News
Microsoft Brings Anthropic's Claude Opus 4.5 to Foundry Preview
Microsoft has introduced Anthropic's Claude Opus 4.5 model in public preview in Microsoft Foundry, while also making it available through GitHub Copilot paid plans and Microsoft Copilot Studio.
The new offering now headlines the Foundry site:
[Click on image for larger view.] Microsoft Foundry (source: Microsoft).
The company positioned the release as another step in expanding Foundry's catalog of frontier models and giving Azure customers access to third-party options alongside other offerings. Microsoft described the current moment as one where AI models are shifting from assistant roles to "genuine collaborators" that can execute multi-tool workflows under constraints, and said Opus 4.5 reflects that direction.
The announcement comes amid a broader branding change, with Azure AI Foundry now renamed Microsoft Foundry. Microsoft again framed Foundry as an integrated platform for deploying and scaling AI applications and agents with centralized governance, security, and observability. The company said that putting Claude models in the same platform as GPT models gives customers flexibility in choosing models for different workloads.
Microsoft characterized Claude Opus 4.5 as Anthropic's newest flagship model and said it is intended for coding, agentic workflows, and enterprise productivity. Microsoft said Opus 4.5 outperforms Claude Sonnet 4.5 and Claude Opus 4.1 while coming at a lower cost, and it emphasized versatility across software engineering, complex reasoning, tool use, and vision. Microsoft attributed the performance claims to Anthropic benchmark reporting and early tester feedback. For example, Microsoft said Opus 4.5 hits a new high on SWE-bench Verified at 80.9 percent and improves results on Terminal-bench 2.0 for agentic terminal coding. It also highlighted tool-use benchmarks in retail and telecom scenarios, along with scaled tool use in MCP Atlas and computer-use gains in OSWorld. Microsoft said these results point to better handling of ambiguous requirements, reasoning over architectural tradeoffs, and finding fixes across multiple systems.
Beyond benchmark numbers, the company focused on new agentic tool-use features in Opus 4.5. Microsoft said the model supports programmatic tool calling, letting tools be executed directly in Python for more deterministic workflows. It also said Opus 4.5 includes tool search, giving agents a way to dynamically discover tools from large libraries without consuming context window space, and tool-use examples designed to improve accuracy with complex schemas. Microsoft said these capabilities can power agents that "work seamlessly across hundreds of tools" and are relevant to multi-tool enterprise scenarios such as cybersecurity, full-stack engineering, and financial modeling.
Microsoft also introduced Foundry-side capabilities tied to Opus 4.5. It announced an Effort Parameter in beta, described as a control for how much computational effort Claude allocates between thinking, tool calls, and responses, enabling tradeoffs among quality, latency, and cost. The company also pointed to compaction control through software development kit helpers meant to support long-running agent sessions by managing context more efficiently. Microsoft presented both as steps toward more predictable and operationally controllable agent deployments in production settings.
On multimodal capability, Microsoft said Opus 4.5 is Anthropic's best vision model to date and improves visual reasoning and multi-step "computer use" automation. The company said desktop task automation is more reliable, and that productivity agents can create spreadsheets, presentations, and documents with more consistency and domain awareness. Microsoft also said the model better leverages memory to keep context across related files, positioning that as useful for work that spans large, multi-document projects.
Safety and security updates were included as part of the release description. Microsoft said Anthropic reports a reduced rate of misaligned responses, better robustness against prompt-injection attacks, and more reliable behavior on complex tasks. Microsoft tied those claims to its own governance expectations for enterprise deployments.
Microsoft listed target use cases for Opus 4.5 on Foundry, including software development agents that handle complex, multi-system tasks with minimal supervision, financial analysis agents that connect insights across filings and internal data, cybersecurity agents that correlate logs and threat intelligence, and operational agents that coordinate workflows across tools and systems.
Microsoft said Opus 4.5 is priced at one-third the cost of previous Opus-class models. For Foundry's serverless pay-as-you-go offering, Microsoft listed global standard deployment and regional availability in East US 2 and Sweden Central. Pricing is set at $5 per million input tokens and $25 per million output tokens. Microsoft said Opus 4.5 is available now in Microsoft Foundry and is "coming soon in Visual Studio Code via the Foundry extension," signaling continued emphasis on in-editor access for developers.
About the Author
David Ramel is an editor and writer at Converge 360.