a useful read on why the company might be willing to ship an orchestrator-heavy mode: cheaper submodels are exactly the inputs an ultra mode needs. It is not, on the public evidence, a verified fact about the new family.
The mechanism matters more than the codename. Subagent orchestration only pays off if the orchestrator is good enough to decompose tasks correctly, and if the submodels are cheap and fast enough that parallel execution beats a single larger call. OpenAI has shown pieces of this in research previews over the past year, but it has not put a number on the trade-off in public. The community is reading tea leaves, not benchmarks, and the company's silence on cost is itself a tell. What is public is the shape of the product: a flagship reasoning model, a balanced model, a fast cheap model, and a multi-agent mode that rides on top. The interesting question is not whether Ultra works. It is how OpenAI prices it, and whether subagent orchestration becomes the default shape of an AI coding tool or stays a power-user mode.
The next milestone is concrete: general availability for Sol, Terra, and Luna is slated for the coming weeks. Until then, the only public test of Ultra is the partner cohort, and the public corpus is thin: a blog post, a confirmation tweet, and a Hacker News thread. If the GA drop includes a Codex mode switch for Ultra, the subagent pitch becomes a product. If it does not, the architectural story waits another quarter.