Function calling
Also written: tool use, tool calling, Programmatic tool calling, Asynchronous tool calling, Async tool calling, Frontier-intelligence function call
AI models can trigger external tools or code, translating requests into function calls with specific arguments.
Function calling lets AI models interact with the outside world. Instead of just generating text, the AI can understand when a task requires using another piece of software or running a specific command.
For example, if you ask an AI to book a flight, it can use function calling to send the details of your request (like destination and dates) to a flight booking system. The AI doesn't actually book the flight itself, but it knows how to ask the booking system to do it.
Function calling enables large language models (LLMs) to act as agents by producing structured output that specifies a function to be called and its arguments. This allows the LLM to integrate with external tools, APIs, or custom code, extending its capabilities beyond text generation. The model identifies user intent that maps to a predefined function and formats the call, which is then executed by an external system. The result is typically returned to the LLM for further processing or to formulate a response.
Function calling is a key feature for making AI models more useful by allowing them to interact with external software and services. Recent announcements highlight how companies are integrating this capability to enable AI agents to perform actions, manage costs for complex tasks, and offer more dynamic user experiences. This points to a trend of moving AI beyond just generating text to actively executing commands and solving problems.
- Stop Overpaying for Intelligence | DevDay 2026OpenAI · OpenAI on YouTube ·
- Discover local models in GitHub Copilot CLIMicrosoft · GitHub Changelog: Copilot ·
- GPT-6.1 SolOpenAI · Model Documentation ·
- OpenAI pauses training of its ‘most capable models’OpenAI · The Verge AI ·
- GPT-6 LunaOpenAI · Model Documentation ·
- GPT-6 SolOpenAI · Model Documentation ·
- Give your voice AI agent real abilities (tool calling)Google · Google Cloud Tech on YouTube ·
- ChatGPT co-creator launches a new kind of AIOpenAI · The Rundown AI ·
- Gemini 2.5 FlashGoogle · Gemini Models ·
- spark Gemini 3.5 FlashGoogle · Gemini Models ·
- Builder Bootcamp: Realtime - EventOpenAI · OpenAI Academy ·
- Introducing GPT-6 Astra for developersOpenAI · OpenAI on YouTube ·
- Updates for developers building with voiceOpenAI · OpenAI Developers Blog ·
- The builder’s guide to GPT‑5.6OpenAI · OpenAI News ·
- Function CallingxAI · xAI Docs: Guides ·
- Tools OverviewxAI · xAI Docs: Guides ·
- Introducing computer use in Gemini 3.5 FlashGoogle · Google DeepMind Blog ·
- Introducing GPT-5.4 mini and nanoOpenAI · OpenAI News ·
- Gemini API Changelog: February 19, 2026Google · Gemini API Changelog ·
- Netomi’s lessons for scaling agentic systems into the enterpriseOpenAI · OpenAI News ·
- Improved Gemini audio models for powerful voice experiencesGoogle · Google DeepMind Blog ·
- How Podium is arming 10,000+ SMBs with AI agentsOpenAI · OpenAI News ·
- How Scout24 is building the next generation of real-estate search with AIOpenAI · OpenAI News ·
- OpenAI o3-miniOpenAI · OpenAI News ·
- Expanded legal protections and improvements to our APIAnthropic · Anthropic News ·
- Introducing Structured Outputs in the APIOpenAI · OpenAI News ·
- GPT-4o mini: advancing cost-efficient intelligenceOpenAI · OpenAI News ·
- Making education data accessibleOpenAI · OpenAI News ·
- New models and developer products announced at DevDayOpenAI · OpenAI News ·
- GPT-3.5 Turbo fine-tuning and API updatesOpenAI · OpenAI News ·
- Function calling and other API updatesOpenAI · OpenAI News ·
- Anthropic Claude
- Anthropic's Claude can trigger other software tools or actions based on user requests, making it more capable of performing tasks.
- Google Gemini
- Google Gemini is enabling voice AI agents to act on user requests instantly through tool calling, making them more responsive and useful.
- Microsoft Copilot
- Microsoft Copilot is allowing users to integrate local AI models within their workflow, suggesting a way to leverage external capabilities.
- OpenAI ChatGPT
- OpenAI's models are being developed with function calling to enable them to translate natural language into structured calls for external tools, helping to manage costs and perform complex tasks.
- Agents
- Agents are broader systems that may use function calling as one of many
- API Integration
- API integration is the general process of connecting software;
- Retrieval Augmented Generation (RAG)
- RAG retrieves information to inform the LL
- Code Generation
- Code generation produces code as output; function calling produ
The concept of enabling language models to call external functions emerged from research into making LLMs more useful and capable of interacting with the real world. Early implementations and research papers explored ways for models to access external knowledge and execute actions. Source ↗
Use a chatbot interface that supports function calling, like ChatGPT or Google Gemini. Ask it to perform a task that requires external information or action, such as 'What's the weather in London?' or 'Set a timer for 5 minutes.' Observe if the AI indicates it's using a tool or provides a response that suggests it has interacted with an external service.
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- HeuristicThe loop: declare, decide, execute, return for AI agentsUnderstand the core cycle of how AI agents with tool calling operate. · 5 min
- Code snippetAgent Development Kit tool calling structure and callbackDemonstrates how to define tools and implement a callback to intercept tool calls before execution in Gemini agents. · 15 min
- Setting tipDiscover and add local Ollama models in GitHub Copilot CLIUse the /model command to find and add local Ollama models to GitHub Copilot CLI without leaving your workflow. · 5 min
- Setting tipEnable function calling in GPT-6 Luna Chat CompletionsUse Chat Completions with `reasoning_effort` set to 'none' for function calling. · 1 min
- RuleKeep every tool instant and put policy in one callbackCentralize policy checks in a single callback for AI agent tool execution. · 10 min
Questions people ask
- Does the AI run the code itself?
- No, the AI generates a request to call an external tool or code. A separate system then executes that call.
- How does the AI know which function to call?
- Developers define a set of available functions and their descriptions for the AI. The AI uses its understanding of the user's request to match it to the most appropriate function.
First seen on Subvolts Jun 13, 2023. Updated Oct 14, 2026. Written by an AI model from the items above and reviewed against them weekly; report an error.