Claude models in Microsoft Foundry

Discover Claude models in Microsoft Foundry. Compare available models, capabilities, hosting options, and supported regions to choose the right model.

Anthropic’s Claude models bring advanced conversational AI capabilities to Microsoft Foundry, providing state-of-the-art language understanding and generation for intelligent applications. Claude models excel at complex reasoning, code generation, and multimodal tasks including image analysis. This article describes the available Claude models, how they’re hosted and billed, supported APIs, capabilities, and best practices.

To deploy and call a Claude model, see Deploy and use Claude models in Microsoft Foundry.

Items marked preview in this article are currently in preview. This preview is provided without a service-level agreement, and Microsoft doesn’t recommend it for production workloads. Certain features might not be supported or might have constrained capabilities. For more information, see Supplemental Terms of Use for Microsoft Azure Previews.

How Claude models are hosted and billed

Microsoft Foundry offers Claude models in two versions:

• Version 1: Hosted on Anthropic infrastructure; these models run on Anthropic’s infrastructure (outside of Azure).
• Version 2: Hosted on Azure; these models run on Azure infrastructure end-to-end and are all Generally available (GA).

Not all models are available in both versions. A model’s lifecycle stage, such as Preview or Generally available, can differ between the two versions. For per-model availability and lifecycle status, see Available Claude models.

To compare both hosting options across data residency, SLAs, support paths, compliance, and purchasing flow, see Compare Azure-hosted and Anthropic-hosted Claude models.

You access Claude models in Microsoft Foundry through Foundry Models from partners and community. Models from partners and community that Anthropic sells and operates are Non-Microsoft Products under the Product Terms.

Claude models in Foundry require an Azure Marketplace subscription and bill through Claude Consumption Units (CCU). Ensure that you have the permissions required to subscribe to model offerings before you deploy. For pricing details, see Claude Consumption Units (CCU) billing in Microsoft Foundry.

Available Claude models

The following Claude models comply with the EU watermarking standard: claude-mythos-5-1, claude-fable-5-1, and claude-fable-5. While claude-fable-5 uses single-key watermarking, claude-mythos-5-1 and claude-fable-5-1 use double key (interwoven 2-key and C2PA) as follows:

• Interwoven text watermarking (second key): a second key is added for interwoven watermark detection per EU standards. Watermarking happens server-side at generation time; there is no change to request or response shapes. A separate watermark detection API is in early access and isn’t part of the Foundry launch scope.
• C2PA watermarking: C2PA content-provenance marking, handled server/client-side by Anthropic surfaces. No API shape change.

The following table compares model availability for both versions of Claude models in Foundry. For details on the features referenced in the table, see the Capabilities and advanced features section.

For errors you might encounter when you deploy or call Claude models, see Deploy and use Claude models: Troubleshooting.

