Microsoft AI Update September 2026

Summary of the “Microsoft AI Update September 2026” Subtitles (Tech + Product Features)

Overall themes (what’s new this month)

• The update introduces jump-to sections and highlights new AI-related videos, including:
  • One on why AI hallucinates
  • One focused on how organizations should approach and use AI to engage customers, employees, and management
• A recurring focus is faster model iteration, paired with the need for processes to evaluate, route, safeguard, and trust model behavior as capabilities change quickly.

---

1) New / Updated AI models and where they’re available

Anthropic “Fable 5.1”

Availability

• M365 Copilot
• GitHub Copilot
• Copilot Studio
• Foundry

Improvements

• Better long-term programming
• Stronger multi-step research
• Works well with documents, spreadsheets, presentations

Quality / behavior

• More favorable performance vs the previous version
• Better judgment in ambiguous tasks
• Fewer confident-but-wrong answers
• If it’s stuck, it says it can’t find a solution rather than fabricating

Guidance

• Recommended for complex logical tasks and long-term agent work

---

Anthropic “Mythos 5.1”

• Same model family as Fable 5.1
• Relaxed restrictions for cyber and biological scenario categories
• Still described as limited within the program

---

OpenAI “GPT-6 Astra”

Positioning

• Described as “level above” Soul in a model hierarchy: Luna → Terra → Soul → Astra (levels mentioned)

Designed for

• Complex problem solving
• Planning
• Generating higher-quality results
• Workflow execution across apps / systems

Computer-use angle

• Useful when there’s no direct API or MCP
• Executes using the same desktop/browser/application as the user

Availability

• Foundry: Global and US data zones
• Also in M365 Copilot, Copilot Studio, GitHub Copilot
• Note: some services may be restricted by region, impacting EU / EFTA / UK availability

---

Gemini “3.8 Flash” (GitHub Copilot)

Availability

• GitHub Copilot

Highlights

• Good results for complex programming via terminal
• Rigorous testing, including recovery from failures requiring action

Licensing

• Available in Pro licenses and above

---

“SpaceX AI Grok” in M365 apps

What’s new

• Added as a subprocessor to M365 Word / Excel / PowerPoint agents

Compliance / security

• Runs in SpaceX AI data centers
• Still adheres to Microsoft security / privacy / data protection standards

Availability restriction

• Not available in EU, EFTA, UK

---

Model updates in Foundry and Copilot ecosystem: Sol and Luna (“6th wave”)

• After Astra, Sol and Luna were updated.

Availability notes

• Foundry: both
• GitHub Copilot: Sol included (Luna not mentioned as available)
• M365 Copilot + Copilot Studio: Sol available; Luna not

Positioning

• Sol: “more balanced” for interactive agent coding
• Luna: smaller / faster with lower cost
• Sol 6.1: further improves agent coding and “computer use,” advancing from “6”

Emphasis

• The pace of innovation requires rapid implementation processes

---

Claude “Opus” and “Sonnet” (versions 5.5)

• Opus 5.5
  • Good for long-term coding
  • Available in Foundry, data zones, globally
• Sonnet 5.5
  • More intelligence per fewer tokens
  • For clearly defined programming tasks
  • Available in Copilot Chat, Word, Excel, PowerPoint, Copilot Co-work, Copilot Studio, Foundry

---

DeepSeek “4.1 Flash”

Availability

• Foundry

Description

• “More refined” V4 Flash
• Mixture of experts with parameter sets depending on the request
• Strong with huge context

Delivery

• Via Fireworks in Foundry (GA)
• Directly through Azure (preview)

---

Microsoft AI Transcribes 2 (speech)

Availability

• Foundry

Claims

• 10x faster than GPT Transcribe
• 5x faster than Gemini 3.5
• #1 on analysis accuracy and latency (Pareto frontier)
• Lowest price positioning

Recommendation

• Try it in Foundry / the Microsoft AI site for automatic speech recognition

---

2) Mentioned new “purposeful decision” model (Jev)

• A new model from Type Space AI called Jev (not yet in the Microsoft ecosystem)

Concept

• Designed for purposeful decision-making, not just text generation
• Acts like a “judge” outputting probabilities / scores for programmatic use

Example behaviors

• Given an email + context, it can output probabilities such as:
  • Does it indicate customer dissatisfaction?
  • Should the system send an answer?
  • Which team should handle it?

