OpenAI x ChatGPT Work launch: The chat box grows an agent, a desktop, and a website builder
A 35-minute keynote turning GPT-5.6 into an enterprise work surface, told entirely through OpenAI’s own demos.
The short version, if you do not have the half hour:
OpenAI launched the GPT-5.6 model family (Sol, Terra, Luna) alongside three products: ChatGPT Work, a new desktop app, and hosted Sites. Sol reaches paid plans over 24 hours, Terra and Luna reach free users.
ChatGPT Work splits the app into “Chat” (quick answers) and “Work” (agentic tasks): a demo had it pull Slack and employee feedback, find people, and schedule meetings, then run a finance variance analysis and update an Excel model, a PowerPoint, and a shareable Site.
The desktop app reaches local files, browser tabs, and other Mac apps. One demo turned a folder of team documents plus three Chrome tabs into a slide deck in about 90 seconds. Another gave ChatGPT its own on-screen cursor to reorganize Apple Notes in the background.
Sites lets the model generate interactive, publishable web pages and 3D visualizations on demand with no code, built from connectors like Slack and Gmail, and one presenter said one was built with “no Figma.”
On cost, OpenAI claimed GPT-5.6 beats competitors on an internal “eval 1.1” at “less than half the cost,” and said experts call it better than anything on the market “while being three times as fast.” A new Ultra mode runs a team of agents in parallel.
Everything here is OpenAI describing and demoing its own product, so the claims and the on-stage runs are exactly that, claims and demos, not independently verified outcomes.
Every one of these numbers and demos comes from OpenAI selling OpenAI, so I am reading the keynote as a set of claims and staged runs, not as verified outcomes, and I will say so once here rather than hedge every line. What is worth tracking is the shape of the product OpenAI just drew, more than whether any single demo worked on stage, because the shape is a direct move on the software companies whose whole value is being the surface where knowledge work happens.
The frame from the opening is that the chat box is no longer the product. OpenAI says almost a billion people use ChatGPT every week to write, get answers, and do research, and the argument is that coding showed the models could do far more than answer.
“We saw a transformation of software development and shortly after internally within OpenAI, everyone was using these models to do so much more than getting great answers. They were using these models to do great things.”
Three products follow from that framing, and each one pushes ChatGPT further onto surfaces that other software vendors currently own.
The models: Sol, Terra, Luna
The launch underneath the products is the GPT-5.6 family. OpenAI split it into three tiers by cost and speed: Sol is pitched as the most powerful model for the hardest agentic workflows, Terra as a faster model for everyday work, and Luna as the fastest and most affordable for high-volume jobs. The researchers framed the gains as the product of two curves multiplied together, reinforcement learning on top of pre-training, dating the reasoning approach to a 2024 announcement.
“In 2024, we announced the reasoning paradigm, which is a reinforcement learning technique to tackle the hardest tasks. And since then, we’ve been scaling both reinforcement learning and pre-training. And we’ve seen exponential improvements in model capabilities as these two multiplied together.”
The tiering does real work for the enterprise pitch. A three-model line lets OpenAI put a cheap model under high-volume automated work and reserve the expensive one for the hard agentic tasks, which is how you get to an aggressive cost claim without running the top model on everything. The one detail I would flag as a genuinely different capability, if it holds, is the self-improvement loop: the researchers said Sol autonomously post-trained Luna from a short, underspecified Codex prompt, work they framed as something a team of senior researchers used to do.
“5.6 Sol actually autonomously post-trained Luna. This is the actual Codex prompt that we, a researcher on our team, used to have Sol kick off a post-training job for Luna. It’s a really short prompt.”
They paired that with a workflow claim, more pull requests per researcher and more experiments run, and said Codex made every plot in the livestream and helped build the slide deck. Read as a claim, it is the same story every AI lab now tells about its own internal productivity, and it is unaudited from the outside. What to watch is whether the “automated researcher is pretty close” line shows up as a measurable release cadence rather than a keynote aside.
