anthropics/anthropic-sdk-python v0.120.0

Anthropic SDK Python v0.120.0: New Model Capabilities and Enhanced Tool Management

Anthropic has released version 0.120.0 of its Python SDK, introducing Claude Opus 5 alongside significant improvements to tool handling and token credit management. The update reflects the company's continued evolution of its Claude AI model family and the infrastructure supporting developers who build applications on top of these models.

TL;DR

• Claude Opus 5: A new flagship model joining Anthropic's model lineup with improved capabilities
• Tool Management: Native support for dynamically adding and removing tools during conversations, with corresponding event tracking
• Token Fallbacks: Enhanced credit token flexibility and standardized server-side fallback handling
• Impact: Developers gain more control over tool-use workflows and better token economy management in production applications

Background

Anthropic's Python SDK serves as the primary interface for developers integrating Claude into applications. Each release incorporates API improvements that reflect user needs and architectural lessons learned from production deployments. The company maintains a regular cadence of updates, with v0.120.0 arriving several months into 2026.

The addition of new models has been consistent with Anthropic's strategy of offering tiered capabilities—from efficient smaller models to more powerful flagship versions. Meanwhile, tool use—the ability for AI models to call functions and interact with external systems—has emerged as a critical capability for production applications. Previously, tools were typically defined upfront; this release adds more granular control.

Token credit management has become increasingly important as organizations look to optimize costs across different deployment scenarios and fallback mechanisms. The SDK improvements address real operational challenges developers face when managing API consumption at scale.

How it works

Claude Opus 5: The New Flagship

Claude Opus 5 represents the latest iteration in Anthropic's capability hierarchy. While the company hasn't detailed exact performance metrics in this release, Opus models historically position themselves as the most capable variants, designed for complex reasoning tasks, extended context windows, and nuanced problem-solving.

The model's addition to the SDK enables developers to select it via standard model identifier strings when instantiating the client or making API requests. This follows Anthropic's convention of naming models consistently across documentation, API endpoints, and SDKs. Developers upgrading to v0.120.0 will find Claude Opus 5 available as a drop-in replacement for earlier Opus variants, though they should evaluate performance and cost characteristics for their specific workloads before migrating production traffic.

Dynamic Tool Addition and Removal

A notable enhancement in this release is the introduction of tool addition and removal blocks, along with corresponding tool_change events. Historically, tool definitions were static—specified once at the beginning of a conversation and remaining constant throughout.

This update enables more sophisticated workflows. A developer might begin a conversation with a basic set of tools, then introduce additional capabilities mid-conversation as context demands. For instance, an AI assistant handling customer support might initially expose only general information retrieval tools, then dynamically add refund or escalation tools once a customer's issue becomes more severe.

The tool_change events provide visibility into these modifications, allowing applications to log, monitor, or react to tool availability changes. This is particularly valuable for debugging, auditing, and understanding model behavior in complex multi-turn scenarios. The event-based architecture aligns with Anthropic's broader approach to providing granular observability into Claude's reasoning process.

Token Credit Fallbacks and Server-Side Defaults

The third major enhancement addresses token economics in fallback scenarios. In production environments, organizations often configure fallback strategies—if one model hits rate limits or encounters errors, automatically attempt with an alternative model or configuration.

Previously, managing credit token types across fallback scenarios required careful client-side orchestration. V0.120.0 expands the range of client-side fallback credit token types, providing developers more flexibility in how they structure fallback logic. Simultaneously, the SDK now supports server-side fallback defaults, shifting some complexity from client implementations to the API platform itself.

This dual approach offers benefits: developers who want fine-grained control over fallback behavior can implement it client-side with expanded options, while teams preferring simpler configurations can rely on sensible server-side defaults. This patterns reduces boilerplate code and makes token management more intuitive, particularly for smaller teams or applications with straightforward failover needs.

Practical implications

For developers actively using the Python SDK, upgrading to v0.120.0 enables access to Claude Opus 5's capabilities immediately. No code changes are required for basic adoption—simply specify the new model identifier in API calls.

Organizations leveraging tool use should assess whether dynamic tool management unlocks new application architectures. Scenarios like conditional tool exposure, progressive disclosure of capabilities based on conversation context, or dynamic workflow orchestration become more tractable with native SDK support.

For infrastructure and platform teams managing Claude deployments at scale, the enhanced token fallback mechanisms simplify cost optimization strategies. Server-side defaults reduce operational overhead while maintaining flexibility for sophisticated use cases.

What happens next

As the Python SDK evolves, expect continued alignment between API capabilities and SDK abstractions. Anthropic typically maintains consistency across SDKs (JavaScript, Python, etc.), so similar features will likely appear in other language implementations in coming releases. Developers should monitor release notes when major features ship to ensure they're not missing capabilities available in sibling SDKs.

For teams considering upgrades, testing Claude Opus 5 in non-production environments first is advisable to understand performance characteristics and cost implications specific to your workloads. The tool management enhancements warrant evaluation if your applications could benefit from more dynamic, context-aware tool exposure patterns. This article does not contain affiliate links.