http://x.com/i/article/2075435927011491840

I still think 5.6 is underhyped.

Real work is not just a repo of markdown files and a git repo

High level, high leverage work, whether you're a developer or not, spans crosses email, slack, multiple, documents, multiple tabs in my browser, and native software.

It's no longer just about finding the context, but actually working on things for long periods of time autonomously, delegating, learning to delegate work, and work across your entire computer.

One of the big surprises that I learned about during the live stream today was that our research team used Sol to help post-train Luna, how much better do we do with multi-agents, and how our product teams turned that same kind of agentic work into something people can use across their applications

Sol trained Luna

To be clear, this was not a model inventing a whole training recipe from scratch. Most of the config already existed from Sol. For this kind of smaller-model setup, our team's estimate is usually one or two researchers for a week or two, not a benchmark result. In this internal run, Sol spun it up in a few hours.

In one internal run, it checked a research branch, created the Luna config, launched the training job, and babysat it. The only human intervention was giving it the right permissions.

I think that is the part we should call out. A model helped with the work of building the next model. If you do model research, setting up jobs, running them, and watching what happens is exactly the kind of work you should be putting in front of Sol.

It has been incredible to see our research teams and product teams work together on this. Once one model can carry a workflow, the next question is obvious: what happens when it can bring in help?

More sub-agents with Ultra Mode

In case folks do not know the lineup yet:

1. Sol is the flagship.

1. Terra is for everyday work.

1. Luna is the fast, affordable one for high-volume work.

All these models now have Ultra mode, which is the higher compute setting across the family. It's effectively an extra mode that changes the way these models use sub-agents. The goal is that you can basically sign up for using Subagents and acknowledge that these will often times be a little bit more expensive.

On the SEC-Bench Pro chart below, the multiagent setup reaches stronger results at lower simulated latency than the one-agent baseline.

So what does that buy you? You spend more tokens to get a hard job back sooner, with a stronger result. Ultra is available in ChatGPT Work for Pro and Enterprise users, and in Codex on Plus and higher plans.

Sol is already efficient before you turn Ultra on. On DeepSWE v1.1, a long-horizon coding benchmark, Sol reaches 72.7%; the launch chart reports 36.2% lower estimated API cost than Fable 5. On the broader Artificial Analysis Intelligence Index, Sol with max reasoning comes within one point of Fable 5 while completing tasks in 61% less time at roughly half the estimated cost.

Using a Computer

We've also made tons of investments in human data, and computer use in 5.6 has been incredible. With often better performance than fable at 3x speed at 1/2 cost .

Here is a good example. Inside OpenAI, finance teams took month-end close and forecasting from days to hours. ChatGPT Work helped find the source data, move it into Excel or Sheets, reconcile it, create slides, and verify the result. The team got to spend more time explaining what changed instead of assembling the data.

This is why I am so excited about computer use and plugins.

Plugins bring in context from Slack, Google Drive, Microsoft 365, and the rest of your work. Computer use handles the steps that do not have a clean API. If you are a provider, we also have a smooth way of getting your plugins in our directories, giving you the distribution you want to succeed.

And the in-app browser now has:

1. auth and logins;

1. multiple browser tabs;

1. browser work inside the ChatGPT desktop app.

Now the agent can work through logged-in sites there while you keep using your computer. Desktop Computer Use is the separate background capability for actions across your local apps, files, and browser.

A good answer is not enough. The job has to get done.

Try it out

The easiest way to see this is to install a few plugins you already use. Open the plugins directory in ChatGPT, connect something like Slack, Google Drive, or Microsoft 365, then ask:

What do you know about me and the work I am doing? Use the plugins I connected. Tell me what projects I am working on, what seems important right now, and what context you are still missing. What should I be thinking about and do next?

You will know pretty quickly whether it has enough context to help. Then give it something real to do.

That is the frontier with 5.6 Sol. A model can help set up the next model, bring in sub-agents when the work gets hard, and use the same logged-in tools you use to move a real job forward.

Download ChatGPT for desktop · Read the GPT-5.6 launch announcement