ChatGPT Work: the app splits in two
The headline product reorganizes ChatGPT into two modes. “Chat” stays the familiar quick-answer surface. “Work” is the agentic side, and the demos are all about a task running for minutes across your tools rather than a reply coming back in seconds. The first demo had an engineer, Jessica, prep for the launch by asking the model to mine internal feedback and then act on it.
“Hey ChatGPT, I would love to get an understanding of how people are using the ChatGPT work feature internally. Look through Slack and employee feedback and find people in diverse roles that have used the product in really interesting ways. I’m in San Francisco this week for the launch. I would love to meet up with a few people and really deep dive on their use cases in person.”
The claim was that it pulled from the named sources, found people, and scheduled the meetings through calendar. The second demo, from the finance team, is the one aimed squarely at enterprise software, and it is worth stating that the presenters twice called the figures “demo numbers.” The workflow ran a variance analysis on a monthly close, updated an Excel forecast model, produced a PowerPoint, built a shareable Site with the same analysis, and sent the link over Slack.
“This used to take so much manual work. We would have to reconcile multiple systems, our Excel forecast model across multiple cases. And now in one pass, ChatGPT can run the variance analysis for me. So we can see why we beat our forecast and where we still have risk. I can even have ChatGPT work go in and propose an updated forecast case for us.”
The spreadsheet-to-deck-to-shareable-page chain is the demo the whole event rests on. It walks straight through the daily output of a finance analyst and lands on a Site instead of a file, which is a plain signal of what OpenAI wants Work to replace. Whether a real month’s messy reconciliation survives “one pass” the way a rehearsed demo did is the open question, and the “few steers” framing suggests it still needs a human driving.
The desktop app: reaching your files, tabs, and other apps
The desktop app is the piece that moves ChatGPT off the browser tab and onto the machine. The presenters said everything from the web app is there, plus access to local files, browser tabs, and other apps. The marquee demo turned a folder of documents from several teams, a launch-readiness PDF, user interviews, a security review with pentest results and compliance controls, plus three open Chrome tabs, into a finished slide deck.
“Can you please look at the materials in this folder that everybody sent me? I also have three open tabs in Chrome that have content for this launch. Go look at it and make me a presentation that I can give to my team in the template that we use all the time at the company.”
The claim was about 90 seconds to a ready deck that adhered to the company template, and the model also drew on memory of prior conversations about the launch. The more striking demo was computer use: the presenter asked ChatGPT to organize a messy Apple Notes, and the model took over its own on-screen cursor and started moving notes and making folders in the background while the human did other things.
“With the incredible computer use that’s built in and with the advances in the new model, it’s going to get its own cursor. It’s going to start operating Apple Notes in the background. You can see here this is not my cursor. This is cursor.”
An agent that drives your actual desktop applications is a different risk surface than a chat reply, and OpenAI is clearly aware of it, which is part of why the safety section runs as long as it does. The competitive read is that a model reaching local files, the browser, and native apps starts to overlap with the operating system’s own assistant layer, well beyond any single SaaS tool.
Sites: generated, interactive web pages on demand
Sites is the third product, available to all paid users, and it is where the finance demo kept landing. The pitch is that the model generates interactive, publishable web pages, dashboards, visualizations, and even 3D scenes as part of an answer, with no code written by the user, and that they are collaborative and shareable in one click. A designer on the team, Ed, showed a rich interactive site built from his connectors.
“I kicked this prompt off this morning and it’s looked through all of my connectors and plugins that we talked about earlier. I’ve connected my Slack, my Gmail, everything that I use every day, and it’s gone through all of those sources and it’s come back and turned it into this super rich interactive website.”
The design-quality claim was pointed. Asked whether he handed the model a Figma file for the layout, the designer said no.
“Did you give it a Figma for this? No, no Figma. This was all just the model.”
The internal examples ran from collaborative dashboards that a web team built to replace an Excel spreadsheet, to interactive prototypes of a new model selector, to a lighthearted 3D version of the pelican-on-a-tricycle test that circulates as a front-end benchmark. The recurring line was that anyone can now build the thing without a designer or a front-end engineer in the loop, which is the same claim being made against Figma-style design tools, no-code site builders, and internal BI dashboards at once. The honest caveat the designer offered is that it is “best when you give it a little bit of guidance,” so the ceiling shown on stage was assisted output that still needed a person guiding it.
