GPT-6 Astra in Microsoft Foundry – Vesa Nopanen – Future Work today

The following period of AI at work is just not going to be about chatting. It’ll be about getting precise work performed — and that’s precisely the pitch Microsoft made yesterday (September 3) after they introduced that GPT-6 Astra, OpenAI’s latest frontier mannequin, began rolling out by means of the Microsoft Foundry Restricted Entry Program.

I after all went to search for it in my very own Foundry instantly. And… not but! The reason being that Astra is behind the Restricted Entry Program, with availability increasing to taking part prospects over the approaching days, so I should not have hands-on time with it to share this time. So this submit is my learn of what was introduced. Palms-on will observe as soon as I get entry!

Let’s take a more in-depth look.

What Astra truly is

That is the half that issues. Astra is constructed to take an open-ended purpose, motive by means of it in a number of steps, construct a plan, and produce a completed consequence. Not a paragraph of recommendation. A deliverable.

Microsoft teams it into three issues:

• Deliberate planning and choice assist – break a problem into steps, consider choices, talk a advice, and determine the following actions for evaluation.
• Polished, purposeful output – apply context, templates and high quality requirements throughout a workflow to supply paperwork, spreadsheets, shows and analyses which are prepared for evaluation.
• Execution throughout purposes – superior instrument use and laptop use, so it might transfer between apps and full multi-step duties with applicable human oversight.

That third one is the large shift. Astra’s computer-use capabilities are designed to work with acquainted purposes even the place there isn’t any devoted API — decoding what’s on display and interacting with accredited interfaces. Updating information, navigating improvement instruments, testing software program, assembling outcomes right into a report.

Take into consideration that for a second for information work. A lot of our day remains to be glue work between techniques that had been by no means designed to speak to one another. That is aimed straight at that. I wrote earlier about laptop utilizing Cowork in Copilot Cowork browser use – First Look article, and this continues the story properly.

Nonetheless, Astra is rather more than simply laptop utilizing mannequin – as fascinating as that side is.

Astra is constructed to take an open-ended problem, motive by means of it in a number of steps, create a plan, and produce a refined consequence. It could possibly weigh trade-offs, incorporate new route as work progresses, and use instruments throughout purposes and techniques. 

The specs – 1.05M context and new reasoning ranges

From OpenAI’s personal mannequin web page, the numbers are:

• 1,050,000-token context window
• 128,000 max output tokens
• Information cutoff: April 30, 2026
• Enter: textual content and picture. Output: textual content.
• Reasoning effort now goes past excessive with two new ranges: xhigh and max
• Device assist covers laptop use, hosted shell, apply patch, abilities, MCP and power search. High-quality-tuning is just not supported.

1,000,000-token window isn’t just an even bigger bucket. It means an entire challenge — the specs, the assembly notes, the backlog, the earlier drafts — can reside in a single working session with out continuously being squeezed right into a abstract. For lengthy agentic runs, that’s the distinction between an agent that remembers why one thing failed and one which quietly forgets.

Having the ability to dial reasoning from low all the best way as much as max can be very sensible. Not each activity deserves most compute, and now that could be a per-call choice.

Pricing

Microsoft printed Astra pricing for Normal deployments, together with World and U.S. Knowledge Zone. It’s consumption-based, so groups can begin constructing with out committing to reserved capability.

(USD per million tokens, per Microsoft.)

So lengthy context is genuinely a premium tier — roughly double on enter. That’s one thing to design for, not stumble into. Microsoft notes Astra is designed for token effectivity on complicated work, and that precise utilization and prices fluctuate by workload and configuration. Which is the trustworthy reply: your invoice relies on the way you construct.

And a be aware for us right here in Europe — the printed deployment choices are World and U.S. Knowledge Zone.

Enterprise controls – as a result of this functionality calls for containment

I actually like that Microsoft didn’t gloss over this half. When a mannequin can function software program in your behalf, the chance floor modifications utterly. Content material in an utility could also be incomplete, deceptive, or intentionally crafted to affect an agent’s habits. That’s immediate injection with palms.

Foundry’s reply is the basics: Microsoft Entra id and entry administration, encryption in transit and at relaxation, non-public networking choices, role-based entry controls, content material filtering, security evaluations, monitoring and governance instruments. Microsoft additionally states plainly that prompts and outputs should not used to coach the fashions.

The design steerage is the bit I’d print out and stick on the wall: scoped credentials, accredited assets, human checkpoints for consequential actions, and exercise information aligned to your danger necessities. And Microsoft is refreshingly clear that none of this eliminates danger or replaces your individual accountability to pick out and configure the best controls.

OpenAI describes Astra as its most aligned mannequin up to now and plans to publish alignment, security and computer-use evaluations in its launch supplies.

The place I’d begin

Microsoft’s personal listing of eventualities reads like a Future Work to-do listing:

• Software program engineering – reproduce complicated bugs, examine possible causes, suggest fixes, put together modifications for developer evaluation
• Enterprise intelligence – construct and refine dashboards in Energy BI, examine knowledge and put together insights to share
• Skilled work – paperwork, spreadsheets and shows that observe current templates and enterprise requirements
• Software workflows – updating buyer information, processing kinds, testing web sites, working by means of accredited interfaces the place APIs are restricted

When entry opens up: discover the mannequin in Foundry Fashions, then construct with the Foundry Agent Service to deliver cross-application activity execution into your workflows.

So what does this imply?

I’m very inquisitive about this one. Not due to a benchmark quantity, however as a result of the form of the product modified. 1,000,000 tokens of context, plus actual laptop use, plus enterprise-grade governance in Foundry — that’s the mixture that lastly makes “an agent completes a unit of labor” one thing you possibly can truly put in entrance of a compliance crew.

The trustworthy caveats: it’s a Restricted Entry Program rollout, so most of us are ready. The long-context tier means the temptation to throw all the pieces into that million-token window comes with an actual invoice hooked up. And the deployment areas are World and US for now.

Work is altering. This one strikes it.

Read Microsoft GPT-6 Astra in Foundry announcement.

Have you ever already acquired GPT-6 Astra in your Foundry? I wish to hear what you might be constructing with it!

Printed by Vesa Nopanen

Vesa “Vesku” Nopanen, Principal Guide and Microsoft MVP (Microsoft 365 and Azure AI Foundry) engaged on Future Work at Sulava MEA.

I work, weblog and talk about Future Work : AI, Microsoft 365, Copilot, Loop, Azure, and different providers & platforms within the cloud connecting digital and bodily and folks collectively.

I’ve 30 years of expertise in IT enterprise on a number of industries, domains, and roles.
View all posts by Vesa Nopanen