Model	Availability	Context window / Max output	Key capabilities	Best for
claude-fable-5-1	- Hosted on Anthropic infrastructure: Preview	1M / 128K	- Adaptive thinking - Long-running work - Scientific research with self verification - Knowledge work; handling challenging knowledge work with less hands-on oversight - Refusal stop_reason on dual-use safeguard policies1 - See What’s new in Claude Fable 5.1 and Claude Mythos 5.1	- Long-running agents - Coding and agents, with deeper reasoning for enterprise workflows - General science and research - Financial analysis - Vision
claude-fable-5	- Hosted on Anthropic infrastructure: Preview	1M / 128K	- Adaptive thinking - Reasoning over entire codebases and multi-day project context - Longer independent work than any prior Claude model - Self verification - Sub-agent orchestration - Refusal stop_reason on dual-use safeguard policies1 - See Migrating to Claude Fable 5.1 and Claude Mythos 5.1	- Cybersecurity - Autonomous coding - Long-running agents - Coding and agents, with deeper reasoning for enterprise workflows
claude-mythos-5-12	- Hosted on Anthropic infrastructure: Gated research preview	1M / 128K	- Adaptive thinking - Image and text input - Microsoft Entra ID authentication only - See What’s new in Claude Fable 5.1 and Claude Mythos 5.1	- Biology and life sciences - Cybersecurity (defensive use cases prioritized): vulnerability discovery, attack-surface auditing, red teaming, threat intelligence - Autonomous coding - Long-running agents
claude-mythos-52	- Hosted on Anthropic infrastructure: Gated research preview	1M / 128K	- Adaptive thinking - Image and text input - Microsoft Entra ID authentication only - See Migrating to Claude Fable 5.1 and Claude Mythos 5.1	- Biology and life sciences - Cybersecurity (defensive use cases prioritized): vulnerability discovery, attack-surface auditing, red teaming, threat intelligence - Autonomous coding - Long-running agents
claude-mythos-preview2	- Hosted on Anthropic infrastructure: Gated research preview	1M / 128K	- Adaptive thinking - Image and text input - Microsoft Entra ID authentication only - See Migrating to Claude Fable 5.1 and Claude Mythos 5.1	- Cybersecurity (defensive use cases prioritized) - Autonomous coding - Long-running agents
claude-opus-5-5	- Hosted on Azure: GA - Hosted on Anthropic infrastructure: GA	1M / 128K	- Cost efficient with lower per-token prices and less expensive cache reads - Agentic coding across long-running sessions; finding root causes before making changes and checking work as it goes - Adaptive thinking always on with effort parameter to steer depth and cost with xhigh and max effort levels; Default is medium; {“type”: “adaptive”} is accepted for thinking; Requests that set thinking to disabled, or to enabled with a manual budget_tokens, are rejected with a 400; Requests that omit thinking run with adaptive thinking - Communicates shorter and easy-to-follow summaries of long runs - Performs knowledge work to produce documents that require limited editing compared to previous models - Per-turn effort controls3 - Mid-conversation3 role:"system" - Token budgets3 (task_budget) - See What’s new in Claude Opus 5.5	- Long-running agents with minimal oversight and clear reports along the way - Enterprise workflows - Financial analysis - Computer use - Vision
claude-opus-5	- Hosted on Azure: GA - Hosted on Anthropic infrastructure: GA	1M / 128K	- Adaptive thinking with xhigh and max effort levels - Reasoning over entire codebases and multi-day project context - Per-turn effort controls3 - Mid-conversation3 role:"system" - Token budgets3 (task_budget) - See Migrating to Claude Opus 5.5	- Near-Fable intelligence for long-horizon coding and complex agentic orchestration - Long-running agents - Enterprise workflows - Financial analysis - Computer use
claude-opus-4-8	- Hosted on Azure: GA - Hosted on Anthropic infrastructure: GA	1M / 128K	- Adaptive thinking with xhigh effort level - Reasoning over entire codebases and multi-day project context - High-resolution image input (up to 2576px / 3.75MP) - See Migrating to Claude Opus 5.5	- Coding - Long-running agents - Financial analysis - Cybersecurity - Computer use
claude-opus-4-7	- Hosted on Anthropic infrastructure: GA	1M / 128K	- Adaptive thinking - Reasoning over entire codebases- High-resolution image input (up to 2576px / 3.75MP) - See Migrating to Claude Opus 5.5	- Coding - Enterprise workflows - Long-running agents - Multimodal reasoning - Financial analysis - Cybersecurity
claude-opus-4-6	- Hosted on Anthropic infrastructure: GA	1M / 128K	- Adaptive thinking - Image and text input - Computer use - Advanced tool use (search, programmatic calling, examples) - See Migrating to Claude Opus 5.5	- Coding - Enterprise agents
claude-opus-4-5	- Hosted on Anthropic infrastructure: GA	200K / 64K	- Extended thinking - Image and text input - Computer use - Advanced tool use (search, programmatic calling, examples) - See Migrating to Claude Opus 5.5	- Coding - Agents - Computer use - Enterprise workflows
claude-sonnet-5-5	- Hosted on Azure: GA - Hosted on Anthropic infrastructure: GA	1M / 128K	- Adaptive thinking - xhigh effort level - Reasoning over entire codebases and multi-day project context - High-res image input (up to 2576px / 3.75MP) are on by default - Mid-conversation3 role:"system" - Token budgets3 (task_budget)	- Coding - Long-running agents - Financial analysis - Cybersecurity - Computer use - See What’s new in Claude Sonnet 5.5
claude-sonnet-5	- Hosted on Azure: GA - Hosted on Anthropic infrastructure: GA	1M / 128K	- Adaptive thinking - xhigh effort level - Reasoning over entire codebases and multi-day project context - High-res image input (up to 2576px / 3.75MP) are on by default - Mid-conversation3 role:"system" - Token budgets3 (task_budget) - See What’s new in Claude Sonnet 5	- Coding - Long-running agents - Financial analysis - Cybersecurity - Computer use
claude-sonnet-4-6	- Hosted on Anthropic infrastructure: GA	1M / 128K	- Adaptive thinking - Image and text input - Computer use - Advanced tool use (search, programmatic calling, examples) - See Migrating to Claude Sonnet 5	- Coding - Agents - Enterprise workflows
claude-sonnet-4-5	- Hosted on Anthropic infrastructure: GA	200K / 64K	- Extended thinking - Image and text input - Computer use - See Migrating to Claude Sonnet 5	- Agents and complex, long-horizon tasks - High-volume workloads
claude-haiku-4-5	- Hosted on Azure: GA - Hosted on Anthropic infrastructure: GA	200K / 64K	- Extended thinking - Image and text input	- Coding - Agents

