The next era of AI at work is not going to be about chatting. It is going to be about getting actual work done — and that is exactly the pitch Microsoft made yesterday (September 3) when they announced that GPT-6 Astra, OpenAI’s newest frontier model, started rolling out through the Microsoft Foundry Limited Access Program.
I of course went to look for it in my own Foundry immediately. And… not yet! The reason is that Astra is behind the Limited Access Program, with availability expanding to participating customers over the coming days, so I do not have hands-on time with it to share this time. So this post is my read of what was announced. Hands-on will follow once I get access!
Let’s take a closer look.
What Astra actually is
This is the part that matters. Astra is built to take an open-ended goal, reason through it in multiple steps, build a plan, and produce a finished result. Not a paragraph of advice. A deliverable.
Microsoft groups it into three things:
- Deliberate planning and decision support – break a challenge into steps, evaluate options, communicate a recommendation, and identify the next actions for review.
- Polished, purposeful output – apply context, templates and quality standards across a workflow to produce documents, spreadsheets, presentations and analyses that are ready for review.
- Execution across applications – advanced tool use and computer use, so it can move between apps and complete multi-step tasks with appropriate human oversight.
That third one is the big shift. Astra’s computer-use capabilities are designed to work with familiar applications even where there is no dedicated API — interpreting what is on screen and interacting with approved interfaces. Updating records, navigating development tools, testing software, assembling results into a report.
Think about that for a moment for knowledge work. So much of our day is still glue work between systems that were never designed to talk to each other. This is aimed straight at that. I wrote earlier about computer using Cowork in Copilot Cowork browser use – First Look article, and this continues the story nicely.
However, Astra is much more than just computer using model – as interesting as that aspect is.
Astra is built to take an open-ended challenge, reason through it in multiple steps, create a plan, and produce a polished result. It can weigh trade-offs, incorporate new direction as work progresses, and use tools across applications and systems.
The specs – 1.05M context and new reasoning levels
From OpenAI’s own model page, the numbers are:
- 1,050,000-token context window
- 128,000 max output tokens
- Knowledge cutoff: April 30, 2026
- Input: text and image. Output: text.
- Reasoning effort now goes beyond high with two new levels: xhigh and max
- Tool support covers computer use, hosted shell, apply patch, skills, MCP and tool search. Fine-tuning is not supported.
A million-token window is not just a bigger bucket. It means a whole project — the specs, the meeting notes, the backlog, the previous drafts — can live in one working session without constantly being squeezed into a summary. For long agentic runs, that is the difference between an agent that remembers why something failed and one that quietly forgets.
Being able to dial reasoning from low all the way up to max is also very practical. Not every task deserves maximum compute, and now that is a per-call decision.
Pricing
Microsoft published Astra pricing for Standard deployments, including Global and U.S. Data Zone. It is consumption-based, so teams can start building without committing to reserved capacity.
| Deployment | Input | Cached input | Cached writes | Output |
|---|---|---|---|---|
| Standard Global (Short context) | $10.00 | $1.00 | $12.50 | $50.00 |
| Standard Global (Long context) | $20.00 | $2.00 | $25.00 | $75.00 |
| Standard Data Zone (US) (Short context) | $11.00 | $1.10 | $13.75 | $55.00 |
| Standard Data Zone (US) (Long context) | $22.00 | $2.20 | $27.50 | $82.50 |
(USD per million tokens, per Microsoft.)
So long context is genuinely a premium tier — roughly double on input. That is something to design for, not stumble into. Microsoft notes Astra is designed for token efficiency on complex work, and that actual usage and costs vary by workload and configuration. Which is the honest answer: your bill depends on how you build.
And a note for us here in Europe — the published deployment options are Global and U.S. Data Zone.
Enterprise controls – because this capability demands containment
I really like that Microsoft did not gloss over this part. When a model can operate software on your behalf, the risk surface changes completely. Content in an application may be incomplete, misleading, or deliberately crafted to influence an agent’s behavior. That is prompt injection with hands.
Foundry’s answer is the fundamentals: Microsoft Entra identity and access management, encryption in transit and at rest, private networking options, role-based access controls, content filtering, safety evaluations, monitoring and governance tools. Microsoft also states plainly that prompts and outputs are not used to train the models.
The design guidance is the bit I would print out and stick on the wall: scoped credentials, approved resources, human checkpoints for consequential actions, and activity records aligned to your risk requirements. And Microsoft is refreshingly clear that none of this eliminates risk or replaces your own responsibility to select and configure the right controls.
OpenAI describes Astra as its most aligned model to date and plans to publish alignment, safety and computer-use evaluations in its launch materials.
Where I would start
Microsoft’s own list of scenarios reads like a Future Work to-do list:
- Software engineering – reproduce complex bugs, investigate likely causes, propose fixes, prepare changes for developer review
- Business intelligence – build and refine dashboards in Power BI, compare data and prepare insights to share
- Professional work – documents, spreadsheets and presentations that follow existing templates and business standards
- Application workflows – updating customer records, processing forms, testing websites, working through approved interfaces where APIs are limited
When access opens up: explore the model in Foundry Models, then build with the Foundry Agent Service to bring cross-application task execution into your workflows.
So what does this mean?
I am very curious about this one. Not because of a benchmark number, but because the shape of the product changed. A million tokens of context, plus real computer use, plus enterprise-grade governance in Foundry — that is the combination that finally makes “an agent completes a unit of work” something you can actually put in front of a compliance team.
The honest caveats: it is a Limited Access Program rollout, so most of us are waiting. The long-context tier means the temptation to throw everything into that million-token window comes with a real bill attached. And the deployment regions are Global and US for now.
Work is changing. This one moves it.
Read Microsoft GPT-6 Astra in Foundry announcement.
Have you already got GPT-6 Astra in your Foundry? I would like to hear what you are building with it!
Vesa Nopanen - Future Work today 