August 2026 Microsoft AI roundup — key updates & concepts

1) New “Technology in 5 Minutes” educational series (tutorial/news-style)

A new YouTube series is introduced to explain AI concepts quickly:

• What AI is, and “big language models” / generative AI
• Tokens and “tokenomics”
• What a “harness” is
• Why AI hallucinates (previewing later explanations via model behaviors)
• Perception project: using AI defensively against malicious uses of AI

---

2) Model releases & improvements (Foundry / GitHub Copilot / platforms)

Microsoft AI / coding & multimodal

• AI coding model v1.1 on GitHub Copilot

  • Lower cost + higher quality
  • 25% fewer tokens, 25% faster token streaming
  • ~4x cheaper than the previous version
  • Presented as an iterative “release → learn → improve” cycle

• Image-capable model update

  • Better results with text, portraits, and 3D images
  • Claimed to be 2nd only to “GPT Image 2”, outperforming other image models

• MAI Thinking 1 (reasoning model) on Foundry

  • Medium-sized, “clean, traceable data”
  • No distillation from another model (as stated)
  • 35B active parameters / 1T total
  • MoE (mixture of experts)
  • Similar performance claim to Claude Opus 4.6
  • ~250k token context window

• MAI Transcribe 2 (speech-to-text)

  • 10x faster than “GPT Transcribe”

  • 1 on accuracy + latency “Pareto frontier”

  • Lowest price claim; positioned as top automatic speech recognition
  • Available on Foundry; also testable in playground.microsoft.ai

• Gemini 3.7 Flash added to GitHub Copilot

  • Web-development focus: agent coding, deeper codebase exploration, verification

• Kimi K3 (open-source coding model, Moonshot AI) added to GitHub Copilot

• Additional model availability via Foundry, emphasizing model selection

Long-context / agent + automation oriented models (via Fireworks / Foundry / Azure)

• DeepSeek

  • Good for agent coding, workflow automation, general agent assistants

• Flash 0731 / Nemotron

  • Similar capabilities, with long context, multi-step execution, multilingual support

• Grok 4.6 (SpaceX AI) on Foundational

  • ~500k context
  • Adjust reasoning level
  • For long-running agents; improved interactive/visual experience
  • Knowledge limited to Jan 2026

• Fable 5.1 (Anthropic) in M365 Copilot / GitHub Copilot / Copilot Studio / Foundational

  • Progress in long-term coding, multi-stage research, working with documents/spreadsheets/slides
  • Cheaper caching rates
  • Behavior for ambiguous tasks: better judgment, less confident but more likely to avoid incorrectness
  • If stuck, it says it’s stuck (explicitly tied to “why AI hallucinates”)

• Mythos 5.1 companion

  • Same model family positioning as Fable, but adapted (“weakened fuses”) for cyber/bio research
  • Still in restricted access program

• “GPT-6 Astra”

  • Presented as a higher tier than “Sol” in an OpenAI-styled naming ladder
  • Open-ended challenge thinking, planning, workflow execution across apps/systems
  • High-performance computing; global deployment with US data zone options

---

3) Foundry platform updates (routing, understanding, libraries, semantics, governance)

• Router model updates

  • Router automatically selects the best model based on complexity, context, tool usage, etc. to optimize cost
  • Now available in 32 regions
  • Supports global use and data-zone constrained operation
  • Adds models (including names like “GPT-56 Soul Tarot” and Luna) alongside Claude Opus 4.8

• Content understanding improvements

  • Handles analysis across images, documents, forms, audio, video
  • Broad support for GPT 5 family
  • Improved “justification” and confidence scoring
  • Goal: more reliable extraction at lower cost

• Client libraries

  • “2.0 preview” introduces a “semantic partitioning suite”
  • Mentions: synchronous reading/markup APIs, improved classifications, agent-based document reasoning

• Claude model hosting shift

  • Claude originally hosted on Anthropic; now hosted on Azure
  • Expanded functionality includes:
    • Structured source data to enforce output conforming to a specified schema (reduces formatting errors)
    • Web search that returns quotes without requiring the user’s own crawler (to help avoid hallucinations)
    • Web scraping by URL
    • MCP connector to avoid running a separate MCP client (point to remote MCP server)
    • “Search for tools” behavior: Claude can load only needed tools instead of overloading context

---

4) Copilot Studio & “harness” for agent workflows

• Copilot Studio GA support for GitHub Copilot harness

  • Positioned as better for long-running business processes requiring intensive reasoning
  • Still supports standard harness types:
    • rules-based agents harness
    • chat harness for extending M365 Copilot

