The Operating Layer Inside Your Working Day
Microsoft just split the Copilot bill in two. A per-user license covers everyday AI. Cowork, Code, Autopilot and the frontier models now run on usage-based billing with Copilot Credits. And from November 2, 2026, thirty-two days out, new Microsoft 365 Copilot Business licenses bought through a CSP partner arrive with usage-based billing switched on by default. The seat price stopped being the whole price. Underneath that, Copilot is still turning into an agentic layer across the tenant. Work IQ supplies context. Chat answers. Cowork executes. Three model vendors sit in the picker. And which Microsoft page you read still decides what you believe.
The teammate that starts with your organization’s context
Copilot’s moat isn’t a model. It’s the work graph, right? Mail, meetings, files, people, activity. Copilot can reason across all of it while respecting the permissions Microsoft 365 already holds.
Context-rich and process-aware
Copilot is strongest when the task leans on what your organization has already said, decided, created and permitted.
- Tenant-aware grounding
- Native business-app outputs
- Enterprise identity and governance
Work that should finish where it began
Meeting follow-up, document creation, spreadsheet analysis, presentation production, inbox work and governed agents.
- Less copy-and-paste
- Fewer tool handoffs
- Work stays connected to its source systems
Context and control outrank model loyalty
Use Copilot when the right answer depends on organizational context, and has to inherit permissions, compliance and brand systems.
- Your work already lives in Microsoft 365
- IT needs visibility and policy control
- The output belongs in a Microsoft app
The bill changed shape
The biggest change is a bill, not a feature. Microsoft has restated how Copilot is sold. A per-user license covers everyday AI. Usage-based billing with Copilot Credits covers Cowork, Code, Autopilot, new advanced work in apps like SharePoint, and the frontier models. Enterprise tenants get a safety catch, because those metered services stay off until an admin writes a spending policy. Copilot Business bought through a CSP partner loses that catch on November 2. And the seat price is now only half of what you’re budgeting.
Copilot is now a license plus a meter Sept 25
The commercial shape changed, and that’s the thing to walk into a budget meeting with. Microsoft now describes everyday AI as priced per user per month, with no change to what you already bought. The advanced half runs on usage-based billing with Copilot Credits. That covers Cowork, Code, Autopilot, new advanced experiences in apps like SharePoint, and frontier models like Fable and Astra. Model access rides along with it: Microsoft says GPT-5.6 and Sonnet 5 are included, and new models like Opus 5 will be included with limits. Here’s the part that protects you. For enterprise customers, the metered services stay off until an admin creates a spending policy in the Microsoft 365 admin center. Admins set budgets at tenant and group level, including user-level caps inside group policies, plus alerts when spend hits a threshold. Cost management for Copilot Studio agents and for Code is described as coming in October.
· Source: “Evolution of the Copilot pricing model,” Microsoft Copilot Blog, September 25, 2026, updated September 28, 2026, by Nicole Herskowitz
- Per-user license for everyday AI; Copilot Credits for the rest
- Metered: Cowork, Code, Autopilot, advanced work in SharePoint, frontier models
- Enterprise metered services stay off until an admin writes a spending policy
- Budgets and alerts at tenant, group and user level
- Copilot Studio and Code cost management stated as coming in October
Usage-based billing arrives switched on, 32 days out Nov 2
Read this one if you buy Copilot through a partner. Starting November 2, 2026, new Microsoft 365 Copilot Business licenses purchased through a Cloud Solution Provider include usage-based billing by default. Microsoft calls pay-as-you-go the default billing configuration, so customers reach eligible usage-based experiences, including Copilot Cowork, Work IQ APIs and GitHub Copilot Harness, with less billing setup. It applies to new purchases of Microsoft 365 Copilot Business, standalone and in bundles, through CSP. Microsoft tells partners to get familiar with the change and with the preset monthly limit. So a seat you can describe as $18.00 a year now ships with a meter nobody asked to turn on. Find the preset limit before the first invoice finds you. Diarize the date, re-read the source page on the same cadence, and keep a dated copy of the row you depend on. A date written down once and trusted is how you miss a deadline that moved.
· Source: Microsoft Partner Center announcements, September 16, 2026 entry, checked October 1, 2026
- Effective November 2, 2026, 32 days after October 1, 2026
- New Microsoft 365 Copilot Business CSP licenses only, standalone and bundles
- Covers Copilot Cowork, Work IQ APIs and GitHub Copilot Harness
- A preset monthly limit applies; get the number from your partner
- Enterprise tenants still need an admin spending policy first
CSP promotions just ended Sept 30
If a quote on your desk was built on a Copilot promotion, it expired. Microsoft 365 E5, Microsoft 365 E7 and Copilot promotions ended September 30, 2026 as part of the move to growth margins, and growth margins reached general availability on October 1, 2026. Nonprofits get a separate path. Eligible nonprofits can combine promotional offers of 15, 20, 30 and 40 percent on top of the 15 percent nonprofit offer already included with the Microsoft 365 Copilot SKU, available through December 31, 2026. There’s also a new Frontier Partner specialization, with Microsoft Agent pre-purchase plan credits and Microsoft 365 E7 licensing among its benefits.
· Source: Microsoft Partner Center announcements, September 23, 24 and 30, 2026 entries, checked October 1, 2026
- E5, E7 and Copilot promotions ended September 30, 2026
- Growth margins generally available October 1, 2026
- Nonprofit offers stack through December 31, 2026
- A Frontier Partner specialization is now live
- Re-price any quote written before September 30
What the Copilot add-on costs, and what it no longer covers Read Oct 1
The commercial shape, in one line. Chat is included, the full Copilot add-on is $30.00 on top of a base plan, and the agentic half runs on a meter beside it. Now the detail. Microsoft lists Microsoft 365 Copilot at $30.00 user/month paid yearly, and also publishes $31.50 paid monthly on an annual commitment. It’s an annual subscription that auto-renews, and a separate license for a qualifying Microsoft 365 plan is required. Microsoft 365 Copilot Chat is listed as included for all Microsoft Entra account users with an eligible Microsoft 365 subscription. What changed is what sits beyond the seat. Microsoft’s own list of usage-based services now runs to Microsoft 365 Copilot, Cowork, advanced work in SharePoint, advanced work in OneDrive, the Work IQ API for custom apps and agents, Copilot Managed Runtime and Teams Phone Agent. Those pricing pages carry no publication date. Prices read October 1, 2026, so confirm with your reseller before a budget is signed.
- Microsoft 365 Copilot add-on at $30.00 user/month paid yearly, $31.50 paid monthly
- A qualifying Microsoft 365 plan license is required separately
- Copilot Chat listed as included with an eligible subscription
- Metered beside the seat: Cowork, SharePoint and OneDrive advanced work, Work IQ API, Copilot Managed Runtime, Teams Phone Agent
- Pricing pages are undated; read October 1, 2026
Copilot Credits, with the rate card that does exist Read Oct 1
“Metered” isn’t a budget line, so here are published numbers. Microsoft publishes a full Copilot Credits rate card for Copilot Studio, set out in the table below. Capacity is bought in packs at $200.00 for 25,000 Copilot Credits. Microsoft describes that as per tenant per month, not a one-time purchase. Now do the division. That’s about $0.008 a credit, eight tenths of a cent. So a generative answer at 2 credits costs about 1.6 cents. That figure is arithmetic, not a Microsoft rate. Microsoft hasn’t published a per-credit price. Commit Units bought up front take up to 20 percent off. And pay-as-you-go coverage kicks in automatically through an Azure meter when prepaid credits run out. Now the hole. No Microsoft page publishes per-task Cowork or Work IQ API consumption rates. And the licensing guide Microsoft names as authoritative isn’t publicly retrievable. You can price a Copilot Studio agent. You still can’t price a Cowork task.
· Rate card read October 1, 2026 from “Billing rates and management,” Microsoft Copilot Studio, stamp 2026-08-03
- Copilot Studio rate card published; see the table below
- $200.00 for 25,000 Copilot Credits, described as per tenant per month
- About $0.008 a credit by division; Microsoft publishes no per-credit rate
- Up to 20 percent off with Commit Units; automatic Azure-meter overflow
- Cowork and Work IQ API per-task rates remain unpublished
Microsoft 365 E7, Agent 365 and the tiers around them Read Oct 1
The Frontier Suite tier was announced March 9, 2026 and reached general availability May 1, 2026. Prices below were read on Microsoft’s own pricing pages on October 1, 2026. E7 at $99.00 user/month, paid yearly on an annual commitment. A no-Teams SKU at $90.45. Agent 365 standalone at $15.00 user/month, paid yearly. Around them sit E5 at $60.00 and E3 at $39.00. Each carries a no-Teams variant that’s easy to drop from a comparison: E5 No Teams at $51.45 and E3 No Teams at $30.45. On the business side the annual and monthly figures diverge enough to matter. Business Premium with Copilot is $32.00 paid yearly or $38.40 monthly. Business Standard with Copilot is $23.50 paid yearly or $28.20 monthly. Copilot Business starts from $18.00 paid yearly, down from $21.00, and Microsoft publishes the add-on at $25.20 user/month, with and without Teams. Cowork is listed as metered on every Business tier that carries it. And from November 2, 2026, a new Copilot Business license bought through CSP arrives with usage-based billing already switched on. These pages are undated, so confirm with your reseller.