¹ Claude Fable 5 and Claude Fable 5.1 apply extra input/output classifiers that might refuse requests if the content triggers dual-use safeguard policies. When a refusal happens, the request returns a successful (200) response with a refusal indicator stop_reason: "refusal" instead of model-generated content. You aren’t billed for input tokens that are refused.

² Claude Mythos 5-1, Claude Mythos 5, and Claude Mythos Preview are only available as gated research preview. Access to the models is granted solely at Anthropic’s discretion and prioritized for defensive cybersecurity use cases. See the Claude Fable 5.1 & Claude Mythos 5.1 system card, Claude Mythos 5 system card, and Claude Mythos Preview system card for responsible use guidance.

³ Per-turn effort controls, Mid-conversation, and Token budgets are currently in Beta state.

API overview

The following table lists the APIs that you can use to interact with both the Hosted on Azure and Hosted on Anthropic infrastructure versions of Claude models in Foundry.

Use the Anthropic SDKs and the following Claude APIs:

The Hosted on Anthropic infrastructure version of Claude models in Foundry supports more APIs than the ones listed in this table. You can see them on the Claude API docs: API overview page.

API	Description
Messages1 (POST /v1/messages)	Core Messages API: Send a structured list of input messages with text or image content, including streaming responses. The model generates the next message in the conversation.
Token counting (POST /v1/messages/count_tokens)	Token Count API: Count the number of tokens in a message before sending it to Claude.

¹You can call the Messages API from the anthropic Python package, the @anthropic-ai/foundry-sdk JavaScript package, or directly through REST. The deployment endpoint follows the shape https://<resource-name>.services.ai.azure.com/anthropic/v1/messages, and REST and JavaScript clients use the anthropic-version: 2023-06-01 header.

Capabilities and advanced features

Claude models in Foundry expose core capabilities for processing, analyzing, and generating content, and tools that let Claude interact with external systems, execute code, and perform automated tasks. Claude’s API surface is organized into five areas:

• Model capabilities
• Tools
• Tool infrastructure
• Context management
• Files and assets

The following sections and tables summarize capabilities available across the Hosted on Azure and Hosted on Anthropic infrastructure versions of Claude models in Foundry. Unless noted, a capability applies to both versions.

The Hosted on Anthropic infrastructure version of Claude models in Foundry supports more capabilities than the ones listed in these tables. You can see the full list of capabilities on Claude Platform Docs: Features overview.

For more information about the available capabilities and advanced features for Claude models in Foundry, see the Microsoft Developer Blog.

Model capabilities

Ways to steer Claude and Claude’s direct outputs, including response format, reasoning depth, and input modalities.