The research read: benchmarks, computer use, and cost
The research segment is where the reusable numbers live, and they are all OpenAI’s own. GPT-5.6 Sol was called state-of-the-art on Terminal-Bench for coding, on BrowseComp for finding hard-to-locate information, and on Agents’ Last Exam for long-horizon professional work. The most investable capability claim is computer use, because it is the thing that turns a chat model into a worker.
“5.6 is particularly good at computer use, so anything that involves navigating a browser, navigating apps on your desktop. It helps medical assistants navigate electronic health records. It helps data scientists analyze if a drug is effective. It can even help investment bankers create financial models.”
The presenter said experts using it in real workflows call it better than anything else on the market “while being three times as fast.” Read against the earlier finance and desktop demos, the message stays consistent. OpenAI is selling an agent that operates the same software an analyst, an assistant, or a banker operates today. The number to hold onto is that every one of these superlatives is self-reported, and none of the named professional use cases came with a named customer.
Cost and Ultra mode
The pricing argument is the one that most directly threatens incumbent software economics. OpenAI claimed token efficiency has been a multi-year focus and that GPT-5.6 out-scores competitors at a fraction of the cost.
“You can see this reflected on eval 1.1 where GPT-5.6 outperforms its competitors at less than half of the cost.”
“Eval 1.1” was not defined on stage, and “competitors” were not named, so I would treat “less than half the cost” as a directional claim rather than a comparable figure. On the other end of the range, OpenAI announced Ultra mode, which runs a team of agents in parallel and, per the demo, gets better and faster as more agents are added. That is the same barbell the model tiering implies: a cheap high-volume floor with Luna, and a compute-heavy ceiling with Ultra for the jobs a customer will pay for. For a buyer, the read is that per-seat pricing and per-token pricing are about to collide with per-task work that used to require a person.
Safety and the cyber angle
OpenAI spent real time on safety, which reads as both genuine and as pre-empting the obvious objection to an agent that drives your desktop. The stated inputs were large: over 700,000 A100-equivalent compute hours on red teaming, six weeks of safety training and testing, and new classifiers and activation probes as monitoring upgrades. The cyber framing was two-sided.
“With GPT 5.6 Sol, researchers have already found vulnerabilities in every major browser and database. As part of the Daybreak umbrella we started Patch the Planet, where we work directly with open source contributors. For Linux, they accepted over half of our patches.”
Finding vulnerabilities in every major browser and database is a capability claim that cuts both ways, and OpenAI’s answer is Project Daybreak plus early access for defenders. The Linux “over half of our patches accepted” figure is the one externally checkable data point in the whole safety section, since kernel maintainers are the ones accepting, so it is the number I would try to verify first. The rest is OpenAI grading its own alignment work.
Who this pressures
The competitive read-through is the point of the event, and it is broad rather than aimed at any one rival. By generating decks, dashboards, and interactive sites from raw files, ChatGPT Work steps onto the same ground as the office-productivity suites, the BI and dashboard tools, and the no-code site and app builders. The “no Figma” claim points at design tooling. The finance demo points at the spreadsheet-plus-slides workflow that is the daily bread of the productivity suites. The connector story, pulling from Slack and Gmail, points at every SaaS tool that assumed it owned its own data surface, since an agent that reads across all of them becomes the layer the user actually lives in.
The demo that should worry incumbents most is the desktop cursor. A model that operates native applications directly does not need each vendor to build an integration, it can drive the existing UI, which weakens the moat of being the app someone already has open. The honest limits are equally clear from OpenAI’s own framing: the good runs needed “a few steers” and “a little bit of guidance,” the finance figures were demo numbers, the cost and speed superlatives were self-reported, and no external enterprise customer was named on stage. What I am watching is the gap between the staged one-pass demo and a real analyst’s messy Monday close, because that gap is exactly where the enterprise-software incumbents still live, and it is the variable that decides whether Work replaces a workflow or just accelerates one.