- E7 $99.00, E7 No Teams $90.45, Agent 365 standalone $15.00
- E5 $60.00 and E5 No Teams $51.45; E3 $39.00 and E3 No Teams $30.45
- Business Premium with Copilot $32.00 yearly or $38.40 monthly
- Business Standard with Copilot $23.50 yearly or $28.20 monthly
- Copilot Business from $18.00 yearly, down from $21.00; $25.20 user/month, with or without Teams
- Cowork metered on top of whichever seat you buy
Microsoft 365 G7 was promised for October 1. It isn’t there Unmet, Oct 1
Check this before you tell a public-sector buyer it shipped. On September 15, 2026 Microsoft announced that Microsoft 365 G7 and Agent 365 would both go on sale to GCC customers on October 1. Capabilities would expand in phases. Here’s what Microsoft’s own government plans page showed on October 1. A banner says G7 and Agent 365 bring the next era of AI innovation to GCC. Below it, the plans you can actually buy are G5, G3, Business Premium for GCC-High, and Office 365 G5 and G3. G7 isn’t one of them. No G7 plan. No G7 price. No confirmation post on the Microsoft Cloud government blog or the Microsoft 365 blog. That isn’t a cancellation, and it isn’t a slip Microsoft has announced. It’s a promise with nothing published behind it. One more thing worth carrying. The announcement publishes no component list for G7. It says G7 brings together productivity, AI, agents, security, identity, compliance and governance. And it treats Agent 365 as sold alongside rather than bundled inside, describing Agent 365 as the control plane for agents in G7. So don’t quote a bundle.
· Source: “Introducing Microsoft 365 G7: Intelligence + Trust for the mission ahead,” Microsoft Cloud Blog, September 15, 2026, by Chris Barry. Government plans page read October 1, 2026
- Promised purchase availability for GCC customers on October 1, 2026
- The government plans page lists no G7 plan and no G7 price
- No confirmation post published since September 15
- The announcement publishes no component list
- Agent 365 is described as sold alongside G7, as its agent control plane
Consumer Copilot walks in the door with your staff Read Oct 1
Your people turn up with opinions formed on the consumer tiers. So you should know what those are. Microsoft’s individuals pricing page lists four. Microsoft 365 Personal at $9.99 per month or $99.99 per year, with advanced Copilot chat at higher usage than free. Family at $12.99 per month or $129.99 per year, for one to six people, with AI for the subscription owner only. Premium at $19.99 per month or $199.99 per year, with extensive Copilot chat usage plus limited agentic AI usage. And Pro at $99.99 per month, monthly billing only, with the highest chat usage and the highest agentic AI usage. Two cautions. “Copilot Pro” as a standalone consumer SKU no longer resolves, so the ladder is Personal, Family, Premium and Pro. And no Microsoft consumer page names which models any of these tiers get, so no consumer model access is stated here. The page is undated. Prices read October 1, 2026.
- Personal $9.99 monthly or $99.99 yearly
- Family $12.99 monthly or $129.99 yearly, one to six people, AI for the owner only
- Premium $19.99 monthly or $199.99 yearly, plus limited agentic AI
- Pro $99.99 monthly, monthly billing only
- The old standalone Copilot Pro page no longer resolves
- No first-party page states which models consumer tiers receive
| Copilot Studio activity | Copilot Credits | Implied cost at $200.00 per 25,000 credits |
|---|---|---|
| Classic answer | 1 | About $0.008 |
| Generative answer | 2 | About $0.016 |
| Agent action | 5 | About $0.04 |
| Tenant graph grounding | 10 | About $0.08 |
| Agent flow actions | 13 per 100 actions | About $0.10 per 100 actions |
| Content processing | 8 per page | About $0.064 per page |
| Text and generative AI tools, basic | 0.1 per 1,000 tokens | About $0.0008 per 1,000 tokens |
| Text and generative AI tools, standard | 1.5 per 1,000 tokens | About $0.012 per 1,000 tokens |
| Text and generative AI tools, premium | 10 per 1,000 tokens | About $0.08 per 1,000 tokens |
| Voice, by tier | 10, 35 or 75 per minute | About $0.08, $0.28 or $0.60 per minute |
The model list outran the page that prints it
Copilot’s model choice grew four times in nine days. The Learn page everyone quotes hasn’t moved since September 14. And Microsoft and OpenAI now print different answers about whether some models exist at all. So stop trusting a copied list. The list that binds is the one in your own picker, right? Everything else is a record of what somebody announced, and when.
The model list grew four times in nine days The current lineup
Don’t publish a Copilot model list copied off a Learn page. Microsoft announced Claude Opus 5.5 and GPT-6 Sol into Copilot model choice on September 22, 2026, then GPT-6.1 Sol and Claude Sonnet 5.5 on September 30, 2026. Those are Microsoft’s Copilot dates, not launch dates. Anthropic launched Opus 5.5 on September 22 and Sonnet 5.5 on September 28. OpenAI shipped GPT-6.1 Sol on September 29. None of the four appear on the Cowork model page, which still reads, verbatim: Auto; GPT 5.5 (Frontier); GPT 5.6 Sol; GPT 5.6 Terra; GPT 6 Astra; Opus 5; Claude Sonnet 5; Claude Fable 5.1. That page hasn’t moved since 2026-09-14. The Cowork what’s-new page, the Cowork FAQ, the Copilot Studio model page and the subprocessor page don’t name the four either. One Microsoft page does name two of them. “Claude models in Microsoft Foundry” lists claude-opus-5-5 and claude-sonnet-5-5 as generally available. But Foundry is a separate Azure purchase, not the Copilot picker. Neither announcement body renders, so the Copilot surfaces, admin gates and retention terms for all four are unestablished. What you can say: Microsoft announced them. What you can’t say: where they’re selectable in Copilot, or on what terms. Your own picker is the only list that binds.
· Sources: “More Models, One Copilot,” September 22, 2026, and “Available today: OpenAI’s GPT-6.1 Sol and Claude Sonnet 5.5 in Microsoft Copilot,” September 30, 2026, both Microsoft Copilot Blog, by Xia_Song
· Launch dates: Anthropic’s “Introducing Claude Opus 5.5,” September 22, 2026, and “Introducing Claude Sonnet 5.5,” September 28, 2026; OpenAI’s “Introducing GPT-6.1 Sol,” September 29, 2026
- September 22: Claude Opus 5.5 and GPT-6 Sol
- September 30: GPT-6.1 Sol and Claude Sonnet 5.5
- Microsoft’s announcement dates; Anthropic launched Sonnet 5.5 on September 28
- The Cowork model page still prints the eight-item list from 2026-09-14
- Absent from the Cowork model, what’s-new and FAQ pages, the Copilot Studio model page and the subprocessor page
- Foundry lists Opus 5.5 and Sonnet 5.5 as GA, as a separate Azure purchase
Microsoft and OpenAI print different answers about the same models Two vendors, one model
GPT 5.6 Terra and GPT 5.6 Luna are gone from OpenAI’s own pricing page and models index, deleted with no deprecation entry. Microsoft still sells them and still documents them. The Cowork model page offers GPT 5.6 Terra as balanced effort for common tasks. Foundry’s “Models sold directly by Azure,” stamped 2026-09-23, lists gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna as available. The Azure OpenAI model retirement schedule, stamped 2026-09-21, gives all three a retirement date of 2028-01-11. So a buyer reading Microsoft sees a model with a retirement date in January 2028. A buyer reading OpenAI sees a model that doesn’t exist. And it’s already happening one generation up. OpenAI shipped GPT-6 Sol on September 22. It superseded it with GPT-6.1 Sol on September 29. GPT-6 Sol is already gone from OpenAI’s models index, again with no deprecation entry. Microsoft announced GPT-6 Sol into Copilot and priced it in Foundry on September 22. Its retirement schedule lists no GPT-6 model at all, so it publishes no end date. Seven days at OpenAI. No exit on Microsoft’s pages. Both sides are first-party, and neither tells you the other disagrees. If you depend on any of these models, the vendor you actually buy from is the only one whose notice period binds you.
