Microsoft let Anthropic move in while OpenAI was sleeping on the couch

RuntimeWire's static analysis of the Claude for Windows 1.49585.0 binary found new Microsoft-specific enterprise authentication infrastructure, including explicit support for Microsoft Entra Continuous Access Evaluation, a local Windows authentication broker and additional enterprise credential paths, none of which appeared in the earlier Claude for Windows 1.46388.4 build distributed two days prior. The new build introduces configuration markers such as McpBuiltinM365Cae, MCP_LOCAL_AUTH_CAE_DISABLE and MCP_LOCAL_AUTH_BROKER_MODE, and an internal configuration description says Claude can request CAE-capable Microsoft Graph tokens lasting up to roughly 28 hours while remaining revocable within minutes. RuntimeWire said it requested comment from Anthropic and the company had not responded by publication time; the reporting found no public evidence OpenAI ever received comparable integration depth in its own Windows client.

Microsoft let Anthropic move in while OpenAI was sleeping on the couch An unreleased Claude for Windows build shows Anthropic wiring itself deeper into Microsoft’s enterprise identity stack, as the partnership that defined the first years of the AI boom becomes increasingly less exclusive. By Ryan Merket /author/ryan-merket · Published RUNTIMEWIRE INVESTIGATION — Original analysis Original reporting by RuntimeWire, based on documents, reverse engineering. Why it matters Enterprise AI distribution increasingly runs through identity, security and procurement. Anthropic is now integrating with all three inside Microsoft’s customer base — at a depth RuntimeWire found no public evidence OpenAI ever received in its own Windows client. Reporting record Finding Claude for Windows version 1.49585.0 contains new Microsoft-specific enterprise authentication infrastructure, including explicit support for Entra Continuous Access Evaluation, a local Windows authentication broker and additional enterprise credential paths. How we verified Methods: documents, reverse engineering. Static analysis of Claude for Windows versions 1.49585.0 and 1.46388.4 identified new CAE configuration markers, default-enabled CAE, Microsoft cp1 capability references, local broker code, enterprise credential options and sovereign-cloud handling; Microsoft and Anthropic documentation provided context. RuntimeWire compared the two distributed Claude for Windows binaries, examining application resources, configuration markers and authentication code, and compared the findings with Microsoft and Anthropic documentation. Tested versions: Claude for Windows 1.49585.0, Claude for Windows 1.46388.4. Reproduction Reproduction does not apply to this reporting document-driven . Company response RuntimeWire requested comment; the company had not responded by publication time. RuntimeWire asked Anthropic whether the broker and CAE implementation is intended for its existing Microsoft 365 connector, a Windows-specific implementation or another enterprise deployment architecture. Microsoft and OpenAI are still together. But Anthropic appears to have a toothbrush in the bathroom and access to Entra. RuntimeWire found new Microsoft-specific enterprise authentication infrastructure inside the latest Claude for Windows build, including explicit support for Microsoft Entra Continuous Access Evaluation, a Windows-native authentication broker and additional credential paths for managed corporate environments. The changes go deeper than Microsoft putting another model in a dropdown. Anthropic is adapting Claude to some of the plumbing Microsoft's enterprise customers already rely on to decide who can access corporate data, from where and for how long. RuntimeWire compared Claude for Windows 1.49585.0 with version 1.46388.4 , distributed two days earlier. The newer binary introduces configuration markers including: McpBuiltinM365Cae McpBuiltinM365CaePh McpBuiltinM365CaeHint MCP LOCAL AUTH CAE DISABLE MCP LOCAL AUTH BROKER MODE MCP LOCAL AUTH BROKER DISABLE Those CAE markers do not appear in the earlier build reviewed by RuntimeWire. An internal configuration description says Claude can request CAE-capable Microsoft Graph tokens, describing them as long-lived tokens lasting up to roughly 28 hours while remaining revocable within minutes. Microsoft's documentation https://learn.microsoft.com/en-us/entra/identity/conditional-access/concept-continuous-access-evaluation says that is exactly what Continuous Access Evaluation is designed to do. Ordinary access tokens generally require periodic