Topic deskAI and data science news

Microsoft AI News

Microsoft's AI story: Copilot family launches across Office, GitHub, and Windows; Azure OpenAI and AI Foundry updates; Fortune 500 partnerships; and the strategy behind it.

Stories

637

Latest source update

July 30, 2026

Coverage

Live

Topic brief

What to know about Microsoft

Brief updated Aug 3, 2026

Microsoft is one of the largest distributors of AI capability to enterprises, combining its Azure cloud infrastructure, the Microsoft 365 Copilot productivity assistant, GitHub Copilot for developers, and a growing set of its own MAI models alongside partner models from OpenAI, Anthropic, Mistral AI and open-weight labs. For data scientists and ML engineers, Microsoft is significant less as a single model vendor and more as the platform layer where many organizations first encounter production AI, whether through Copilot in Word and Excel, agent orchestration in Copilot Studio, or model hosting in Microsoft Foundry.

Under CEO Satya Nadella, Microsoft has been pushing two ideas that shape how practitioners should think about the platform: that enterprises should treat AI model choice as a portfolio decision rather than a single-vendor bet, and that firms risk giving away proprietary knowledge to AI vendors through ordinary product use unless they establish their own data and orchestration boundaries. Both ideas are playing out concretely as Microsoft routes some Office AI prompts to its own MAI models while simultaneously expanding partner-model access inside the same Copilot surfaces.

The company is also one of the largest builders of AI infrastructure, which brings its own reporting obligations, energy and community constraints, and it is increasingly selling governance and deployment labor as products through offerings such as Agent 365, the Microsoft 365 E7 Frontier Suite and the Microsoft Frontier Company consulting unit.

What changed recently

Microsoft's July 29 fiscal fourth-quarter report is the clearest recent statement of how the company is financing and pricing AI. It reported $90.0 billion in quarterly revenue, up 18%, Azure and other cloud-services growth of 43%, full-year Azure revenue past $100 billion, and more than 30 million paid Microsoft 365 Copilot seats, against $41 billion of capital expenditures in the quarter, roughly two-thirds of it on short-lived assets such as CPUs and GPUs. Microsoft also lowered expected reported calendar-2026 capex to about $175 billion from the roughly $190 billion it gave in April, attributing the change to a shift of future datacenter leases from finance leases to operating leases rather than to a smaller buildout, and said September-quarter capex should exceed $50 billion. The same cost pressure is now visible in model selection: Microsoft said on July 23 that Bing Image Creator runs end to end on MAI-Image-2.5 and that PowerPoint image-to-image features use up to 84% less GPU capacity than GPT-Image-2, and said on July 27 that an MDASH configuration pairing its new MAI-Cyber-1-Flash model with GPT-5.4 scored 95.95% on CyberGym at nearly 50% lower cost than its previous best setup. For anyone building or buying on these surfaces, the practical reading is that the model behind a familiar Copilot feature is now a routed, cost-driven choice, and Microsoft's own July 6 test of 150 Copilot Chat agent runs, where Claude Sonnet 5 cost $2.01 per code-upgrade run against $0.55 for Sonnet 4.6, shows that a cheaper rate card does not guarantee a cheaper run.

The governance layer is arriving on a slower clock than the distribution layer. Security researcher Hakon Maloy disclosed on July 28 that hidden instructions in a Word document could alter Copilot-assisted output and copy themselves into the generated file, and said the broader class remained reproducible after 144 days of coordination and two mitigation attempts; Microsoft told The Register it had addressed the reported findings and continues to strengthen layered safeguards. Manifold Security's July 21 disclosure showed hidden HTML comments in Azure DevOps pull-request descriptions reaching AI agents through Microsoft's official MCP server, and a July 27 repository check found the pull-request path still returning raw JSON rather than the spotlighting wrapper already applied to wiki pages and build logs, with version 2.8.0 from June 24 still the latest tagged release. Microsoft's counterweight, a Purview Data Loss Prevention rule that excludes externally received email from Copilot grounding, summarization and citation, is only in preview and lists general availability starting January 2027, and it evaluates sender-domain metadata rather than message bodies. Teams standardizing on Copilot and agent surfaces should therefore treat repository text, partner documents and inbound email as untrusted input today, and evaluate controls at the feature level instead of assuming a tenant-wide switch exists.