· Readings taken October 1, 2026 from the Cowork model page, “Models sold directly by Azure” (stamp 2026-09-23) and “Azure OpenAI model retirements” (stamp 2026-09-21)
· OpenAI: “Introducing GPT-6 Sol and Luna,” September 22, 2026, updated September 29, and “Introducing GPT-6.1 Sol,” September 29, 2026. OpenAI’s models index is a living page, read October 1, 2026
- OpenAI removed GPT-5.6 Terra and Luna from its pricing page and models index, with no deprecation entry
- Microsoft still lists both as available and sells both
- Azure retirement date for GPT 5.6 Sol, Terra and Luna: 2028-01-11
- GPT-6 Sol: shipped September 22, superseded by GPT-6.1 Sol September 29, absent from OpenAI’s models index with no deprecation entry
- The Azure retirement schedule lists no GPT-6 model at all
- Get the commitment in writing from whichever vendor holds your contract
The slow page is now the one that prints the model list The new instance
Which Microsoft page you read still decides what you believe. The example just flipped. The tenant-wide “Release Notes for Microsoft 365 Copilot” page carries an updated stamp of 2026-09-24, with a newest entry of September 23, 2026. The Cowork what’s-new page is stamped 2026-09-25. One day apart. The page that fell behind is the Cowork model page, stamped 2026-09-14 and 17 days old, and that’s the page people quote as the current lineup. It’s four models short. Worse sits next door. The Copilot Studio page “Select a primary AI model for your agent,” stamped 2026-09-18, lists nothing newer than GPT-5.5 Chat, GPT-5.5 Reasoning and Claude Opus 4.7. No GPT-6 of any kind. No Claude 5 of any kind. Which means it contradicts Microsoft’s own announcement that Astra rolled out to Copilot Studio. A newer stamp is not a newer list. So check which page you’re on, then check your picker.
· Readings taken October 1, 2026 on pages Microsoft rewrites in place
- Release notes stamped 2026-09-24, newest entry September 23: 7 and 8 days old
- Cowork what’s-new stamped 2026-09-25: 6 days old
- Cowork model page stamped 2026-09-14: 17 days old and four models short
- Copilot Studio model page stamped 2026-09-18 lists no GPT-6 and no Claude 5
- A fresh stamp records an edit, not a complete list
GPT 6 Astra in Copilot, and the page that doesn’t know Sept 4
Astra, not GPT 5.6 Sol, is the top OpenAI option on the Cowork model page. Microsoft says “GPT-6 Astra is rolling out today to users in Copilot Cowork and Copilot Studio.” Access is administrator-managed: “Admins can manage access and configure availability for their organization through the Microsoft 365 admin center,” and “Availability may vary by region and organization.” The Cowork model page carries Astra as the latest model for tough problems. Now the catch. Copilot Studio’s own model page, stamped 2026-09-18 and therefore newer than the announcement, lists no Astra and no GPT-6 of any kind. Its newest entries are GPT-5.5 Chat, GPT-5.5 Reasoning and Claude Opus 4.7. The dated announcement is the record of what Microsoft announced. The live Studio page governs what’s supported now, and as read on October 1, 2026 it doesn’t list Astra. It’s also the page your agent builders will read. Confirm Astra in Studio in your own tenant before you design an agent around it.
· Source: “Available today: OpenAI GPT-6 Astra in Microsoft Copilot,” Microsoft Copilot Blog, September 4, 2026, by Bas Brekelmans. Studio model list read October 1, 2026
- Copilot Cowork and Copilot Studio, per Microsoft
- Admin-managed through the Microsoft 365 admin center
- Availability may vary by region and organization
- The Copilot Studio model page names no GPT-6 and no Claude 5
- No pricing and no retention language in the post
Full GPT-6 token pricing in Foundry, and none of it is a Copilot price Sept 22
Microsoft publishing a GPT-6 price isn’t Microsoft repricing Copilot. Those two get run together constantly. Per million tokens, Global Standard, short context: GPT-6 Astra at $10.00 input, $1.00 cached input, $12.50 cached writes and $50.00 output. GPT-6 Sol at $2.00, $0.20, $2.50 and $10.00. GPT-6 Luna at $0.10, $0.01, $0.125 and $0.50. Long context: GPT-6 Astra $20.00, $2.00, $25.00 and $75.00. GPT-6 Sol $4.00, $0.40, $5.00 and $15.00. GPT-6 Luna $0.20, $0.02, $0.25 and $0.75. The US Data Zone is exactly 10 percent more on every line. The EU Data Zone is 20 percent more, which is the number most comparisons miss. Standard deployment covers all 28 Global regions plus the US and EU Data Zones. One trap. GPT-6.1 Sol is a different model, and it isn’t on this table. OpenAI prices its cached input at $0.10 against a $2.00 input price, a 95 percent cache discount. GPT-6 Sol’s is 90 percent. Don’t price one off the other’s row. These are Microsoft Foundry token prices. No Copilot seat price changed alongside them, and the post publishes no retirement for GPT 5.6 Terra or Luna.
· Source: “GPT-6 Astra, Sol, and Luna for production agents in Microsoft Foundry,” Azure Blog, September 22, 2026, by Naomi Moneypenny
· GPT-6.1 Sol pricing: OpenAI, “Introducing GPT-6.1 Sol,” September 29, 2026
- GPT-6 Astra short context: $10.00 in, $1.00 cached in, $12.50 cached writes, $50.00 out
- GPT-6 Sol short context: $2.00, $0.20, $2.50, $10.00
- GPT-6 Luna short context: $0.10, $0.01, $0.125, $0.50
- US Data Zone plus 10 percent; EU Data Zone plus 20 percent
- GPT-6.1 Sol is not on this table; OpenAI prices its cached input at $0.10, a 95 percent discount
- Foundry token prices, not Copilot seat prices
Effort levels, now documented with a default On the model page
Higher reasoning effort uses more credits. Put that in front of a budget owner before anything else. Effort is a cost control as much as a quality control, and whoever runs the task sets it. The ladder itself is documented on the Cowork model page, with detail the August 2026 blog roundup didn’t carry. Medium is the default, and the page notes that the default can vary across models to meet the Cowork quality bar. Light is there for lighter tasks. High and Extra High for harder ones. Max for your hardest work.
- Medium is the default effort level
- The default can vary across models
- Light, Medium, High, Extra High and Max
- Higher effort consumes more Copilot Credits
- Model page, stamped 2026-09-14; confirm directly
Opus 5 Bare name on the page
Still on the Cowork model page, and still as the bare string “Opus 5” rather than “Claude Opus 5,” while the rows beside it read “Claude Sonnet 5” and “Claude Fable 5.1.” It matters for one reason: it defeats a search. A compliance reviewer hunting the full product name on that page finds nothing. Scope the absence to what was actually checked, though. The Azure Databricks page “Databricks-hosted foundation models,” stamped 2026-08-28, writes “Claude Opus 5” in full and exposes the endpoint databricks-claude-opus-5. So one Microsoft property names the model properly and the page your Copilot admins read doesn’t. The full name also appears on Microsoft 365 Roadmap item id 568372, dated July 24, 2026, status Launched. The roadmap still can’t confirm it. It’s a client-rendered application, a direct feature-id lookup returns nothing, and the roadmap RSS feed is a rolling window that doesn’t include the item. Carry that as an unconfirmed reading.
· Readings taken October 1, 2026
- Renders as bare “Opus 5” on the Cowork model page
- Azure Databricks page, stamped 2026-08-28, names it in full
- Endpoint
databricks-claude-opus-5 - Roadmap item id 568372 still cannot be confirmed by a direct fetch
- The absence claim covers the pages checked, not the whole doc set
Claude Sonnet 5 Everyday tier
Announced on July 2, 2026, rolling out in Copilot Cowork and PowerPoint. It’s still carried on the Cowork model page, verbatim, as “For everyday tasks and fast responses such as drafting, quick lookups, and day-to-day work.” Claude Sonnet 4.6 stays off the table. Read that beside the Opus 5 description and Microsoft hands you its own routing hint. Sonnet 5 is the everyday tier. Opus 5 is the escalation for complex, high-stakes work. And Fable 5.1 sits at the top of the Anthropic ladder on that page. One thing the page hasn’t caught up with. Microsoft announced Claude Sonnet 5.5 into Copilot model choice on September 30, 2026, two days after Anthropic launched it on September 28.
- Replaced Sonnet 4.6, which is no longer listed
- Documented for everyday tasks and fast responses
- Escalate to Opus 5 when the stakes justify it
- Fable 5.1 sits above both on that page
Sonnet plus Opus Advisor is gone Resolved
The paired mode, where Sonnet handled the turn and Opus reviewed the result, came off the Cowork model page in August and survived on the admin governance page. It doesn’t survive anywhere now. The governance page rewritten on 2026-09-14 names no Claude, Sonnet or Opus model, and never mentions Advisor. Notice how Microsoft resolved the disagreement. It deleted the losing side rather than reconciling the two lists. A contradiction can close because a page was corrected, or because a page was emptied. Those aren’t the same thing. So if you built a review workflow on Advisor, the documented basis for it is gone from both pages.
- Absent from the model page since August 14
- Now absent from the governance page as well
- The contradiction closed by deletion, not by reconciliation
- Design your own review pass instead
Retention and licensing depend on the route you buy through
Same model name, different contract. Fable 5.1 bought through Microsoft and Fable 5.1 bought from Anthropic carry different retention answers. And Anthropic’s own pages disagree about one of them. Add a third model vendor in the Office apps, a new data-retaining Anthropic family, and an agent security cliff that already passed on July 1. Hand these to privacy and licensing before anyone flips a switch.