expiration before some changes can be enforced. CAE lets supported services react to security events while a token is still alive. Microsoft says CAE sessions can use access tokens lasting up to 28 hours , while events such as account disablement, password changes, elevated user risk and explicit session revocation can be enforced close to real time. The Claude implementation also references Microsoft's cp1 client capability. Microsoft says applications declare cp1 https://learn.microsoft.com/en-us/entra/identity/conditional-access/concept-continuous-access-evaluation-workload?utm source=chatgpt.com when they can understand the claims challenges required for CAE. Microsoft Graph sends those challenges only to clients that explicitly advertise the capability. In the Claude code reviewed by RuntimeWire, the CAE option is enabled by default. That makes the discovery especially interesting because Microsoft's currently published certification record for Claude for Microsoft 365 says something different. Microsoft's App Certification page https://learn.microsoft.com/en-us/microsoft-365-app-certification/word/claude-by-anthropic-for-office-microsoft-365?utm source=chatgpt.com says Claude supports Entra ID and Conditional Access. It specifically asks, however, whether the certified application supports Continuous Access Evaluation. The answer is "No." The information was last updated by the developer on August 4th and published by Microsoft on August 7th. There is an important caveat. Microsoft's certification appears to cover Claude's Microsoft 365 Office application and its nested authentication flow. It may represent a different Claude surface or authentication path than the Microsoft 365 MCP code bundled in Claude Desktop. RuntimeWire therefore is not treating the discrepancy as evidence that Anthropic provided Microsoft with inaccurate information. It does show something simpler and more useful: a capability Microsoft's current public Claude documentation says is absent now exists explicitly inside a new Claude Windows binary. Claude is getting closer to Windows itself CAE is one part of a larger authentication change in the build. Anthropic has replaced an office365-mcp.mjs resource with: office365-mcp-stdio.mjs The new implementation includes machinery for running Microsoft authentication through a local broker. RuntimeWire found references including: MCP PARENT WINDOW HANDLE MCP LOCAL AUTH BROKER HOST MCP LOCAL AUTH BROKER MODE The supported broker modes are: auto disabled required The required path is designed to fail closed if broker authentication cannot be established. References to the parent Windows window handle are particularly suggestive. Microsoft's native authentication stack uses the Windows environment to broker organizational identity rather than forcing every application to independently implement an isolated browser login. The build also contains enterprise credential options including: MCP AUTH CLIENT SECRET MCP AUTH CLIENT CERT PATH MCP AUTH CLIENT ASSERTION FILE along with code handling Microsoft sovereign-cloud environments. RuntimeWire cannot determine from the binary alone whether Anthropic intends to expose all of these mechanisms to customers, which enterprise tiers will receive them or whether some exist for internal and managed deployments. Anthropic's current Microsoft 365 documentation https://support.claude.com/en/articles/12684923-microsoft-365-connector-security-guide?utm source=chatgpt.com describes a substantially different architecture. Its public connector operates as an Anthropic-hosted integration . Credentials are managed by Anthropic's backend, and Microsoft's Azure SDK performs an On-Behalf-Of token exchange before the connector accesses Microsoft Graph. Anthropic also warns that Conditional Access behaves differently https://support.claude.com/en/articles/12684923-microsoft-365-connector-security-guide?utm source=chatgpt.com through that architecture. After the initial sign-in, later Microsoft 365 requests originate from Anthropic's servers rather than necessarily appearing to come from the user's current device or network. The new Windows code points toward Anthropic doing significantly more Microsoft authentication work locally. RuntimeWire has asked Anthropic whether the broker and CAE implementation is intended for its existing Microsoft 365 connector, a Windows-specific implementation or another enterprise deployment architecture. The revenue opportunity is hiding in the plumbing Anthropic is already operating at a scale where enterprise compatibility can translate directly into