OpenAI didn’t just ship a feature update. They reorganized the entire product around a single idea: go from chat to done. The result is a ChatGPT that feels less like a conversation tool and more like a lightweight operating system for knowledge work.
Here’s what’s new, what’s actually useful, and what’s still rough around the edges.
The Work Tab Is the Headline Feature
A new Work tab sits alongside the standard chat interface. Think of it as a structured task runner: you define a source (uploaded files, connected plugins, web search), tell it what output you want, and it executes in steps.
Say you hand it a folder of quarterly reports in different formats—PDFs, spreadsheets, a few Word docs—and ask for a single executive summary deck. Work mode maps out a plan, pulls from each source, and builds toward the output rather than just responding to a single prompt.
This is available on all plans, including free. That’s a meaningful differentiator. Comparable agentic workflows in competing tools sit behind paid tiers.
Inputs, Steps, Outputs: The Mental Model
The interface makes the pipeline explicit. On the left: what’s coming in. On the right: what’s being produced. In between: the steps the model is taking, visible in real time. If you’ve ever tried to wrangle a multi-step AI task through a regular chat window, this structure alone saves significant friction.
Sites: Hosted Web Pages, No Disclaimer Banner
One of the cleaner new features is Sites—the ability to generate a fully hosted web page directly from a prompt. The URL is readable and shareable. The page opens without any platform branding or “this content is AI-generated” overlay cluttering the top.
For practical use: a freelancer could mock up a simple project proposal page for a client call, or a small team could spin up a clean internal resource page in minutes. It won’t replace a real CMS, but for throwaway or lightweight publishing, it’s genuinely fast.
Sites is paid plans only.
Presentations Got a Real Upgrade
The new model family handles document synthesis noticeably better than previous versions. A workflow that’s immediately useful: drop in a brand style guide image alongside your raw notes, ask for a PowerPoint deck, and the model generates custom images (using ChatGPT’s image generation, which is best-in-class right now), builds slides around them, and delivers an editable .pptx file.
The first pass may be generic. The second pass—where you point it at a visual reference and say “apply this aesthetic”—is where it gets interesting. Color palette, layout density, background texture: it picks up on those cues well.
Expect a 10–15 minute generation time for a full deck with custom images. Not instant, but largely hands-off.
The New Models: GPT-5.6 Saul (and the Ultra Setting)
The updated model lineup includes GPT-5.6 Saul, which OpenAI positions at roughly the same capability tier as Gemini’s top models on standard benchmarks. Early user reports suggest it’s strong at synthesis and document-heavy tasks, though it may not yet match the best competing models on complex multi-step planning.
The Ultra setting, available on the $200/month Pro plan, coordinates a minimum of four model instances simultaneously on your task. It’s powerful—one user reportedly used it to consolidate ten large PDFs into a single 700-page comprehensive document for $20—but it burns through usage limits fast. Use it deliberately, not as the default.
For most people: the standard Saul model with Work mode will cover 90% of use cases.
The Desktop App Now Does Three Jobs
OpenAI consolidated several standalone tools into one desktop app:
- Atlas (their separate web browsing app) is discontinued; browsing is now built in
- Codex (the standalone coding agent) is folded in
- The original desktop chat app is replaced
The result is a single app that handles chat, agentic work tasks, screen recording for context, and deep web research. If you’ve been juggling multiple OpenAI apps, that’s over.
Voice Mode Now Talks Over You (On Purpose)
The updated voice mode supports interruption and overlap—you can speak while it’s responding, and it will adjust in real time rather than stopping and restarting. The practical application OpenAI highlighted is live translation.
In back-and-forth testing between English and another language, the turnaround felt close to instantaneous and the translation held up under natural conversational pacing. For anyone who does cross-language calls or interviews, this is worth testing.
What This Means in Practice
The through-line across all of these changes is consolidation and accessibility. OpenAI is making the capable stuff harder to miss and easier to use without a learning curve. The free tier getting Work mode access is a strategic move—it lowers the barrier to experiencing agentic AI for the first time.
If you’re on a paid plan, start with Sites and the presentation workflow. If you’re on free, open the Work tab and try a simple multi-source task: a few uploaded documents, a clear output format. The structure will make the capability click faster than any demo.
