Edit

Claude models in Microsoft Foundry

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, quotas, and best practices.

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

Important

Items marked (preview) in this article are currently in public preview. This preview is provided without a service-level agreement, and we don'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.

Note

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 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.

Warning

1M context beta on Claude Sonnet 4.5 was retired on April 30, 2026.

Starting May 1, 2026:

  • Requests greater than 200K tokens that include the context-1m-2025-08-07 beta header on Sonnet 4.5 return an error.
  • Requests 200K tokens or fewer remain unaffected, even with the header present.

To migrate, remove the context-1m-2025-08-07 beta header from your requests. For workloads that require 1M context, migrate to Claude Sonnet 4.6 (where 1M context is generally available) or to Claude Opus 4.6 or Claude Opus 4.7 for higher-intelligence workloads.

Model Availability Context window / Max output Key capabilities Best for
claude-mythos-51
  • Hosted on Anthropic infrastructure: Gated research preview
1M / 128K
  • Biology and life sciences
  • Cybersecurity (defensive use cases prioritized): vulnerability discovery, attack-surface auditing, red teaming, threat intelligence
  • Autonomous coding
  • Long-running agents
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 policies2
  • See See Introducing Claude Fable 5 and Claude Mythos 5
  • Cybersecurity
  • Autonomous coding
  • Long-running agents
  • Coding and agents, with deeper reasoning for enterprise workflows
claude-mythos-preview1
  • Hosted on Anthropic infrastructure: Gated research preview
1M / 128K
  • Cybersecurity (defensive use cases prioritized)
  • Autonomous coding
  • Long-running agents
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 What's new in Claude Opus 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
  • 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
  • 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
  • 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
  • Coding
  • Agents
  • Computer use
  • Enterprise workflows
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
  • 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 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 Mythos Preview system card and Claude Mythos 5 system card for responsible use guidance.

2 Claude Fable 5 applies additional input/output classifiers that might refuse requests whose content triggers dual-use safeguard policies. When a refusal occurs, the request returns a successful (200) response with a refusal indicator stop_reason: "refusal" instead of model-generated content. You're not billed for input tokens that are refused.

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:

Tip

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:

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.

Tip

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-mythos-5 Yes No No
claude-fable-5 Yes No No
claude-mythos-preview Yes Yes No
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 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-mythos-5 Yes Yes Yes Yes No
claude-fable-5 Yes Yes Yes Yes No
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 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

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, claude-opus-4-8, and claude-sonnet-5.

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

Quotas and rate limits

This section explains how deployments share quota and what rate limits apply to them. Subscription-level management handles the deployment quota. Resources and regions share the quota instead of allocating it separately for each resource or region.

  • All Global Standard deployments of the same model and version in a subscription draw from one shared quota pool across all regions.
  • All Data Zone Standard deployments of the same model and version in a subscription draw from a shared quota pool within each data zone (for example, US).

For more information about quota management for Foundry Models, see Microsoft Foundry Models quotas and limits.

Cache-aware ITPM

Claude models in Foundry measure rate limits in requests per minute (RPM), uncached input tokens per minute (ITPM), and output tokens per minute (OTPM) for each model.

For most Claude models, only uncached input tokens count toward your ITPM rate limits. These tokens include:

  • Input tokens — tokens in the request after the last cache breakpoint (uncached input).

  • Cache creation input tokens — tokens being written to cache, which comprises:

    • Cache write 5m TPM — tokens being written to the 5-minute prompt cache.
    • Cache write 1h TPM — tokens being written to the 1-hour prompt cache.

Tip

The total input tokens is the sum of Input tokens, Cache creation input tokens, and Cache read input tokens (the tokens read from cache). However, the Cache read input tokens don't count towards ITPM. OTPM also doesn't count towards ITPM.

For more information about rate limits and cache, see Claude API Docs: Rate limits.

Rate limits by subscription type

Your Azure subscription type determines your rate limits. The Version 2: Hosted on Azure and Version 1: Hosted on Anthropic infrastructure columns indicate whether quota is available for that model and deployment type combination. Yes means quota is available. N/A means the model and version combination don't have quota for that deployment type.

The following table lists rate limits. To increase your quota beyond the default limits, submit a request through the quota increase request form.

Pay-as-you-go

Model Deployment type Version 2: Hosted on Azure Version 1: Hosted on Anthropic infrastructure RPM ITPM OTPM
claude-fable-5 Global Standard N/A Yes 0 0 0
claude-opus-5 Global Standard Yes Yes 40 40,000 8,000
claude-opus-5 Data Zone Standard (US) Yes N/A 40 40,000 8,000
claude-opus-4-8 Global Standard Yes Yes 40 40,000 8,000
claude-opus-4-8 Data Zone Standard (US) Yes N/A 40 40,000 8,000
claude-opus-4-7 Global Standard N/A Yes 40 40,000 8,000
claude-opus-4-6 Global Standard N/A Yes 40 40,000 8,000
claude-opus-4-5 Global Standard N/A Yes 40 40,000 8,000
claude-sonnet-5 Global Standard Yes Yes 40 40,000 8,000
claude-sonnet-5 Data Zone Standard (US) Yes N/A 40 40,000 8,000
claude-sonnet-4-6 Global Standard N/A Yes 80 80,000 16,000
claude-sonnet-4-5 Global Standard N/A Yes 80 80,000 16,000
claude-haiku-4-5 Global Standard Yes Yes 80 80,000 16,000

Responsible AI considerations

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

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.