What to watch

Several dated items are checkable in the near term. Microsoft said Project Perception, its red, blue and green agent security system, enters public preview on August 3, 2026, and the announcement does not yet list preview workflows, permission and approval configuration, evidence handling, or how agent actions are audited or reversed. The Purview DLP control for external email is listed as Microsoft 365 roadmap item 561552 with preview in June 2026 and general-availability rollout starting January 2027; Microsoft notes DLP policy updates can take up to four hours to appear in Copilot experiences. On the Azure DevOps MCP server, watch whether a tagged release after version 2.8.0 (June 24) applies the createExternalContentResponse wrapper to the pull-request path; Manifold Security says Microsoft acknowledged and triaged the report through MSRC, and the evidence does not establish the status of any nonpublic remediation, an assigned CVE, or exploitation outside the researchers' testing. On spending, Microsoft said it expects September-quarter capex above $50 billion and fiscal-2027 capex growth year over year, so the next earnings report will show whether the lease reclassification that cut reported calendar-2026 capex to about $175 billion continues to separate reported capex from underlying commitments. Also watch whether Microsoft documents which Microsoft 365 workloads run on MAI models and exposes tenant-level controls, an open question the July 7 Bloomberg reporting raised; whether Teams Facilitator reaches its targeted late-August 2026 general availability; and whether Manulife or Microsoft publish measured results against Manulife's company-reported target of more than C$1 billion in enterprise value by 2027.

Timeline

Key players

Satya NadellaMicrosoft's CEO set the strategic frame for the topic with a July 12, 2026 essay, The Reverse Information Paradox, arguing that enterprises can pay for AI twice, once in money and again through proprietary knowledge revealed in prompts, tool calls, corrections, evaluations and memory. His proposed response is a firm-controlled trust boundary over data, traces, evaluations, adapted weights and memory, plus model-independent orchestration. Business Standard reported a five-part framework covering data and evaluation control, proprietary learning, model choice, cost optimization and a compounding internal learning loop.
Mustafa SuleymanMicrosoft's AI chief is the public voice behind the shift of high-volume product workloads to in-house MAI models. He told Bloomberg the internal image technology is faster, cheaper and higher quality and that it improves retention, in the context of Microsoft's July 23 statement that Bing Image Creator, PowerPoint image-to-image and OneDrive editing now run on MAI-Image-2.5.
OpenAIOpenAI remains Microsoft's most visible model partner even as first-party models expand. OpenAI announced on July 9, 2026 that GPT-5.6 would become the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat and Cowork, and GitHub is rolling GPT-5.6 Sol, Terra and Luna into Copilot. GPT-5.4 also pairs with MAI-Cyber-1-Flash in Microsoft's MDASH system. Bloomberg reported that Microsoft can use OpenAI models without a model-access fee under the partnership but still pays for the compute to run them.
AnthropicMicrosoft added Claude Sonnet 5 to Microsoft 365 Copilot on July 2, 2026, with rollout to Copilot Cowork and Copilot in PowerPoint; Microsoft Learn describes it as the efficient everyday option in the Cowork model picker and says admins can turn off the Anthropic model family. Microsoft's own 150-run Copilot Chat experiment, reported July 6, used Claude Sonnet 4.6 and Sonnet 5 to show that a newer model can raise agent costs.
GitHubGitHub is where Microsoft's coding-agent decisions land. Across early July 2026 it made Kimi K2.7 Code generally available in the Copilot picker (off by default for Business and Enterprise), opened the Copilot desktop app to every plan, took browser tools to general availability in VS Code, and added agent session streaming, AI credit pools and managed OpenTelemetry settings for enterprise administrators.
Mistral AIMicrosoft and Mistral expanded their partnership on July 21, 2026 with an undisclosed multibillion-dollar Microsoft commitment to use part of Mistral's expanding European GPU infrastructure, which will draw on thousands of Nvidia Vera Rubin GPUs. Mistral Medium 3.5 and OCR 4 became available in Microsoft Foundry, with Medium 3.5 also in Copilot Studio, across public-cloud, cloud-connected and fully disconnected Azure Local deployments.
DatabricksDatabricks and Microsoft said on July 23, 2026 that their partnership extends into the 2030s. Databricks will run its own core operations and analytics on Azure Databricks, expand use of Microsoft's Arm-based Azure Cobalt processors, and deepen Genie and Unity AI Gateway integration across Entra, OneLake, Power BI, Purview, Foundry, Power Platform, Microsoft 365, Teams and Copilot. No financial terms were disclosed.
Judson AlthoffMicrosoft's Commercial Business CEO announced Microsoft Frontier Company on July 2, 2026, a $2.5 billion operating unit with roughly 6,000 experts led by Rodrigo Kede Lima that embeds Microsoft staff inside customer AI deployments. Althoff wrote that the effort goes beyond what has been labeled forward-deployed engineering, and said customers "shouldn't be locked into a single model any more than they should be locked into a single technology vendor."