Feature	Description
Streaming messages	When creating a Message, set "stream": true to incrementally stream the response using server-sent events (SSE).
Thinking	Enhanced reasoning capabilities for complex tasks, providing transparency into Claude’s step-by-step thought process before delivering its final answer. See Thinking and effort for thinking parameter values per model.
Adaptive thinking	Let Claude dynamically decide when and how much to think. This feature is the only thinking mode on Claude 4.7 and later models. Use the effort parameter to control thinking depth.
Effort	Control how many tokens Claude uses when responding, trading off between response thoroughness and token efficiency. See Thinking and effort for effort parameter values per model.
Citations	Ground Claude’s responses in sources, including search results content blocks search_result.
Images and vision	Process and analyze content from images. Hosted on Azure deployments only accept base64 encoded or URL-based images.
PDF support	Process and analyze text and visual content from PDF documents. Provide PDFs as base64 or URL.
1M context window	Up to 1 million tokens for processing large documents, extensive codebases, and long conversations. Support is subject to model eligibility.
Structured outputs	Constrain Claude’s responses to follow a specific JSON schema, using two complementary features: JSON outputs (the output_config.format parameter) for structured responses, and strict tool use (strict: true) for validated tool inputs. For Hosted on Azure deployments, structured outputs also support the legacy parameter for JSON outputs (the output_format parameter).

Thinking and effort

The Thinking feature allows specific values for the thinking parameter type, depending on the model, as described in the following table. The adaptive type configures the adaptive thinking feature, allowing the model to decide whether to think, based on query complexity and effort level. For example, thinking={"type": "adaptive"}.

Model	adaptive	enabled	disabled
claude-fable-5-1	Yes	No	No
claude-fable-5	Yes	No	No
claude-mythos-5-1	Yes	No	No
claude-mythos-5	Yes	No	No
claude-mythos-preview	Yes	Yes	No
claude-opus-5-5	Yes	No	Yes1
claude-opus-5	Yes	No	Yes1
claude-opus-4-8	Yes	No	Yes
claude-opus-4-7	Yes	No	Yes
claude-opus-4-6	Yes	Yes	Yes
claude-sonnet-5-5	Yes	Yes	Yes
claude-sonnet-5	Yes	No	Yes
claude-sonnet-4-6	Yes	Yes	Yes

¹ Thinking can be disabled only at effort high or below

The Effort feature allows specific effort levels for each model, as described in the following table. The xhigh level produces the same result as max.

Model	low	medium	high	xhigh	max
claude-fable-5-1	Yes	Yes	Yes	Yes	No
claude-fable-5	Yes	Yes	Yes	Yes	No
claude-mythos-5-1	Yes	Yes	Yes	Yes	No
claude-mythos-5	Yes	Yes	Yes	Yes	No
claude-opus-5-5	Yes	Yes	Yes	Yes	Yes
claude-opus-5	Yes	Yes	Yes	Yes	Yes
claude-opus-4-8	Yes	Yes	Yes	Yes	Yes
claude-opus-4-7	Yes	Yes	Yes	Yes	Yes
claude-opus-4-6	Yes	Yes	Yes	No	Yes
claude-sonnet-5-5	Yes	Yes	Yes	Yes	Yes
claude-sonnet-5	Yes	Yes	Yes	Yes	Yes
claude-sonnet-4-6	Yes	Yes	Yes	No	Yes

Tools

Let Claude take actions on the web or in your environment. This feature consists of built-in tools that Claude invokes through tool_use. The platform runs server-side tools, and you implement and execute client-side tools.

Feature	Description
Tool use with client-executed tools	Custom tools plus Anthropic-defined bash, text editor, computer use, and memory. For more information about these tools, see Bash, Text editor, Computer use, and Memory.
Web search	Discover current real-world data from across the web to use to augment Claude’s knowledge. For Hosted on Azure deployments, only the web_search_20250305 tool version is supported.
Web fetch	Retrieve and perform in-depth analysis of full content from specified web pages and PDF documents, augmenting Claude’s context with live web content. On Foundry, web fetch requires a Hosted on Anthropic infrastructure deployment. For Hosted on Azure deployments, only the web_fetch_20250910 tool version is supported.

Tool infrastructure

Discover, orchestrate, and scale tool use.

