Anthropic has introduced major new features to its Claude Developer Platform, significantly enhancing how its AI models, particularly the new Claude Opus 4.5, interact with external tools. These updates aim to make AI agents more efficient, capable, and easier for developers to build. The changes, announced around November 24-25, 2025, include three key beta features: the Tool Search Tool, Programmatic Tool Calling, and Tool Use Examples. Developers can now build more sophisticated AI applications that handle complex, multi-step tasks with greater precision and lower operational costs.
Smarter Tool Discovery Reduces AI Costs
One of the most significant new additions is the Tool Search Tool, designed to tackle the problem of "context bloat" in AI models. Previously, large language models often needed to load all available tool definitions upfront, regardless of whether they were immediately relevant to the task. This process consumed valuable "token" space within the model's context window, limiting the amount of information Claude could process for a given request.
The Tool Search Tool changes this by allowing Claude to dynamically discover and load tools only when they are needed. This "just-in-time" approach means Claude only sees the tools relevant to the current task. Internal testing by Anthropic showed dramatic improvements in efficiency. For instance, when working with a set of 58 tools, the Tool Search Tool reduced context size by an impressive 85%, dropping from approximately 77,000 tokens to just 8,700 tokens. This reduction frees up substantial space for longer conversations and more complex reasoning. Furthermore, this dynamic tool discovery led to significant accuracy gains on internal evaluations, with Claude Opus 4 improving from 49% to 74% and Claude Opus 4.5 improving from 79.5% to 88.1% when the Tool Search Tool was enabled. The feature also helps to preserve a larger context window, maintaining 191,300 tokens compared to 122,800 with the traditional method.
Code-Driven Automation for Complex Workflows
Anthropic also rolled out Programmatic Tool Calling, a feature that enables Claude to orchestrate multiple tools through Python code rather than requiring individual API calls for each step. In traditional setups, each tool invocation often requires a separate "inference pass" by the AI model, leading to increased latency and token usage as intermediate results fill the context window.
With Programmatic Tool Calling, Claude can generate a block of Python code that calls several tools, processes their outputs, and manages the flow of information. This means that instead of many back-and-forth interactions with the model, Claude can execute a series of actions within a single code block. This new method significantly reduces latency, as it can eliminate over 19 inference passes when Claude orchestrates 20 or more tool calls at once. It also cuts down on token usage, with average usage dropping by 37% on complex research tasks, from 43,588 to 27,297 tokens. The ability to write explicit orchestration logic, including loops, conditionals, and data transformations, also leads to improved accuracy. Internal knowledge retrieval improved from 25.6% to 28.5%, and GIA benchmarks saw an increase from 46.5% to 51.2%. This capability allows Claude to handle messy data and conditional logic more effectively, focusing on actionable results.
Learning Proper Tool Usage from Examples
The third new beta feature, Tool Use Examples, provides a standardized way for developers to demonstrate how to effectively use a given tool. While JSON schemas define the structural validity of tool inputs, they often fall short in conveying usage patterns, such as when to include optional parameters, which combinations make sense, or specific API conventions.
By providing concrete examples, developers can help Claude learn the correct and most effective ways to interact with tools. This direct instruction improves Claude's ability to make accurate and appropriate tool calls, leading to more reliable and predictable outcomes in complex agentic tasks. This feature helps bridge the gap between abstract tool definitions and practical application, ensuring Claude can use tools as intended by developers.
Broader Impact and Efficiency Controls
These advanced tool use capabilities are tightly integrated with Anthropic’s latest flagship model, Claude Opus 4.5, which the company positions as a leading AI model for coding, agents, and computer use. Claude Opus 4.5 has shown state-of-the-art performance on real-world coding benchmarks, achieving an 80.9% score on SWE-bench Verified. The new model excels at complex, multi-step tasks that require sustained reasoning and autonomous execution.
To further enhance developer control, Anthropic also introduced a new "effort parameter" on the Claude API. This parameter allows developers to manage the trade-offs between efficiency and capability. With a "medium effort" setting, Claude Opus 4.5 can match the performance of Sonnet 4.5 while using 76% fewer output tokens. At a "high effort" setting, it can surpass Sonnet 4.5's performance by 4.3 percentage points, using 48% fewer tokens. This flexibility allows developers to optimize for speed and cost or maximize capability based on the specific requirements of their tasks. The Claude Developer Platform is being redesigned around this modularity, giving developers more precise control over efficiency, context, and tools.
These updates signify a major step forward for building advanced AI agents. They enable Claude to work seamlessly across vast libraries of tools, handle complex orchestration logic, and learn tool usage more effectively. Developers can now create more robust, efficient, and intelligent AI applications that automate critical workflows and solve challenging problems across various industries.