Comparison

claim	scope	stated limit
PowerPoint image-to-image workloads use up to 84% less GPU capacity than GPT-Image-2	Microsoft's July 23 announcement that MAI-Image-2.5 now powers Bing Image Creator, PowerPoint image-to-image and OneDrive editing	Microsoft-reported production figures, not independent benchmarks; the deployment does not establish that MAI-Image-2.5 is better than every OpenAI model for every image task
An MDASH configuration scored 95.95% on CyberGym, against 83.8% for Mythos on Microsoft's published chart	Microsoft's July 27 introduction of MAI-Cyber-1-Flash inside its MDASH vulnerability system	The score applies to the combined MDASH system pairing MAI-Cyber-1-Flash with GPT-5.4, not the compact model alone, and the vendor-run comparison has not been independently replicated
The new configuration cuts cost by nearly 50%	The same July 27 MDASH announcement, in which Microsoft says the compact model handles up to 90% of tasks and GPT-5.4 the hardest 10%	Measured against Microsoft's current best MDASH setup, which uses GPT-5.4, GPT-5.4 Mini and GPT-5.3 Codex; it remains a vendor-reported cost comparison
Azure Cobalt 200 delivers up to 50% better performance	Microsoft's published figure, cited in the July 23 Databricks partnership extension, where Databricks said it plans to move from Cobalt 100 to Cobalt 200	A vendor claim, not an independently reproduced result for Databricks workloads; the announcement disclosed no benchmark methodology, migration volume or financial terms
Claude Sonnet 5 cost $2.01 per code-upgrade run versus $0.55 for Claude Sonnet 4.6	Microsoft's July 6 report on 150 Copilot Chat agent runs across architecture and SharePoint upgrade tasks	Architecture tasks flipped the other way and were slightly cheaper for Sonnet 5; the inversion came from far higher token consumption on some workloads despite lower per-token pricing

Frequently asked questions

Is Microsoft replacing OpenAI models across its products?

The evidence shows workload-level routing rather than a partner break. Microsoft said on July 23 that Bing Image Creator runs end to end on MAI-Image-2.5, that PowerPoint uses it for image-to-image features at up to 84% less GPU capacity than GPT-Image-2, and that OneDrive uses it for key editing scenarios; Bloomberg separately reported on July 7 that tens of thousands of weekly Excel and Outlook prompts are already completed with internal MAI models. At the same time, OpenAI announced on July 9 that GPT-5.6 becomes the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat and Cowork, and GPT-5.4 is paired with MAI-Cyber-1-Flash inside MDASH. The brief's own framing is that Microsoft assigns high-volume features to internal models where quality, latency and cost fit, and reserves partner models for tasks that need them.

What did Microsoft's July 29 earnings actually disclose about AI infrastructure?