Feature	Description
Fine-grained tool streaming	Stream tool use parameters without buffering or JSON validation, reducing latency for large parameters. Requires the anthropic-beta header fine-grained-tool-streaming-2025-05-14.
MCP connector	Connect to remote MCP servers directly from the Messages API without a separate MCP client.
Tool search	Scale to thousands of tools by dynamically discovering and loading tools on demand using regex- and BM25-based search, optimizing context usage and improving tool selection accuracy. For Hosted on Azure deployments, both the tool_search_tool_bm25_20251119 and tool_search_tool_regex_20251119 tool versions are supported. The legacy aliases tool_search_tool_bm25 and tool_search_tool_regex are also accepted.

Context management

Control and optimize Claude’s context window for long-running sessions.

Feature	Description
Automatic prompt caching	Simplify prompt caching to a single API parameter. The system automatically caches the last cacheable block in your request, moving the cache point forward as conversations grow.
Prompt caching (5m)	Provide Claude with more background knowledge and example outputs to reduce costs and latency.
Prompt caching (1hr)	Extended 1-hour cache duration for less frequently accessed but important context, complementing the standard 5-minute cache.
Context editing	Automatically manage conversation context with configurable strategies, including clearing tool results and managing thinking blocks. Requires the anthropic-beta header context-management-2025-06-27.
Token counting	Token counting enables you to determine the number of tokens in a message before sending it to Claude, helping you make informed decisions about your prompts and usage.

Files and assets

Manage the documents and data you provide to Claude.

Feature	Description
Files API	Currently available only on Hosted on Anthropic infrastructure deployments. Upload and manage files to use with Claude without re-uploading content with each request. Supports PDFs, images, and text files.

Agent support

• Microsoft Agent Framework supports creating agents that use Claude models.
• Build custom AI agents with the Claude Agent SDK.

Deployment types and regions

Claude models in Foundry are available for the following deployment types in specific Azure regions:

• Global Standard: All Claude models (Hosted on Azure and Hosted on Anthropic infrastructure).
• Data Zone Standard (US): Hosted on Azure versions of claude-opus-5-5, claude-opus-5, claude-opus-4-8, claude-sonnet-5, and claude-sonnet-5-5.

For the exact Azure regions where Claude models are available for deployment, see Region availability by deployment type.

Quotas and rate limits

Rate limits for Claude models vary by model, deployment type, hosting version, and Azure subscription type. Limits are measured in requests per minute (RPM), uncached input tokens per minute (ITPM), and output tokens per minute (OTPM).

For current limits, shared quota behavior, and prompt cache accounting, see Claude model quotas and rate limits.

Responsible AI considerations

When using Claude models in Foundry, consider these responsible AI practices:

• Review Data, privacy, and security for Claude models in Microsoft Foundry to understand how your data is processed and retained when you deploy Claude models.

• Configure AI content safety during model inference, because Foundry doesn’t provide built-in content filtering for Claude models at deployment time.

• Ensure your applications comply with Anthropic’s Acceptable Use Policy. Also, see details of safety evaluations for Claude Fable 5.1 & Claude Mythos 5.1 , Claude Fable 5, Claude Mythos 5, Claude Mythos Preview, Claude Opus 5.5, Claude Opus 5, Claude Opus 4.8, Claude Opus 4.7, Claude Opus 4.6, Claude Opus 4.5, Claude Sonnet 5, Claude Sonnet 4.6, Claude Sonnet 4.5, and Claude Haiku 4.5.

Best practices

Follow these best practices when working with Claude models in Foundry:

Prompt engineering

• Clear instructions: Provide specific and detailed prompts.
• Context management: Use the available context window effectively.
• Role definitions: Use system messages to define the assistant’s role and behavior.
• Structured prompts: Use consistent formatting for better results.

Cost optimization

To optimize your usage and avoid rate limiting:

• Implement retry logic: Handle 429 responses with exponential backoff.
• Batch requests: Combine multiple prompts when possible.
• Monitor token usage: Track your token consumption and request patterns.
• Use appropriate models: Use the most cost-effective model for your use case. See Available Claude models.

Related content

• Deploy and use Claude models in Microsoft Foundry
• Deploy Claude models in Microsoft Foundry using Bicep or Terraform
• Foundry Models from partners and community
• Claude Consumption Units (CCU) billing in Microsoft Foundry
• Data, privacy, and security for Claude models in Microsoft Foundry