GPT-6 Astra in Copilot Cowork and in Foundry – Vesa Nopanen – Future Work today

After I wrote about GPT-6 Astra in Microsoft Foundry a number of weeks in the past, I didn’t but have entry. There was loads to debate about what the mannequin promised, however essentially the most attention-grabbing half was nonetheless lacking: utilizing it in my very own work.

Now I can begin doing that! I’m writing this text with GPT-6 Astra chosen in Copilot Cowork, utilizing info I collected with Copilot because the beginning materials and my blog-writing ability to information the draft. A barely meta instance: utilizing Astra to assist write about utilizing Astra. However let’s preserve the excellence clear. It is a actual writing activity, not but a comparability proving that Astra is best than the opposite out there fashions. The attention-grabbing query is not simply what Astra can theoretically do. It’s what adjustments once I put it to work.

Let’s take a more in-depth look.

From a mannequin announcement to a piece possibility

For me, the most important change will not be one other set of specs. I coated these within the earlier article. It’s the shorter distance between listening to a few mannequin and utilizing it for one thing I really have to get executed.

My start line right here was not an empty immediate asking for an article about GPT-6. I introduced analysis, my earlier submit, a course for the follow-up, and my very own writing steering. I additionally recognized particulars that also wanted checking.

That’s a lot nearer to how I need to work with AI.

Right here is the context. Here’s what I’m making an attempt to attain. Assist me transfer it ahead, with out pretending the lacking items are already identified.

For this text, meaning separating bulletins from observations, avoiding one other launch recap, and leaving room for checks I’ve not accomplished but. A refined draft that quietly invents these outcomes can be worse than an unfinished one which exhibits precisely what’s lacking.

The writing is simply a part of the duty. Understanding what to not embrace and refining the outcome from the AI draft is essential.

Foundry and Cowork – the mannequin is simply a part of the expertise

I see two completely different beginning factors right here.

With Microsoft Foundry, my query is about constructing an answer. How ought to an utility or agent be designed, related, evaluated, deployed, and operated?

With Copilot Cowork, my query is concerning the work in entrance of me. What can I delegate, what context ought to I present, and the way a lot checking will the outcome want?

These questions are associated, however they aren’t interchangeable.

Choosing Astra in Cowork will not be the identical factor as deploying Astra in Foundry and constructing an utility round it. The encircling expertise issues: directions, instruments, out there info, permissions, and the way in which the work is coordinated.

That additionally issues when evaluating outcomes.

If an article draft seems effectively, I can’t robotically credit score the mannequin for all the things. Background info supplied are important a part of the setup. Likewise, a disappointing outcome would possibly reveal lacking context slightly than a limitation in reasoning.

I need to consider the entire means of working, not simply admire the mannequin badge.

For my very own work, Cowork is the place to start out once I need assist finishing a considerable activity in an current work expertise. Foundry turns into related once I have to construct and function a repeatable answer with express architectural and deployment decisions.

This text is a place to begin, not a verdict

Writing a follow-up is definitely a helpful first activity. There may be an current argument to proceed with out repeating it. There are technical particulars with completely different ranges of certainty. There’s a private voice to protect. And there’s a temptation to show an enthusiastic introduction right into a conclusion earlier than the proof exists.

For this draft, I would like Astra to assist with three issues:

• Discover the brand new story. What has modified sufficient to justify one other article?
• Preserve the boundaries seen. What’s documented, what have I personally noticed, and what stays untested?
• Produce one thing price enhancing. Not merely extra textual content.

That final level is necessary. An extended response will not be robotically a greater outcome.

If I spend extra time eradicating repetition, checking unsupported statements, and placing my very own perspective again into the draft, the obvious productiveness achieve can disappear fairly shortly.

That is additionally why I discover the mixture of a mannequin and a reusable writing ability attention-grabbing (sure, I’ve a ability in Cowork that helps me write weblog posts). It supplies steering about tone, construction, and learn how to deal with uncertainty slightly than clarify all the things once more from scratch.

Astra versus Auto – make it earn the choice

For my subsequent checks, I need to evaluate Auto with explicitly chosen Astra.

I’m not beginning with the idea that manually selecting Astra should be higher. If Auto will get me an appropriate outcome with much less effort, that’s helpful too. And it typically is – I’ll normally simply use Auto for many duties.

I might begin with duties that resemble my precise work:

• Flip analysis and private notes into an article with a transparent argument.
• Evaluate a number of paperwork and produce a suggestion that preserves disagreements and uncertainties.
• Construct a workshop define from a quick, with a coherent circulate and sensible workouts.

I need to preserve the supply materials, directions, and acceptance standards constant. I might additionally use separate duties in order that one try doesn’t profit from corrections made throughout the different.