Fable 5 is still selectable, behind an admin switch Gated, not gone
One naming point first, because it defeats searches. Fable 5 and Fable 5.0 are the same model. Microsoft writes it as Fable 5 on the Cowork pages and as Fable 5.0 on the subprocessor page. Now the substance. Fable 5 hasn’t left Cowork. The Copilot Cowork FAQ, stamped 2026-09-29, says the choices include GPT 5.5, GPT 5.6 and GPT 6 variants, Claude Opus, Claude Sonnet, and Claude Fable 5.1 and 5 (Preview). It adds that Claude Fable 5 (Preview) is off by default, so it appears only after an admin turns it on in the Microsoft 365 admin center under Copilot settings. The Cowork model page, 15 days older, simply omits it. Two Microsoft pages, one list, two answers. The FAQ is the newer one, and it matches the off-by-default gate the subprocessor page already described. So treat Fable 5 as live and gated, and look at the switch in your own admin center before you tell anyone it’s gone. Fable 5.1 is still the more advanced of the two.
· Readings taken October 1, 2026 from the Copilot Cowork FAQ (stamp 2026-09-29) and the Cowork model page (stamp 2026-09-14)
- Fable 5 and Fable 5.0 are the same model under two Microsoft spellings
- The Cowork FAQ, stamped 2026-09-29, lists Fable 5 (Preview) as a choice
- Off by default; it appears only once an admin enables it
- The Cowork model page omits it, and the FAQ is 15 days newer
- Scope a pilot to the picker in your tenant, not to either page
Retention depends on the route you buy through Two contracts, one model name
“The Fable line is the retained line.” Stop saying that. It’s too flat to survive contact with either vendor’s documentation. Microsoft treats the two Fable models differently inside its own services. Fable 5.0 carries extra controls in the Microsoft 365 admin center. It’s off by default for all scenarios, and nobody in your tenant can touch it until an admin explicitly opts in. Fable 5.1 can run another way. For certain organizations it’s available with Anthropic acting as a subprocessor, under Microsoft’s Product Terms and Data Protection Addendum. On that route Anthropic doesn’t retain customer content, prompts and responses included. Now read that for what it is. It’s Microsoft describing the terms Microsoft contracted for as the buyer of record. It isn’t a statement about Fable 5.1 wherever you meet it. Buy direct from Anthropic and the same model sits under Anthropic’s own terms, starting with 30-day retention. Zero data retention is where it gets messy, because Anthropic’s own pages disagree. Its September 1 API release note says no zero data retention on the 5.1 models unless Anthropic expressly authorizes it. Its two September 1 news pages say eligible customers get zero data retention on Fable 5 and Fable 5.1 until Enterprise Frontier Safeguards ships. Anthropic doesn’t publish who counts as eligible. So Microsoft and Anthropic aren’t contradicting each other here. They’re describing two contracts. The live conflict sits inside Anthropic. Same model name, two routes, and more than one possible answer on the direct one. The route you buy through decides which is yours.
· Anthropic: “Claude Fable 5.1 and Claude Mythos 5.1” and “Developing Enterprise Frontier Safeguards with our customers,” both September 1, 2026. The API release note is a page Anthropic rewrites in place, so read it directly
- Source: “Anthropic models in Microsoft Online Services,” Microsoft Learn, stamped 2026-09-18. Verbatim: Anthropic models with Data Retention, “such as Fable 5.0,” carry additional controls in the Microsoft 365 admin center and “are off by default for all scenarios,” with no user access “until the tenant admin explicitly opts in.” And: “For example, for certain organizations, Fable 5.1 is available with Anthropic acting as a subprocessor and Microsoft’s Product Terms and Data Protection Addendum (DPA) apply,” under which “Anthropic doesn’t retain customer content (including prompts and responses)”
- Inside Microsoft, Fable 5.0 is the gated, off-by-default retained model
- Through Microsoft, Fable 5.1 can run with Anthropic as a subprocessor under Product Terms and the DPA
- On that path Microsoft states Anthropic does not retain customer content
- Direct from Anthropic: 30-day retention, and two Anthropic answers on zero data retention
- A flat “the Fable line is retained” answer is unqualified rather than simply false
The retention tails privacy review actually asks about Numbers to carry
Thirty days is the headline, and it’s the least interesting number here. Microsoft says Anthropic, not Microsoft, stores most inputs and outputs for up to 30 days before deleting them. Anthropic’s own wording runs the other way: at least 30 days, then deletion, unless a safety investigation or a legal requirement keeps the data. Up to is a ceiling. At least is a floor. On the subprocessor route you contract with Microsoft, so Microsoft’s up to 30 days is the wording your agreement carries. Buy direct and Anthropic’s at least 30 days governs. Either way, get the ceiling in writing. Then the long tails. If Anthropic’s trust and safety classifiers identify potential violations of its Usage Policy, Anthropic may retain inputs and outputs for up to two years. And it may keep trust and safety classification scores for up to seven years. Those are the figures a privacy review stops on. The same Microsoft page now names a second data-retaining family too: Claude Mythos 5, alongside Claude Fable 5.
· Microsoft’s wording: the subprocessor page, stamped 2026-09-18. Anthropic’s wording: its API and data retention documentation, a living page with no date line
- Both pages are rewritten in place, so confirm the current wording before you paste it into a data protection impact assessment
- Up to 30 days for most inputs and outputs, in Microsoft’s wording
- At least 30 days, in Anthropic’s wording
- Up to two years on a usage-policy classifier hit
- Up to seven years for classification scores
- Anthropic holds the content, not Microsoft
A new Anthropic family nobody has reviewed yet Mythos
Add a row to the subprocessor register. Microsoft’s subprocessor page, stamped 2026-09-18, now says Anthropic models with Data Retention are advanced models, such as Claude Fable 5 and Claude Mythos 5, that require data retention by Anthropic. Mythos is new to that sentence and new to most registers. In Microsoft Foundry, claude-mythos-5-1, claude-mythos-5 and claude-mythos-preview are listed as gated research preview, while claude-opus-5-5 and claude-sonnet-5-5 are generally available both hosted on Azure and hosted on Anthropic infrastructure, and claude-fable-5-1 and claude-fable-5 are preview on Anthropic infrastructure. Foundry bills Claude models through Claude Consumption Units and requires an Azure Marketplace subscription, so it’s a separate purchase from a Copilot seat. One warning about that Foundry page. It carries a 2026-09-21 stamp, yet it lists Sonnet 5.5, which Anthropic launched on September 28. So read the stamp as a review date, not a content date. A gated, data-retaining model family becomes a privacy review item the moment it has a name, not the moment somebody requests it.
· Readings taken October 1, 2026 from “Anthropic models in Microsoft Online Services” (stamp 2026-09-18) and “Claude models in Microsoft Foundry” (stamp 2026-09-21)
- Mythos 5 named on the subprocessor page as a data-retention model
- Mythos 5.1, 5 and preview are gated research preview in Foundry
- Opus 5.5 and Sonnet 5.5 are GA in Foundry, on Azure and on Anthropic infrastructure
- Foundry bills Claude through Claude Consumption Units
- Foundry access is bought separately from a Copilot seat
- The Foundry page’s 2026-09-21 stamp is a review date, not a content date
The Anthropic switch has a default that depends on when your tenant was created Check this one
There’s a specific setting, in a specific place, with a date attached. Microsoft 365 admin center, then Copilot, Settings, View All, then “AI providers operating as Microsoft subprocessors.” You need the AI Administrator role to change it. It’s on by default for EU, EFTA and UK tenants created after March 25, 2026, and tenants created before that are told to check Message Center for their own state. One more thing catches people. The Anthropic independent processor setting was decommissioned on May 1, 2026, so Anthropic has to be enabled as a Microsoft subprocessor to reach Anthropic models at all. If your data protection impact assessment still describes an independent-processor route, it describes something that no longer exists.
· Reading taken October 1, 2026 from “Copilot in Microsoft 365 apps with Anthropic models,” stamp 2026-09-03
- Setting name: AI providers operating as Microsoft subprocessors
- Requires the AI Administrator role
- On by default for EU, EFTA and UK tenants created after March 25, 2026
- Older tenants: check Message Center for the current state
- The independent processor route was decommissioned May 1, 2026
Grok from SpaceXAI, a third vendor in the Microsoft 365 picker Sept 12
Microsoft 365 Copilot’s own model picker is no longer a two-vendor product. But scope that claim before you repeat it. On the wider Copilot brand, SpaceXAI was already a model vendor inside a Copilot-branded Microsoft product. xAI’s own dated articles put Grok 4.5 in GitHub Copilot on July 28, 2026 and Grok 4.6 on August 14, 2026. No Microsoft page confirms either. Microsoft’s own GitHub Copilot model documentation names no xAI model as a curated choice. It mentions xAI only as a bring-your-own-model provider. So carry those two dates as xAI’s, not Microsoft’s. Different product. Different admin gate. Different purchase. So September 12 is the date for Microsoft 365 Copilot, not the date SpaceXAI became a Copilot vendor. Here’s what landed. Grok in Microsoft Word, Excel and PowerPoint, through the Microsoft Frontier Program. Not Cowork. Not Copilot Studio. The admin gate is stricter than the Anthropic one: “Access to SpaceXAI models is managed through a dedicated administrative setting that is disabled by default.” Geography is limited too: “Grok is not available to Frontier customers in the European Union (EU), European Free Trade Association (EFTA), and the United Kingdom during the preview,” with no end date given. And Microsoft has published no retention terms for SpaceXAI models. That makes this a third subprocessor question for privacy review, not a fourth model in the same picker.