Role in architecture

• Not a replacement for large general LLMs
• Intended as a supplement called from code to decide which tools to use

Strategic takeaway

• Microsoft emphasizes diversity in model/lab sources to avoid being tied to one innovation stream

---

3) Microsoft Foundry updates

Model Router updates

• Adds a single endpoint/API to route among models based on conditions like:
  • Complexity / reasoning needed
  • Context size
  • Tool usage
  • Compute resources
  • Latency requirements, etc.

New features

• Session affinity: keep the same matching/routing model across turns for conversational apps
• Retrieve routing metadata (routing modes, latency, attempts, errors, rollback mechanisms) to improve troubleshooting / observability

Availability

• Router now in 32 regions

---

Foundry Dev Pad

• Integrates developer tools for:
  • Terminal
  • VS Code
  • Coding agent workflows
• Goal: reduce setup complexity when moving from experiments to stable agent development

---

Voice agents (Foundry) — preview

Described as:

• Real-time voice agent capabilities
• Deployment workflows + observability + evaluation in one platform
• < 1 second latency
• Multilingual support
• Integrated knowledge tools and model selection
• Supports “talking heads” avatars (ready-made and custom)
• Lifecycle management: observe / track / evaluate / optimize

Challenges mentioned

• Recognition errors
• Interruptions
• Ordering of utterances

---

Publishing Foundry agents to M365 Copilot and Teams — GA

General availability

• Lets users place agents where they already work

Access control

• Control who can see and call agents (including both agent types: “tipped” and “posted” as mentioned)

Hosted agents: outbound traffic control

• Rules constrain communication targets (example: wildcard host)
• Motivation: responsible AI and preventing agent “runaway” behavior that could damage third-party sites
• Can operate in audit or enforcement modes

---

Foundry Procedures — preview

• Run published agents on a schedule or event basis
• Triggers may include:
  • Time / frequency
  • Teams messages
  • GitHub tasks
  • Required actions
• Advantage: keeps execution inside Foundry projects (no extra function / logic app / queue-style glue)

---

Agent Interaction Protocol v1.0 — GA

• Supports A2A (agent-to-agent communication)
• Agents can coordinate with other agents that support the protocol
• Foundry agents now appear in Agent 365 admin center
  • Lets org admins enable/disable them like other agent types

---

4) Copilot Studio updates (natural language app creation)

Natural language app creation — preview

• Users describe an app in conversation:
  • Data
  • Users
  • Actions
• Then test / refine / publish / manage it

Apps can use

• Connectors
• Work IQ
• Entra identification
• Client policy compliance

Runtime

• Runs in Copilot managed runtime
• Client-integrated via Entra authentication
• Includes versioning, uses its own Git, and provides lifecycle controls

Administration

• Apps show in the M365 admin panel for centralized access verification, monitoring, and health policy management
• Note: this is application hosting, not agent hosting

---

Copilot in SharePoint and OneDrive

• Copilot can search / summarize / compare / create content from cloud files directly

---

5) “Copilot app” UI/workflow and pricing model updates

Copilot app reorganized into three tabs

1. Home
   • Chat and Co-work (collaboration / delegation workspace)
2. Chat
   • Helps with tasks; tied to the user license (also mentioned for Copilot in Word/Excel/PowerPoint)
3. Co-work
   • For delegating more complex or complete results
   • Consumption-based pricing (pay for work performed)

Additional notes

• Microsoft says functionality will be added to the appropriate license tiers as economically feasible
• Some advanced features may have initial limits for smaller/advanced user groups
• An “intent” feature is “on the way” to route users to the right place (chat vs co-work vs code)

Code tab

• Build with natural language
• Runs in Copilot managed environment

Autopilot tab (formerly Scout)