Then I can look past which response feels extra spectacular at first look. Did it observe the transient? Did it protect necessary {qualifications}? Did it miss a supply? How a lot rewriting was needed? Was the deliverable usable?

Reasoning effort – what does this activity deserve?

Cowork’s mannequin steering describes 5 reasoning-effort decisions: Gentle, Medium, Excessive, Further Excessive, and Max. Larger effort entails a trade-off in response time and Copilot Credit. Microsoft explains this beneath Set the reasoning effort degree to steadiness high quality, pace, and price in Choose a model for Copilot Cowork.

That makes this extra attention-grabbing than merely selecting a mannequin.

My proposed comparability is to run the identical substantial activity at Medium, Excessive, and Max, then file:

• Time to an preliminary deliverable.
• Credit consumed, the place utilization will be reliably attributed to the duty.
• Clarification questions and correction rounds.
• Factual errors or missed necessities.
• My very own evaluate and enhancing time.
• Whether or not I accepted the outcome, revised it, or began once more.

If I can’t measure task-level credit score consumption reliably, I’ll say so slightly than estimate it.

The query I would like answered is easy: does the additional reasoning (and price) cut back the overall effort wanted to get an appropriate outcome?

A slower run could possibly be worthwhile if it saves substantial evaluate. A quicker one could possibly be the higher selection when the duty is simple. And Max ought to need to display its worth—not change into my default as a result of it sounds reassuring.

The identical applies to value. A less expensive particular person run will not be essentially a less expensive accomplished activity if it wants a number of retries and intensive guide correction. However, spending extra doesn’t robotically purchase an enchancment that issues.

I need to measure the entire journey to one thing I can use.

A observe for us in Europe – documented, however not but seen in my Foundry EUR Datazone

In my first Astra submit, I famous that the printed deployment choices have been World and U.S. Knowledge Zone. There may be now an replace for us in Europe.

Microsoft’s September 22 article, GPT-6 Astra, Sol and Luna: For production agents in Microsoft Foundry, explicitly consists of Astra in EU Knowledge Zone availability.

The Foundry Models region availability page additionally listed Astra for the EU Knowledge Zone once I checked.

After all, I went to look in my very own Foundry. And… not but at the very least for me!

On the morning of September 23, 2026, Astra was not out there for me to deploy to the EU Knowledge Zone. I might deploy it to World Commonplace or U.S. Knowledge Zone.

GPT-6 Sol, alternatively, was already out there to deploy to the EU Knowledge Zone.

It appears to me like EUR Datazone availability for Astra should be rolling out. Given the printed availability, I count on this to occur quickly.

If processing location issues on your venture and buyer, verify the mannequin and deployment sort in your personal subscription earlier than committing to an structure or promising availability to a buyer.

Another distinction: Foundry deployment choices don’t set up the place a Copilot Cowork activity utilizing GPT-6 Astra is processed. These preparations must be checked individually. At this second GPT-6 Astra in Cowork is a preview mannequin and runs utilizing OpenAI subprocessor, which doesn’t outline the datazone. Now that Astra is in Foundry, this hopefully adjustments quickly.

What did writing this text value?

I’ve additionally been following the Copilot Credit consumed whereas engaged on this text, utilizing /value in Cowork.

Earlier than including this value info and additional updates, the overall was 232 Copilot Credit. That was a lot lower than I anticipated.

That is an commentary from this explicit writing session with GPT-6 Astra chosen—not a set worth for an article, and never a comparability with Auto or one other mannequin. It additionally excludes the updates made after that studying.

Nonetheless, it provides me a concrete start line. As an alternative of solely discussing what deeper reasoning may cost, I now have an precise utilization determine from my very own work.

The following query is how that consumption compares with the standard of the outcome and the time I spend reviewing and enhancing it. Credit are one a part of the associated fee. My very own time is one other.

From a great reply to accepted work

I’m excited that I can now embrace Astra in my very own Cowork experiments. However I don’t want the conclusion of this follow-up to be {that a} new mannequin appeared and subsequently all the things improved. The helpful conclusion will come from the work.

For this text, success means a follow-up that provides one thing new, retains unsure particulars seen, and takes much less effort to complete with out decreasing the standard. For one more activity, the acceptance standards shall be completely different.

That’s the place I need to focus subsequent: the associated fee and reliability of a accomplished, reviewed, accepted final result.

Not simply what the mannequin produced however what I might really use.

That could be a extra helpful dialog about the way forward for work than selecting a favorite mannequin and assuming it belongs in each activity. We have to be taught the place deeper reasoning helps, the place a lighter strategy is sufficient, and the place human judgement stays important.

Now, on to the testing!