- Source: “Expanding model choice in Copilot with Grok,” by xiasong, published September 12, 2026 and updated September 14
- Word, Excel and PowerPoint via the Microsoft Frontier Program
- Not in Cowork and not in Copilot Studio
- Dedicated admin setting, disabled by default
- Not available to Frontier customers in the EU, EFTA and the UK during the preview
- Grok in GitHub Copilot: dated by xAI, July 28 and August 14, 2026; Microsoft’s GitHub Copilot docs name no xAI model
- No published retention terms; treat as an open subprocessor question
Grok’s absence from Cowork is unstated, not denied Boundary, not conflict
The Cowork model page doesn’t mention Grok, SpaceXAI or the Frontier Program. That reads as consistent with the Grok post, which scopes the rollout to Word, Excel and PowerPoint. The Grok post was itself updated on the same day the model page was rewritten. But record it accurately. No Microsoft page states that Grok is unavailable in Cowork. The Cowork page just leaves it out. An unstated boundary is a weaker thing to plan around than a documented one. So if a Cowork policy has to exclude SpaceXAI, write the exclusion into your admin setting. Don’t infer it from a page that never mentions the vendor.
- Grok, SpaceXAI and Frontier Program all absent from the Cowork model page
- No page says Grok is unavailable in Cowork
- Treat as an omission, not a documented limit
- Enforce the boundary in the admin setting, not by inference
The Agent 365 licensing question has an answer now Service description
Stop carrying two contradictory pages to your licensing desk. Carry one. The Microsoft Agent 365 service description, stamped 2026-09-15, says Agent 365 is available as a stand-alone subscription for use with eligible Microsoft 365 subscriptions. It’s also included with Microsoft 365 E7. Then it splits the feature set, which is the part that decides your design. Eligible Microsoft 365 plans, meaning Enterprise, Business, Education and Front Line Worker, get basic registry, governance actions, lifecycle rules and external sync. Ten more features need Microsoft 365 E7, G7 or Agent 365 standalone. They’re listed below. The older comparison page, stamped 2026-06-30, still calls Agent 365 a component included exclusively in Microsoft 365 E7. But a footnote on that same page says Agent 365 is also available as a standalone add-on. Sloppy wording, not a contradiction. And the Agent 365 overview still says it works best when using Microsoft E5 as a pre-requisite. That’s a recommendation, not a requirement.
· Reading taken October 1, 2026 from the Microsoft Agent 365 service description, stamp 2026-09-15
- Policy templates
- Observability
- Tool control
- Graph API access to the agent registry
- Access packages
- Conditional access
- Audit logs
- eDiscovery
- Data lifecycle management
- Defender and Intune integration
Agent security already switched off for tenants without Agent 365 July 1, 2026
This happened on July 1, 92 days before October 1, and almost nobody carried it. Since July 1, 2026, Microsoft requires an Agent 365 license for AI agent security on Copilot Studio and Foundry agents. Tenants without an eligible license lost those capabilities that day. What goes dark: agent discovery and posture, agent threat detection and real-time protection, and investigation of agent activity in Advanced Hunting. The AIAgentsInfo table is deprecated in favor of AgentsInfo. And here’s the trap. Tenants currently configured to Block on existing Agent 365 rules stop blocking on July 1, 2026. A control that quietly stops blocking is worse than one you never had, because the dashboard still looks fine. Go and check which side of that line your tenant sits on.
· Reading taken October 1, 2026 from “Transition Microsoft Copilot Studio and Microsoft Foundry agent security capabilities to Microsoft Agent 365,” Microsoft Defender XDR
- Effective July 1, 2026
- Agent discovery, posture, threat detection and Advanced Hunting investigation all require Agent 365
AIAgentsInfodeprecated in favor ofAgentsInfo- Existing Block rules stop blocking without the license
- Verify your own tenant rather than assuming the rules still fire
Most new features are a spend event or a governance surface
Read the September feature list with a budget owner beside you. Home, Code and Autopilot are metered. App building is metered. The plugin registry is a policy surface, not a catalog. And two of the most useful facts, the Cowork region limit and the Researcher cap, sit in FAQs nobody puts in a rollout plan.
Home, Code and Autopilot Sept 25
Copilot grew two new products and renamed the front door. Home combines chat and collaboration. Code lets people build applications and automations using conversational language. Autopilot is described as a persistent, proactive and personal agent hosted in the cloud, one that continues work when the user is offline. Rollout is staged: Home and Code roll out gradually to the Microsoft Frontier program over the coming weeks, and Autopilot expands to private preview at the end of the month. Microsoft tells partners the new Copilot is rolling out through the Frontier program, with capabilities arriving continuously. Here’s why this belongs in a budget conversation rather than a feature briefing. Code and Autopilot are both named in the usage-based billing list. An agent that works while nobody is watching is a spend profile, not a feature.
· Source: Microsoft Partner Center announcements, September 25, 2026 entry, checked October 1, 2026
- Home combines chat and collaboration
- Code builds apps and automations from conversational language
- Autopilot is a cloud-hosted agent that continues work offline
- Home and Code through the Frontier program; Autopilot in private preview
- Code and Autopilot are both metered with Copilot Credits
A plugin registry, and it’s a governance surface Sept 30
Plugins just got a front door and an off switch. Microsoft introduced a plugin registry as a shared place to discover, publish, manage and govern plugins across supported Copilot experiences. That means Copilot Home, which is Chat and Cowork, plus Copilot Code, Copilot Autopilot, Word, Excel, PowerPoint and SharePoint, with Copilot Studio, GitHub Copilot and Microsoft Foundry to follow. Admins can review Microsoft, custom-built and partner plugins, enable or disable which plugins are available, and block or assign them to users or groups. Microsoft says basic registry and management capabilities for Copilot experiences are included with qualifying Microsoft cloud subscriptions. Over 100 plugins at launch, including HubSpot, Asana, Linear, Notion, Canva and MongoDB. Treat the registry as where your plugin policy now lives, not as a catalog to browse.
· Source: “Introducing the plugin registry: one place to discover and govern plugins for Microsoft Copilot,” Microsoft Copilot Blog, September 30, 2026, by Rob Howard
- One registry across Home, Code, Autopilot, Word, Excel, PowerPoint and SharePoint
- Copilot Studio, GitHub Copilot and Foundry to follow
- Enable, disable, block or assign per user or group
- Basic registry capabilities included with qualifying subscriptions
- Over 100 plugins at launch
Three Cowork limits that aren’t in anyone’s rollout plan Read the FAQ
Region, students and what counts as spend. Microsoft says access to Anthropic models through Microsoft services is limited to the regions Anthropic currently supports, that Cowork isn’t an exception, and that Cowork use and access is currently limited to Anthropic-supported regions. On education: faculty and staff users in the education industry can access Cowork, and students aren’t eligible for access at this time. And on the meter: activities such as model responses, tool and skill calls, image generation and browser tasks count toward consumption, with administrators seeing consumption in the Microsoft 365 admin center and setting limits per user or group. That third one is the useful one. Browser tasks burn credits. So does every tool call inside a plan the user never sees.
· Readings taken October 1, 2026 from the Copilot Cowork FAQ, stamp 2026-09-29
- Cowork access is limited to Anthropic-supported regions
- Education: faculty and staff only; students are not eligible
- Model responses, tool and skill calls, image generation and browser tasks all consume credits
- Admins can set consumption limits per user or group
- Check regional coverage before you promise Cowork to a global team
Researcher has a hard monthly cap 25 queries
Twenty-five. That’s the number nobody quotes when they recommend Researcher. Microsoft documents a maximum of 25 Researcher queries per user per month. Researcher is part of the core default experience for Microsoft Copilot licensed users, Claude can be enabled inside it, it can’t process images as input, and it can’t be customized or extended using Copilot Studio. PowerPoint and PDF generation, memory, agent-to-agent conversations, report-length control and sovereign cloud availability are all listed as to be released soon. So if your research workflow assumes a few deep reports a week per analyst, do the arithmetic before you build a process on it.