• Delegate responsibilities to an ongoing agent
• Uses Copilot credits

Co-work file editing improvements

• Edit Word/Excel/PowerPoint files directly in OneDrive/SharePoint without creating extra copies
• Supports version history for collaboration and recovery
• Requires permissions; local device files are not supported

---

6) Fabric IQ in Copilot (data + business logic grounding)

• Fabric IQ concept: treat AI like an “intelligent intern” that needs tools and enterprise context
• Grounding mechanism:
  • Connects Copilot/Co-work chat to Power BI semantic models
  • Semantic models define enterprise objects and relationships
  • Uses OneLake for data virtualization to reference enterprise data
• Benefits:
  • Answers reuse shared business logic from reporting
  • Maintains existing access controls
• Framed as “state of the business” understanding

---

7) Branding kits / skills + plugin registry (Copilot ecosystem management)

PowerPoint branded kits and skills — GA

• Branded sets provide consistent org templates (fonts, colors, icons, etc.)
• Skills turn common workflows into repeatable actions that guide Copilot

Unified registry of plugins (agents / skills / connectors / MCP)

• A single registry for packaging and managing:
  • agents, skills, connectors, and MCP capabilities
• Applies across Copilot areas like Home, Code, Autopilot, M365 apps, and SharePoint

Work IQ developer tools

• Public preview team interface for importing/building/testing/packaging/sharing plugins

---

8) GitHub Copilot updates

“Hydra Fusion” — preview (research routing)

• Accessed via /experimental (slash experimental)
• Routes models balancing performance, cost, latency
• Three templates:
  • Singular: one model solves
  • Cascade: draft model + quality control/refinement by a stronger model
  • Reviewer/edit loop: one family drafts, another reviews, author model edits

Enterprise permission management for agent operations — GA

• Admins can prohibit / require approval / allow agent actions involving:
  • files, shells, and network
• “Managed rules” enforce corporate boundaries so user/workspace settings can’t be weakened

VS Code agent for code merging — preview

• Handles:
  • review feedback
  • failed reviews
  • merge conflicts
• Extends from code creation to pull-request-ready merges

Teams + Slack integration improvements

• Better context handling:
  • conversations, files, attachments can serve as context for GitHub Copilot actions

Cost/quality model auto-selection levels

• “Auto” mode includes levels:
  • efficient, balanced, intelligent
• Trade-off: cost vs quality vs latency
• Initially limited to some surfaces

Budget increase request flow — GA

• Previously: when credits ran out, users were blocked
• Now: request more budget, with payer approval (approve/modify/reject)

---

9) AI skills and data features (Azure SQL, PostgreSQL, MCP)

PostgreSQL and MCP skills — preview

• Skills help agents integrate with PostgreSQL
• MCP-related skills assist with getting started/configuring/troubleshooting SQL access
• Goal: improve SQL accuracy/usage in assistants like GitHub Copilot

SQL Agent skills for migration (mentioned)

• Supports SQL migration for:
  • Azure SQL DB
  • Azure SQL MI

Azure SQL Database vectors — GA

• Vector search via:
  • native vector indexes
  • vector search using Microsoft Research disk ANN
• Motivation: embeddings represent semantic meaning for similarity search (nearest neighbors)

---

10) ACA sandboxes — GA (agent isolation + outbound control)

• ACA sandboxes are now GA
• Architecture:
  • micro-VMs isolated at kernel level by hardware
  • separate outbound proxy for controlling external communications
• Purpose:
  • run Linux container images for agents needing strong management control
  • supports ephemeral use cases (launch < 1 second)
  • supports steady state via persistent storage (blobs/disks)
• Mentioned as similar technology to “Foundry, GitHub sandboxes, and others.”
• Includes portals depending on interaction type

---

Main speakers / sources mentioned

• Microsoft (speaker not named in subtitles) — “AI Update for September”
• Technology in Five Minutes (video series referenced)
• Antropic (Fable / Mythos models)
• OpenAI (GPT-6 Astra)
• Google (Gemini 3.8 Flash)
• SpaceX AI (Grok)
• Type Space AI (Jev model)
• GitHub / VS Code / Teams / Slack (integration context)
• Azure / Microsoft Foundry / Copilot / Fabric (Microsoft product suite)