• Workflow builder upgrades

  • Visual designer + AI actions
  • Agent-to-agent handoffs
  • Connectors and steps requiring human intervention
  • Node-level testing
  • Focus: tying adaptive agent behavior to clearly defined process steps

• Supports delegating to specialized agents, workflows, MCP services, tools

• Mentions use of Windows 365 Cloud PC as a desktop within workflows

---

5) Microsoft 365 Copilot features (knowledge trust, control, cost clarity)

• Authoritative SharePoint sites

  • Admins can mark up to 100 sites as “authoritative” so Copilot treats them as official sources (HR, corporate news, etc.)
  • Improves response trustworthiness and signals authority to users

• Homework / “thinking effort” control

  • Customizable effort level (easy → maximum with increments)
  • Gives more control over how hard the AI “thinks”

• Expenses credit clarity

  • Shows monthly credit balance and reset timing, plus current-session usage

• Chat UX/iteration

  • Renamed “Scheduled” → Automations
  • Ability to select a portion of a response and ask Copilot to work only on that context
  • A “try again” button (circle icon)

• Copilot notebooks

  • Can be used either in Copilot or switched to OneNote workspace without losing context

---

6) App-level updates (Word / PowerPoint / Excel / SharePoint / Planner / OneDrive)

• Word Copilot

  • Adds hyperlinks
  • Understands images inside documents
  • Highlights the specific word rather than the whole modified section

• PowerPoint Copilot

  • Assigns tasks through comments
  • Enforces corporate identity (colors/layouts) to avoid creating new layouts
  • Supports custom skills and action control via notes in a special area

• Excel Copilot

  • Describes changes using workbook history
  • Saves past conversations so users can resume

• SharePoint skills

  • SharePoint supports skills usable in OneDrive
  • Skills saved as Markdown files

• Planner integration

  • Copilot can create tasks and request information

• GitHub Copilot credit usage

  • Shown in an “Agent 365 dashboard” with paid credits, active users, adoption trends, etc.

---

7) Agent & automation capabilities (Scout, Copilot automations, Azure Copilot agents)

• Scout autopilot available in “Frontier preview”

  • Acts on the user’s behalf
  • Can check system status and run on schedules/events/conditions or via natural language query

• GitHub Copilot public code review

  • New risk-based setting: light / balanced / maximum
    • “Light” for simple PRs (replacing earlier “low”)
    • “Balanced” for deeper analysis (replacing earlier “medium”)
    • “Maximum” for thorough but more costly review
  • Now also covers merge requests from bots and very large requests

• Cloud agent delegation controls

  • Configurable “level of consideration” when delegating to control thought process and therefore credit usage

• Corporate governance

  • Admin control of MCP servers: allow/deny in supported Copilot clients
  • Policies cover app, CLI, and VS Code; unverified configs blocked
  • Global model enablement policy:
    • Unconfigured public models inherit corporate defaults
    • Models requiring “scales to be kept open” remain excluded from automatic activation (manual enablement required)

• Copilot automations via comments

  • Use a comment text as the trigger to run automations (e.g., create/update documents, investigate errors, generate follow-ups)
  • Can be launched from Copilot app and GitHub CLI

• Multi-session management

  • Better support for switching sessions between GitHub Copilot app and GitHub CLI; improved simultaneous-session handling

• Usage/reporting

  • “Usage Metrics API” separates third-party agent activity, per-agent task launches, and aggregate sessions
  • Reports input/output and cached token usage per model and credits (aimed at explaining costs, not individual employee performance)
  • Impact dashboard adds ROI section:
    • Estimates potential ROI based on depth of Copilot adoption (chat/code additions vs agents)
    • Includes salary selector and payroll assumptions for modeling returns

• Azure Copilot

  • Now exposes specific agents directly (not only chat)
  • Agents for troubleshooting, deployment, optimization, resiliency, etc., for faster specialized help

---

Main speakers / sources mentioned

• Narrator/speaker: “I” / the video host (no personal name given in subtitles)

• Microsoft AI / product sources referenced:

  • Microsoft (Foundry, GitHub Copilot, M365 Copilot, Copilot Studio)

• Other organizations/models referenced:

  • OpenAI (described via GPT-6 Astra naming ladder)
  • Anthropic (Claude Opus 4.6/4.8, Fable 5.1, Mythos 5.1)
  • Google (Gemini 3.7 Flash)
  • Moonshot AI (Kimi K3)
  • DeepSeek (DeepSeek model)
  • SpaceX AI (Grok 4.6)
  • Fireworks (model serving)
  • Hugging Face (Nemotron managed computing)
  • MCP (Model Context Protocol ecosystem; referenced via connectors/policies)