· Reading taken October 1, 2026 from “Microsoft Copilot Researcher agent frequently asked questions,” stamp 2026-08-18
- Maximum 25 Researcher queries per user per month
- Included in the core experience for licensed Copilot users
- No image input, and no Copilot Studio customization
- Claude can be enabled in Researcher
- PowerPoint and PDF output listed as still to come
The /app skill and app building Sept 10
App generation is a spend event, not a free feature. Start there. A new agent surface arrived on September 10, 2026: the /app skill builds a working app scaffold from a plain description. In Cowork it comes through the Microsoft Frontier program. In Copilot Studio it entered public preview, described as rolling out over the following week. Admins get centralized visibility in the Microsoft 365 admin center, and published apps live at managedapps.cloud.microsoft.com. It’s billed on the usage-based model with Copilot Credits. The Cowork what’s-new September section corroborates it as “App skill (Frontier).” One loose end, still open. The stated one-week Studio rollout window closed on September 17, and 14 days later Microsoft has published no confirmation. The Cowork what’s-new page still lists only “App skill (Frontier),” and the Copilot Studio release-wave plan carries no app-skill entry. So verify availability in your own tenant.
- Source: “Build business apps with Copilot Cowork and Copilot Studio,” September 10, 2026, by Ryan Cunningham, CVP for Copilot Studio and Power Platform
/appgenerates a working app scaffold from a description- Cowork through the Frontier program; Copilot Studio in public preview
- Centralized admin visibility; published apps at managedapps.cloud.microsoft.com
- Billed with Copilot Credits
- The stated window closed September 17; still unconfirmed 14 days later
Brand Kit in PowerPoint Sept 14
Brand Kit covers templates, imagery, icons, fonts and other assets, and adds a stricter mode where Copilot follows approved layouts aligned with company standards. But the governance half is the part to plan for. Admins can publish approved Skills to specific users, groups or the entire organization. That turns deck production standards into something you distribute centrally instead of training into people. It’s generally available with a Microsoft 365 Copilot license.
- Source: “Brand Kit and Skills in Copilot in PowerPoint,” September 14, 2026, by Derek Johnson
- Templates, imagery, icons, fonts and other assets
- A stricter mode that follows approved layouts
- Admins publish approved Skills to users, groups or the whole organization
- GA with a Microsoft 365 Copilot license
Domain Exclusion is back Sept 9
This belongs on the governance checklist, not the feature list. Domain Exclusion is one of the few controls that changes what Copilot may read from the open web, as opposed to what it may read from your tenant. It had been withdrawn. On September 9, 2026 Microsoft restored it, described as available again. So if your web-grounding policy was written during the withdrawal, it was written without this lever. Worth reopening.
- Source: “Update: Domain Exclusion for Microsoft 365 Copilot,” September 9, 2026
- Admin control over Copilot web grounding
- Withdrawn earlier, now available again
- Governs open-web reading, not tenant content
- Re-check any policy written while it was unavailable
New connectors August 31 roundup
The August 2026 roundup names new connectors including Harvey, iManage Work, Mercury, Xero, Statista and ZoomInfo. Harvey and iManage Work are the two to notice if you run a legal or professional-services tenant. They put matter and document systems into the grounding set. That changes what a Copilot answer can be built from, and therefore what your permission review has to cover.
- Harvey, iManage Work, Mercury, Xero, Statista, ZoomInfo
- Legal and finance systems entering the grounding set
- Every new connector widens the permission review, not just the answer
Local browser use reaches GA August 28
An agent that drives a browser on somebody’s machine is a different risk surface from one that reads the graph. Review it before it becomes a habit. The August 28, 2026 Cowork what’s-new entry moves local browser use from Frontier to general availability for all Microsoft 365 Copilot tenants. It requires Edge. The same entry adds workspace file input for plugin tools, plus event-driven tasks that fire on a matching email or Teams message, @mentions included.
- Local browser use GA for all Microsoft 365 Copilot tenants
- Requires Edge
- Workspace file input for plugin tools
- Event-driven tasks on a matching email, Teams message or @mention
Visual creation in flow
Cowork can make deck graphics, document illustrations and email imagery without shunting the user into a separate creative app. The August 25, 2026 release notes add image generation in Cowork via ChatGPT Images 2.0. And Brand Kit now ships in PowerPoint. So the approved-asset path and the generated-asset path meet inside one deliverable, which means your brand review has to cover both.
- Image generation in Cowork via ChatGPT Images 2.0
- Integrated into deliverables
- Brand Kit and generated assets now coexist in one deck
- Still requires brand and truth review
Consumer-side retirements Underway
These retirements get almost no coverage, and they land on the consumer side, where your employees notice first. Group Chat: account updates began on August 18, 2026 and roll out gradually, so that one has come due. Mico, the animated character in Copilot Voice, is retiring in waves that started in August, with no fixed end date. Copilot Labs is retiring as part of a Copilot update, with timing varying by account and no published date. All of it comes from Microsoft’s support FAQ on retiring Copilot features. That page shows no date on its face, but its metadata carries a document date of August 10, 2026 and an update of August 18, 2026, which makes it 44 days stale. None of the three entries has moved since. So confirm directly before you build a communication plan on any of them.
- Group Chat: account updates began August 18, 2026
- Mico retiring in waves from August, no fixed date
- Copilot Labs retiring with timing varying by account
- The FAQ shows no visible date; its metadata says August 18, 2026, now 44 days old
The context layer is the real product advantage
Stop asking what the model knows. Work IQ connects the people, content and activity already sitting in Microsoft 365. So the question changes, right? It becomes: what can the system retrieve, permission and use for this one employee?
Organizational relationships
Who owns the work, who attended, who collaborates and which roles matter to the task.
Files and communications
Documents, presentations, spreadsheets, emails, chats and meeting artifacts. The August 25, 2026 release notes add Viva Engage private content as a Copilot grounding source.
What changed and when
Recent edits, decisions, conversations and project movement provide operational context.
The security boundary
Copilot should surface only the information the requesting user is already entitled to access.
What Copilot may read outside the tenant
Domain Exclusion, the admin control over Copilot web grounding, came back on September 9, 2026 after being withdrawn. It’s the lever that limits open-web sources. Worth revisiting if your policy was written while it was gone.
Plugins and Dynamics
Fabric, Dynamics 365 and partner plugins extend work beyond the Microsoft content graph. New connectors in the August 2026 roundup include Harvey, iManage Work, Mercury, Xero, Statista and ZoomInfo.
Skills and instructions
Custom skills encode the process, rules and output expectations the organization wants repeated. From September 14, 2026 admins can publish approved PowerPoint Skills to specific users, groups or the entire organization.
Brand Kit and asset libraries
Brand Kit in Copilot in PowerPoint covers templates, imagery, icons, fonts and other assets, with a stricter mode in which Copilot follows approved layouts aligned with company standards.
Work IQ REST endpoint
The August 25, 2026 release notes add a unified Work IQ REST endpoint for agents and workflows. That turns the context layer into something an in-house agent can call directly, rather than only something Copilot consumes.
Using the launch brief, the most recent steering meeting, the finance workbook and the project channel, build the leadership decision pack. Reconcile conflicting dates, name missing owners and end with the decisions required this week.
From today’s meeting, update the project plan, draft the stakeholder email and create the five-slide progress deck using our approved template. Do not send or publish anything until I approve.
Cowork changes Copilot from assistant to operator
The value isn’t a longer answer. Cowork takes responsibility for the sequence of work, right? Planning, tool use, file creation, delivery. Meanwhile the organization keeps policy and cost control.
Skills
Custom instructions that make Cowork follow the organization’s preferred process and output standards.
- Create and edit in the Customize area
- Combine personal, partner and organization-wide skills
- Use for recurring work
The /app skill Sept 10
The /app skill builds a working app scaffold from a plain description. In Cowork it comes through the Microsoft Frontier program. In Copilot Studio it was announced into public preview on September 10, 2026, and that preview is still unconfirmed on any Microsoft page. Published apps appear at managedapps.cloud.microsoft.com, with centralized visibility in the Microsoft 365 admin center. And generation is billed with Copilot Credits.
- Scaffolds an app from a description
- Frontier in Cowork; the Copilot Studio preview is still unconfirmed
- Centralized admin visibility
- Consumes Copilot Credits
Plugins
Wire Cowork into partner data and business systems without hard-coding every integration into the prompt. The August 28, 2026 what’s-new entry adds workspace file input for plugin tools. The September entry adds plugins on the mobile app, discoverable through the attach menu under Skills. And from September 30, 2026 there’s a plugin registry: one place where admins review Microsoft, custom-built and partner plugins, enable or disable them, and block or assign them to users or groups. That’s where plugin policy lives now.
- Dynamics and Fabric support
- A plugin registry spanning Home, Code, Autopilot and the Office apps
- Enable, disable, block or assign per user or group
- Over 100 plugins at launch
- Admin-managed availability
Automations Renamed Aug 31
The Scheduled tab is now called Automations, per the August 2026 Copilot roundup. Same idea as always: work that runs over time, which makes the home page a task-management view rather than a chat history. Update your internal training. The old label is in every screenshot you own.
- Review in-flight work
- Adjust automated tasks
- Keep ownership visible
Event-driven tasks Restated Aug 28
Cowork tasks can fire on an event rather than a clock. A matching email arriving. A Teams message. An @mention. The August 28, 2026 what’s-new entry restates this alongside the new capabilities. So Cowork stops being something a person starts and becomes something the work starts.