revenue. The company says it serves more than 300,000 business customers , with more than 500 customers spending at least $1 million annually https://www.anthropic.com/news/anthropic-raises-30-billion-in-series-g-funding-at-380-billion-post-money-valuation and eight of the Fortune 10 using Claude https://www.anthropic.com/enterprise . Those are the organizations where identity, security, compliance, deployment controls and procurement can determine whether an AI product remains a departmental pilot or expands across the company. Entra, Conditional Access, native authentication brokering, enterprise certificates, centralized policy and sovereign-cloud support are not decorative features in that sales process. They are often deployment requirements. Microsoft 365 Commercial remains one of the world's largest enterprise software footprints, and Microsoft continues to report growth https://www.microsoft.com/en-us/Investor/earnings/FY-2026-Q4/press-release-webcast in both commercial seats and Microsoft 365 commercial cloud revenue. Anthropic is therefore building compatibility for the same Microsoft-heavy organizations it is already trying to convert into larger Claude accounts. The commercial inference is straightforward, even if the resulting revenue cannot be calculated from the binary. Deeper Microsoft-native integration can remove identity and compliance blockers that keep deployments small. Removing one of those blockers can turn a limited team deployment into a broader enterprise rollout. For a company with more than 500 customers already spending seven figures annually, even a modest number of expanded contracts could be financially meaningful. RuntimeWire cannot estimate the incremental revenue these changes could generate, and the binary does not reveal pricing, customer commitments or whether every path is shipping. But the opportunity is substantial because Anthropic appears to be reducing deployment friction inside one of the world's largest enterprise software environments. How the Windows work compares with ChatGPT OpenAI publicly documents ChatGPT connections to Microsoft 365 services https://help.openai.com/en/articles/11487775-connectors-in-chatgpt through Microsoft Graph and Microsoft Entra-approved permissions. RuntimeWire found no public evidence that the ChatGPT Windows client exposes the same combination of CAE-aware local brokering, parent HWND handling, fail-closed broker modes, enterprise certificate and assertion paths, and sovereign-cloud handling found in the Claude binary. Microsoft's public documentation https://learn.microsoft.com/en-us/copilot/microsoft-365/use-openai-models also says OpenAI-operated models in Copilot are currently unavailable in GCC, GCC High, DoD and sovereign clouds. The Claude Windows binary, by contrast, contains sovereign-cloud paths. That is a finding about code present in an unreleased build, not proof that Anthropic has shipped sovereign-cloud support or that Microsoft granted Anthropic preferential treatment. Microsoft already invited Claude inside That engineering work lands against a remarkable reversal in Microsoft's AI relationships. In 2019, Microsoft invested $1 billion in OpenAI and announced an exclusive computing partnership https://news.microsoft.com/source/2019/07/22/openai-forms-exclusive-computing-partnership-with-microsoft-to-build-new-azure-ai-supercomputing-technologies/?return=https%3A%2F%2Fnews.microsoft.com%2Fabout%2F%3Fmsockid%3D2a52ae9cd2fd63fe0d82b857d3246257&utm source=chatgpt.com . OpenAI would build on Azure, the companies would jointly develop supercomputing technology, and Microsoft would become OpenAI's preferred partner for commercializing its technology. By January 2023, Microsoft was describing Azure as OpenAI's exclusive cloud provider https://blogs.microsoft.com/blog/2023/01/23/microsoftandopenaiextendpartnership/?utm source=chatgpt.com . It announced another multiyear, multibillion-dollar investment and said it would deploy OpenAI models throughout Microsoft's consumer and enterprise products. For a while, Microsoft's AI strategy and OpenAI's commercial ascent were difficult to separate. Microsoft supplied enormous amounts of computing infrastructure. OpenAI supplied frontier models. Azure sold them to enterprises. OpenAI technology spread into Bing, GitHub, Office and Microsoft's Copilot products. Then the relationship started acquiring escape hatches. In January 2025, Microsoft said OpenAI could build additional computing capacity elsewhere https://blogs.microsoft.com/blog/2025/01/21/microsoft-and-openai-evolve-partnership-to-drive-the-next-phase-of-ai/?utm