OpenAI's Thibault Sottiaux on GPT-5.6: multi-agent Ultra mode, 3x faster computer use, and a 'Jarvis-style' voice experience

GPT-5.6 Sol: multi-agent Ultra mode, 3x faster computer use, and a Jarvis-style voice layer

Thibault Sottiaux, OpenAI's Head of Core Products and Platform, sat down amid what he describes as five to ten simultaneous war rooms to walk through the GPT-5.6 Sol launch. The headline capability is Ultra mode, which orchestrates up to eight agents working in parallel — OpenAI's answer to scaling test-time compute without simply throwing more tokens at a single model.

Multi-agent Ultra mode

Sottiaux says the benchmark results were "blown away" territory across coding, cybersecurity, long-context reasoning, website generation, and document production. Ultra mode is the public abstraction of what had previously required significant prompt engineering and bespoke multi-agent scaffolding. The pitch is that Sol now understands human intent well enough that users don't need to decompose work into subtasks themselves — the model reasons over constraints like deadlines and dependencies on its own. For genuinely hard problems, Ultra can run for two days and "leave no stone unturned."

“For the first time, we also really cracked multi agent setups, which we shipped as the ultra mode. And when you just see that going and you've got, like, eight agents collaborating together, communicating, and getting the same work done faster — it's just another way to scale test time compute... Compare it to five five, it's just really, like, three times faster [for computer use].”

Computer use

Computer use on GPT-5.6 Sol is roughly 3x faster than GPT-5.5. Sottiaux attributes this to two compounding forces: a dedicated team building bespoke improvements for Windows, Mac, and mobile; and general model efficiency gains that reduce latency and token cost, which directly benefit computer use by cutting correction loops when the model misclicks or misreads a screen.

Efficiency and enterprise controls

Sottiaux flags token efficiency as the defining commercial theme of the year. Enterprises pushed back on what he calls the "token maxing" era, asking for steerable, controllable models with spend dashboards and predictable ROI. GPT-5.6 Sol is positioned as the answer — significantly fewer tokens for equivalent results.

For users who want the opposite trade-off, OpenAI is running a Cerebras-accelerated version at roughly 750 tokens per second, about an order of magnitude faster than the standard API, though at higher cost.

ChatGPT Work and the product layer

The consumer-facing integration is ChatGPT Work, which connects to email, calendar, and documents from within the existing ChatGPT mobile app. Sottiaux frames this as deliberate simplicity — no new app, just a toggle. The goal is a conversation that feels like delegating to a smart coworker rather than configuring an automation.

Voice

The voice layer shipped alongside the model and is already live in the ChatGPT app, though not yet in the desktop app. Sottiaux describes the target experience as "a modern-day Jarvis" — walking into a room, speaking naturally, and having the model act on the computer with the same precision currently available over text.

Cross-device continuity

Sottiaux's framing on mobile versus desktop is that the device boundary itself is the problem to eliminate. A task started on a phone shouldn't be stranded there. The model should surface approvals on mobile when something sensitive needs authorization, even if the underlying work is running elsewhere.

The through-line across all of it is that the interaction model is shifting away from prompt engineering toward intent delegation. Whether that holds as power users find the edges of Ultra mode is the open question — Sottiaux acknowledges the frontier will keep moving, and novel agent configurations in cybersecurity and other domains will continue to reward users who push harder.