- Trigger on a matching email
- Trigger on a Teams message or @mention
- Define the stop condition before you enable it
Local browser use GA Aug 28
Local browser use moved from Frontier to general availability for all Microsoft 365 Copilot tenants on August 28, 2026. It requires Edge. An agent driving a real browser on a real machine is a very different risk surface from one reading the graph. So decide the policy before the capability shows up by default.
- GA for all Microsoft 365 Copilot tenants
- Requires Edge
- Review what a browsing agent may sign in to
Cost management
Usage-based execution needs budgets, monitoring, and a named answer to who gets to spend organizational agent capacity. For enterprise customers the metered services stay off until an admin creates a spending policy in the Microsoft 365 admin center, where budgets and alerts can be set at tenant, group and user level. The /cost skill reports the percentage of the monthly credit limit remaining. Capacity is bought as Copilot Credits: a $200.00 pack of 25,000 credits, described as per tenant per month. That’s about $0.008 a credit by division, though Microsoft publishes no per-credit rate. Commit Units take up to 20 percent off, and an Azure meter covers overflow automatically when prepaid credits run out.
- Write the spending policy before the pilot, not after the invoice
- Budgets and alerts at tenant, group and user level
- $200.00 for 25,000 Copilot Credits per tenant per month, about $0.008 a credit
- Up to 20 percent off with Commit Units
- Cowork per-task rates are still unpublished
DLP controls
Purview can restrict sensitive information and email behavior inside Cowork workflows.
- Policy follows the work
- Protect external actions
- Test real edge cases
Completed deliverables
The success metric is a usable file, update or outcome. Not the sophistication of the plan. Microsoft published its own framing of this on September 9, 2026, under the title “Measuring the value of Cowork: From AI interactions to completed work.” That’s the right unit of measurement, even before the post body is readable.
- Inspect the result
- Require completion evidence
- Count completed work, not interactions
- Keep human sign-off where needed
The interface disappears into the applications people already use
Copilot gets more valuable the moment the user stops exporting context, explaining the file structure and translating the output into whatever format the app wants.
Draft, edit and reshape
Create and revise documents in context, with tighter control over structure, tone and targeted edits. Word for Web gained an “Enable Sonnet 5 in models menu” setting in the August 25, 2026 release notes. Word is also one of the three apps where SpaceXAI’s Grok became selectable on September 12, behind an admin setting that’s disabled by default.
- Use existing templates
- Anthropic model selection since August 11
- Grok available via the Frontier Program, off by default
- Review tracked changes
Analyze with skills and rules
Use agent-style assistance, personalization and explicit rules for spreadsheet work. The August 25, 2026 release notes add Python in Edit with Copilot for Excel. And Excel is a Grok surface as of September 12, 2026.
- Inspect formulas
- Model scenarios
- Grok selectable through the Frontier Program
- Explain assumptions
Build on Brand Kit
Brand Kit went generally available on September 14, 2026 with a Microsoft 365 Copilot license. It covers templates, imagery, icons, fonts and other assets, with a stricter mode where Copilot follows approved layouts aligned with company standards. PowerPoint is also a Grok surface.
- Approved layouts and assets
- Admin-published Skills
- Source-aware narrative
- Generated visuals in flow
Move from inbox to action
Customize and refine responses using the surrounding thread, calendar and work context.
- Draft in the user’s style
- Summarize decisions
- Confirm before sending
Capture meeting reality
Use transcripts, chat and meeting artifacts to create summaries, actions and follow-up materials.
- Name owners
- Resolve ambiguous commitments
- Link back to the discussion
Curate a working context
Bring a deliberate set of references together for a sustained piece of research or planning. The August 2026 roundup adds a dual OneNote experience for Copilot Notebooks.
- Control the source set
- Add references over time
- Keep citations attached
Turn language into accountable work
Planner agents can help translate decisions into tasks, ownership and progress.
- Set dates and owners
- Review dependencies
- Avoid silent task creation
Measure and manage
Adoption, policy, cost and agent activity become part of the deployment, not an afterthought. Model access for Anthropic and SpaceXAI is set here, and the SpaceXAI setting ships disabled.
- Usage and value signals
- Model and plugin policy
- Risk and spend controls
A third vendor lands in the Office apps, not in Cowork
Grok from SpaceXAI became available in Word, Excel and PowerPoint through the Microsoft Frontier Program. It sits behind a dedicated administrative setting that’s disabled by default, and it isn’t available to Frontier customers in the EU, EFTA or the UK during the preview. It isn’t in Cowork and not in Copilot Studio. So model choice in the Office apps is now a three-vendor question.
- Word, Excel and PowerPoint only
- Admin setting disabled by default
- EU, EFTA and UK excluded during the preview
- No published retention terms
Unified chat and a Work IQ button
The August 25, 2026 batch unifies web and work chat behind one surface with a dedicated Work IQ button, and adds model selection for Researcher in Chat. The button is the interesting part. It makes the grounding decision explicit, instead of leaving it to a mode the user never noticed they were in.
- Unified web and work chat
- Dedicated Work IQ button
- Model selection for Researcher in Chat
Viva Engage becomes grounding
Viva Engage private content can now ground Copilot answers. If you run private Engage communities as a knowledge channel, that widens what a Copilot answer can be built from. Which means it widens what your permission review has to cover.
- Private Engage content in the grounding set
- Check community membership as an access control
- Re-run the oversharing review after enabling it
A Work IQ endpoint for your own agents
A unified Work IQ REST endpoint for agents and workflows moves the context layer from a Copilot feature to a platform your own team can call. That’s the difference between buying Copilot and building on Work IQ.
- REST access for agents and workflows
- Plan authentication and scope before the first prototype
- Governance follows the endpoint, not the app
A dashboard for credit consumption
The Consumption Dashboard tracks Copilot credit usage for Cowork and WorkIQ API services. If you’re running metered Cowork, this is the surface that tells you the cost before the invoice does. The /cost skill adds a percentage-remaining view inside the session itself. A second dashboard landed in the September 23, 2026 release-notes batch: a GitHub Copilot AI cost dashboard in Insights, showing GitHub Copilot AI consumption, cost and adoption metrics across the organization. Neither surface publishes a per-task Cowork rate.
- Covers Cowork and WorkIQ API services
- GitHub Copilot AI cost dashboard in Insights, September 23, 2026
- Pair both with admin budgets
- Cowork per-task rates still unpublished
- Measure cost against completed work
MCP agents reach the Office apps
Agents built on the Model Context Protocol are available inside Word, Excel, PowerPoint and Outlook. That’s the point where an external tool stops being a Chat sidebar and starts working on the document in front of the user.
- Word, Excel, PowerPoint and Outlook
- Scope tool access per agent
- Review what the agent can change, not only what it can read
Audit this workbook, correct only verified formula issues, run the agreed sensitivities, then create the executive deck using our Brand Kit template. Every chart must link back to the workbook source.
Turn this meeting into an action plan. Distinguish decisions, proposed actions and unresolved discussion. Create tasks only for explicit commitments and ask me to approve owners and dates.
Review this thread and the related project context. Draft a concise response in my normal style, identify the decision being requested and list any commitments the draft would create. Do not send.
Review this document against our policy and the cited source materials. Suggest changes with reasons, preserve the author’s voice and flag every statement the evidence cannot support.
Eight habits for making Copilot an operating layer, not another sidebar
The value shows up when teams change three habits. How they point at context. How they hand work over. And how they standardize the things they do over and over.
Reference, do not retype
Point to the authoritative meeting, file, person or project context.
Choose Chat or Cowork
Chat when you want an answer. Cowork when the job is a completed deliverable and you accept a metered bill for it.
Let Auto route inside Cowork
Auto routing is still Cowork-only. Model choice isn’t. Word gained Anthropic model selection on August 11. Researcher in Chat and Word for Web followed on August 25. Grok reached Word, Excel and PowerPoint on September 12, behind an admin setting that’s off by default. And Microsoft announced four more models on September 22 and September 30 that neither the Cowork nor the Copilot Studio model page lists yet. Everywhere else, pick on purpose, and only when the task actually benefits.
Finish in the native app
Ask for the Word file, Excel model, PowerPoint deck or Planner update you’ll use.
Preserve provenance
Keep source links, workbook references and meeting context attached to claims.
Standardize successful work
Turn repeated prompts into governed skills, plugins or agent templates. Then publish the approved ones centrally instead of training them into people.
Treat permissions as quality
Too little access and too much access both wreck grounding. One starves it. The other leaks.
Measure outcomes, and watch the meter
Track completed work, decision quality and time saved, not chat volume. Then put a second number beside it, because Cowork, Code and Autopilot spend Copilot Credits per task. Write the spending policy before the pilot, not after the first invoice.
The context advantage becomes a liability without information hygiene
Copilot inherits your Microsoft 365 environment, warts and all. So the AI experience is only ever as good as the housekeeping, right? Permissions, labels, stale files, duplicate sources and ownership decide it. So does knowing which Microsoft page to believe, and which of three model vendors is processing the request.