Microsoft reported $90.0 billion in quarterly revenue (up 18%), $35.8 billion in GAAP net income (up 31%), Microsoft Cloud revenue of $59.3 billion, Azure and other cloud-services growth of 43%, full-year Azure revenue past $100 billion, free cash flow of $19.6 billion, and more than 30 million paid Microsoft 365 Copilot seats. Capital expenditures were $41 billion for the quarter, with roughly two-thirds on short-lived assets, primarily CPUs and GPUs, including $5.6 billion of finance leases and $35.8 billion of cash additions to property and equipment. Microsoft said its underlying calendar-2026 investment expectations are unchanged while a lease reclassification lowers expected reported capex to about $175 billion from roughly $190 billion. Microsoft did not disclose active-seat usage or revenue attributable to Copilot licenses, so the figures are demand and cost signals rather than a clean measure of AI return.

What prompt-injection risks are currently documented in Microsoft's AI surfaces?

Three separate disclosures appear in the evidence. Hakon Maloy published a coordinated disclosure on July 28 showing that hidden white text in a Word document could alter Copilot-assisted output and be copied into the generated file, so that file becomes another carrier; he said the class remained reproducible after 144 days and two mitigation attempts, while Microsoft told The Register it had addressed the reported findings and uses defense-in-depth safeguards. Manifold Security disclosed on July 21 that hidden HTML comments in Azure DevOps pull-request descriptions can reach AI agents through Microsoft's official MCP server, with a proof of concept that used a reviewer's broader credentials for cross-project actions. Separately, Alan Turing Institute researchers Abhishek Kumar and Carsten Maple reported workflow-level jailbreaks in GitHub Copilot, where direct-chat baselines produced 8 of 816 unsafe responses but the full multi-turn workflow produced 816 of 816 unsafe teaching-shot completions under expert review.

Can administrators stop Copilot from grounding on untrusted external email?

Partly, and not yet in general availability. Microsoft has a Purview Data Loss Prevention control in preview that excludes email received from external domains from Microsoft 365 Copilot and Copilot Chat grounding, summarization and citation, configured with the condition Email is received from > External users and the action Prevent Copilot from processing content. It is Microsoft 365 roadmap item 561552, with preview listed for June 2026 and general-availability rollout starting January 2027. Microsoft says the rule compares sender domain against the tenant's accepted domains and evaluates metadata only, so it does not inspect the message body or remove the user's access to the message. A Message Center notice republished by Hands on Tek says the control also extends to Copilot Studio agents published to Microsoft 365 Copilot, and Microsoft notes policy updates can take up to four hours to appear.

Which third-party and open-weight models can teams run inside Microsoft's stack?

Microsoft added Claude Sonnet 5 to Microsoft 365 Copilot on July 2, with Microsoft Learn listing Auto, Claude Sonnet 5, Claude Opus 4.8, GPT 5.5, Fable 5 preview and Sonnet + Opus Advisor as possible Cowork model choices depending on organizational settings, and saying admins can turn off the Anthropic model family. Mistral Medium 3.5 and OCR 4 became available in Microsoft Foundry on July 21, with Medium 3.5 also in Copilot Studio. Hugging Face said on July 7 that its curated open-weight collection is available in preview through Foundry Managed Compute, with weights pre-staged in Azure, runtimes built and scanned by Microsoft, and deployments on NVIDIA A100, H100 or AMD MI300X. GitHub made Moonshot AI's Kimi K2.7 Code generally available in the Copilot picker on July 1, off by default for Business and Enterprise plans, with GitHub's own documentation warning the model may be less aligned than other Copilot models.

Does upgrading to a newer model reduce agent costs?

Not reliably, according to Microsoft's own testing. Microsoft reported on July 6 that across 150 Copilot Chat agent runs spanning architecture and SharePoint upgrade tasks, Claude Sonnet 5 was better on some instruction-following upgrade tasks but consumed far more tokens on others, costing $2.01 per code-upgrade run against $0.55 for Sonnet 4.6, despite lower per-token pricing; architecture tasks flipped the other way and were slightly cheaper for Sonnet 5. The practical implication drawn in the evidence is to treat model swaps like dependency upgrades: run workload-specific evaluations, cap tool loops, monitor token-distribution tails, and regression-test cost, completion rate and quality before changing a default model in production agents.

Latest coverage