GPT-6 Astra is now available
GPT-6 Astra has joined the model list, with support for EU hosting and ready for deployment. Enterprise admins can enable it in the admin settings and make it available to their teams.
For everyday work, try turning research notes into a structured management report, preparing a client proposal using your company's template, or analyzing spreadsheet data to identify trends and explain the findings.
It is also designed to handle evolving requests more consistently. For example, you can ask it to shorten a report for an executive audience, add a new requirement, or revise a recommendation while keeping the original constraints in place, rather than starting over.
Your tenant's default model remains under admin control. Adding GPT-6 Astra does not automatically change it.
Model pricing and task-cost estimates for enterprise admins
Choosing models for your team also means understanding their cost. Tenant admins can now see input and output prices alongside an estimated cost per task in the model settings, making it easier to compare options before enabling them.
The cost per task is an estimate, not a fixed price or a final bill. Use it as a comparison guide. Actual charges depend on usage and may differ from the estimate shown.
Agent tasks: clearer plans for complex work
For more complex requests, your agent can now organize the work into a structured task plan before it begins. Each task includes a clear objective and description, helping the agent work through longer projects in a more focused and reliable way.
You can follow the plan as it progresses, see which tasks are complete, which one the agent is currently working on, and what comes next. This gives you more transparency into what the agent is doing at any moment without interrupting its work.
Zeno's computer (Beta): a separate machine for heavy work
Zeno's computer is available on paid plans, and rolling out gradually.
Big multi-step jobs used to strain the regular sandbox and sometimes stall partway.
Zeno's computer runs that work on its own separate machine: a supervised coding agent that explores, edits, and verifies across many steps, so long tasks like building a financial model, analyzing several data files, or generating a full report actually finish. You give it the goal, watch it work in a live activity view, and get the finished files back in the chat.
How Zeno's computer gets triggered
It's automatic in the sense that the assistant decides to call it based on the task, not the user.
AI Presentations (Beta): review the outline before the build
Beta feature. AI Presentations is in beta. Access is available on request and requires an extra usage plan. Reach out to us to enable it for your account or team.
Building a deck is expensive, and a wrong storyline wastes the whole build. So it now plans with you before it generates a single slide.
First it researches your topic and asks a few real qualifying questions (including ones where more than one answer applies). Then it shows you a proposed outline: each slide's position, its action title, the one message it needs to land, and the layout chosen for it, plus alternatives where the choice is genuinely open. It flags any slide where a figure is missing or a source is shaky, so you catch gaps while fixing them is still cheap. You approve or correct the plan, then it builds.
Together with July's self-check, that's the whole loop: plan the story up front, verify the slides at the end.
Sort models by what matters to you
The model list can be long, and if you don't follow model names closely it's hard to know which one to pick. You can now sort the model picker by what you actually care about, instead of scanning names.
Open the model picker, choose Sort, and pick one:
- Default: alphabetical (A to Z).
- Quality: highest benchmark intelligence first.
- Speed: fastest models first.
- Quality + Speed: the best balance of the two.
If you're not sure what to choose, sort by Quality + Speed and pick from the top.
Edit skill files right from the skill page
Keeping a skill's contents current used to be awkward. Now you can browse a skill's files from its detail page, add new files to an existing skill, and edit them in place, with upload progress as you go. It's a proper editing workflow instead of a workaround.
Smarter connector recovery
When a connected account needs attention, the in-chat reconnect prompt now names the exact account that failed instead of a generic message. Handy if you have several Microsoft or other accounts connected.