Permission sprawl
Copilot makes overshared information easier to find. Fix access before you celebrate retrieval. Every connector added in the August 2026 roundup widens the surface this applies to, and so does Viva Engage private content as a grounding source.
Stale authority
Several plausible versions of the same file, and Copilot grounds itself confidently in the wrong one.
Model availability
Don’t treat the documented lineup as the current one. The Cowork model page, stamped 2026-09-14, reads Auto, GPT 5.5 (Frontier), GPT 5.6 Sol, GPT 5.6 Terra, GPT 6 Astra, Opus 5, Claude Sonnet 5 and Claude Fable 5.1. Four models have been announced into Copilot model choice since: Claude Opus 5.5 and GPT-6 Sol on September 22, GPT-6.1 Sol and Claude Sonnet 5.5 on September 30, by Microsoft’s dates. None of them is on that page. Foundry lists Opus 5.5 and Sonnet 5.5, but Foundry isn’t the Copilot picker. The Cowork FAQ, stamped 2026-09-29, adds Claude Fable 5 (Preview) as an admin-gated choice the model page leaves out. The Copilot Studio model page, stamped 2026-09-18, lists no GPT-6 and no Claude 5 at all. Claude, OpenAI, SpaceXAI and Frontier capabilities all depend on plan, rollout stage and administrator policy. So read the picker first. Then the dated announcement. Then the Learn page. And check the stamp on all three.
Retention, and which regime your tenant gets
Rebuild any note that says the Fable line is the retained line, full stop. The answer depends on the route you buy through. Microsoft’s subprocessor page, stamped 2026-09-18, gates Fable 5.0. Extra admin-center controls. Off by default for all scenarios. No user access until a tenant admin explicitly opts in. For Fable 5.1 it says something else. For certain organizations the model is available with Anthropic acting as a subprocessor, under Microsoft’s Product Terms and Data Protection Addendum. On that path, Anthropic doesn’t retain customer content. Where retention does apply, Microsoft says Anthropic stores most inputs and outputs for up to 30 days. Anthropic’s own wording is at least 30 days. Through Microsoft, Microsoft’s wording is the one in your contract, so get that ceiling confirmed in writing. It may retain content for up to two years on a usage-policy classifier hit. And it may keep classification scores for up to seven years. Now the gap. Neither Microsoft page maps a tenant to a regime, and none of it describes the direct route. Bought from Anthropic rather than through Microsoft, the same Fable 5.1 carries 30-day retention. And Anthropic’s own pages split on zero data retention. Its September 1 API release note says only with express authorization. Its September 1 news pages promise it to eligible customers until Enterprise Frontier Safeguards ships. So the answer comes from your contract. One addition for the register: that same Microsoft page now names Claude Mythos 5 alongside Claude Fable 5 as an advanced model requiring data retention by Anthropic.
A third subprocessor
Your subprocessor register needs a third row. SpaceXAI is now a model vendor inside Copilot, in Word, Excel and PowerPoint through the Frontier Program. The access setting is dedicated and disabled by default, and Grok isn’t available to Frontier customers in the EU, EFTA or the UK during the preview. Microsoft has published no retention terms for SpaceXAI models. Until it does, flipping that setting is a data protection decision, not a productivity one.
Web grounding controls
Domain Exclusion for Microsoft 365 Copilot came back on September 9, 2026 after being withdrawn. It’s the admin control over what Copilot may ground on from the open web. So reopen any web-grounding policy written during the withdrawal. It was written without this lever.
Agent 365 licensing, now documented
The Microsoft Agent 365 service description, stamped 2026-09-15, settles the question two older pages used to muddle. Agent 365 is a standalone subscription usable with eligible Microsoft 365 subscriptions, and it is also included with Microsoft 365 E7. Basic registry, governance actions, lifecycle rules and external sync come with eligible Microsoft 365 plans. Ten more features need E7, G7 or the standalone license. They’re listed in the Agent 365 card under Retention and Licensing. Separately, and this is the one to act on, agent security for Copilot Studio and Foundry agents has required an Agent 365 license since July 1, 2026. Tenants without one lost agent discovery, threat detection and Advanced Hunting investigation that day, and existing Block rules stopped blocking. Check which side of that line your tenant is on.
Plugin governance has a home
The plugin registry, introduced September 30, 2026, is where plugin policy now lives: one place to discover, publish, manage and govern plugins across Copilot Home, Code, Autopilot, Word, Excel, PowerPoint and SharePoint, with Copilot Studio, GitHub Copilot and Microsoft Foundry to follow. Admins can enable or disable plugins and block or assign them to users or groups. Over 100 plugins shipped at launch. Every one of them is another path into your grounding set, so review the registry the way you review a connector list.
Where Cowork can actually run
Three facts, three different places to check before a global rollout. Cowork use and access is limited to the regions Anthropic supports, because Anthropic models sit under it. Students aren’t eligible; faculty and staff in education are. And the Anthropic subprocessor setting in the Microsoft 365 admin center, under Copilot, Settings, View All, is on by default for EU, EFTA and UK tenants created after March 25, 2026, with older tenants pointed at Message Center. The independent processor route was decommissioned on May 1, 2026, so Anthropic as a Microsoft subprocessor is the only path to Anthropic models.
Usage-based cost
Cowork creates value and variable spend at the same time, and the metered list keeps growing: Cowork, Code, Autopilot, advanced work in SharePoint and OneDrive, Work IQ APIs, Copilot Managed Runtime and Teams Phone Agent. Enterprise tenants get a catch. Those services stay off until an admin creates a spending policy in the Microsoft 365 admin center, where budgets and alerts are set at tenant, group and user level. Copilot Business bought through a CSP partner does not get that catch: from November 2, 2026, new Copilot Business licenses purchased through CSP include usage-based billing by default, with a preset monthly limit. Capacity is bought as Copilot Credits at $200.00 for a 25,000-credit pack per tenant per month. That’s about $0.008 a credit, though Microsoft publishes no per-credit rate. Commit Units take up to 20 percent off, and an Azure meter picks up overflow. Two gaps remain. Per-task Cowork and Work IQ consumption rates are unpublished. And reasoning effort changes consumption quietly, because higher effort uses more credits.
External actions
Sending, publishing, changing records, contacting customers. All of that needs explicit authority and review. And local browser use reached GA on August 28, 2026, which stretches this from your records to anything a browser session can reach.
DLP edge cases
Policy must be tested against real agent workflows, not only ordinary document sharing.
Generated brand risk
Brand Kit and approved layouts cut visual drift. They do nothing for facts or narrative. A deck can be perfectly on brand and completely wrong.
Agent and app proliferation
Skills, plugins, agents and now generated apps all need ownership, lifecycle management and retirement. Creation is the easy part. The /app skill drops the cost of making an app to one sentence, which raises the cost of governing what exists.
| Control layer | Question to answer | Useful evidence |
|---|---|---|
| Identity | Who is asking and what can they access? | Effective permission review |
| Information | Which source is authoritative and current? | Labels, ownership and version policy |
| Model | Which models may process this task? | Admin configuration and the live picker first. Then the dated announcement. Then the Learn page and its stamp. Remember that the Cowork model page and the Copilot Studio model page both run behind the announcements |
| Subprocessor | Which vendors process our content, and under what terms? | Anthropic, OpenAI and SpaceXAI are all in scope; the SpaceXAI setting is off by default and carries no published retention terms; the Anthropic subprocessor setting is on by default for EU, EFTA and UK tenants created after March 25, 2026 |
| Retention | Does this model retain our prompts and responses? | Subprocessor page, stamped 2026-09-18: Fable 5.0 gated, Fable 5.1 terms vary by organization, Mythos 5 retains. Up to 30 days in Microsoft’s wording, at least 30 in Anthropic’s; two years on a classifier hit; seven years for scores |
| Commercial | Do we hold the prerequisite licenses? | Agent 365 service description, stamped 2026-09-15: standalone or included with E7; ten governance features need E7, G7 or standalone. Agent security for Copilot Studio and Foundry agents needs Agent 365 since July 1, 2026 |
| Cost | What will execution consume? | Copilot Credits at $200.00 per 25,000-credit pack per tenant per month, about $0.008 a credit; Commit Units; the Consumption Dashboard; effort level per task; CSP Copilot Business billing default-on from November 2, 2026 |
| Agent | What can it change and spend? | Tool scope, budgets, action logs and whether local browser use is enabled |
| Outcome | Did the work create measurable value? | Completed artifact, cycle time and quality review |
Where these facts come from
Every item below is a dated first-party article. Microsoft’s come first, then the Anthropic, OpenAI and xAI articles that date the models Microsoft resells. Microsoft’s reference documentation is a different animal. Those pages are rewritten in place and carry a last-updated stamp rather than a publication date. So they’re attributed inline with their stamp instead of listed as sources, and you should confirm them directly. The callouts below name exactly which claims rest on that kind of page. And where Microsoft’s own pages disagree with each other, both sides are cited, with a note of which one is older.
AI Mindset