source=chatgpt.com , while Microsoft retained a right of first refusal on new capacity. The core IP, API and revenue-sharing arrangements remained intact. By September, Microsoft was putting Anthropic models directly beside OpenAI models in Microsoft Copilot Studio . Microsoft announced Claude Sonnet 4 and Claude Opus 4.1 as choices https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/anthropic-joins-the-multi-model-lineup-in-microsoft-copilot-studio/?msockid=1fffa93510546a242a24bf8d11d96b7f&utm source=chatgpt.com for orchestration, chat and deep-reasoning workloads, while OpenAI remained the default for new agents. Days later, Microsoft added Claude Sonnet 4.5 https://www.microsoft.com/en-us/microsoft-copilot/blog/copilot-studio/available-today-claude-sonnet-4-5-in-microsoft-copilot-studio/?msockid=3b67cc9af3ab63213021dae0f2b26207&utm source=chatgpt.com . The message from Redmond was "model choice." From OpenAI's old seat at the table, it probably looked a lot like Microsoft dating again. Then the marriage opened up The contractual changes eventually became much larger. In October 2025, Microsoft and OpenAI announced a revised relationship https://blogs.microsoft.com/blog/2025/10/28/the-next-chapter-of-the-microsoft-openai-partnership/?cmid=30fb0089-3f89-41a0-a801-7464d0de6a4a&utm source=chatgpt.com that gave both companies greater freedom to work independently. Microsoft could pursue AGI itself or with third parties. OpenAI gained additional freedom around third-party product development and infrastructure, while Microsoft retained extensive IP and commercial rights. The relationship still mattered enough that Microsoft and OpenAI published a joint statement on February 27th, 2026 https://blogs.microsoft.com/blog/2026/02/27/microsoft-and-openai-joint-statement-on-continuing-partnership/?utm source=chatgpt.com reiterating that they continued to work closely together and that the companies were designed to have room to pursue opportunities independently. Two months later, they loosened it again. Under an amended agreement announced April 27th https://openai.com/index/next-phase-of-microsoft-partnership/?utm source=chatgpt.com , Microsoft remains OpenAI's primary cloud partner and OpenAI products are supposed to ship first on Azure when Microsoft can support them. But OpenAI can now serve its products through any cloud provider . Microsoft's license to OpenAI model and product intellectual property runs through 2032, but the license is now non-exclusive . Microsoft also no longer pays OpenAI a revenue share, while OpenAI's revenue-sharing payments to Microsoft continue through 2030 subject to a cap. Microsoft remains a major shareholder. This is hardly a breakup. It is a very different relationship from the one Microsoft described in 2023, when Azure was OpenAI's exclusive cloud and Microsoft was putting OpenAI technology across its entire product portfolio. As OpenAI gained freedom from Microsoft, Microsoft gained freedom from OpenAI. And Anthropic was waiting. Anthropic has a toothbrush in the bathroom Microsoft adopting Claude as another foundation model supplier could be dismissed as ordinary procurement strategy. Depending on one AI lab would give that lab enormous pricing power and strategic influence over Microsoft's products. The new Claude Windows code feels different. Anthropic engineers are accounting for Microsoft's token revocation semantics. They're handling Entra claims challenges. They're building around native Windows authentication brokers. They're adding client certificate and assertion paths for enterprises. They're dealing with sovereign Microsoft environments. This is what it looks like when a product begins learning the house rules. The most interesting contrast may be that both sides of the original Microsoft-OpenAI partnership are now doing essentially the same thing. OpenAI has spent years reducing its dependence on Microsoft's infrastructure and distribution. Microsoft has spent years reducing its dependence on OpenAI's models. OpenAI gained other clouds. Microsoft gained other frontier labs. Claude is the clearest beneficiary. Microsoft still owns a major stake in OpenAI. It still has valuable OpenAI IP rights. OpenAI remains deeply integrated across Microsoft products. The companies continue building infrastructure and products together. Nobody has filed for divorce. But while OpenAI has been establishing a life outside Microsoft, its fiercest rival has been getting awfully comfortable inside. Now Anthropic appears to know where Microsoft keeps the Entra tokens.