← Volver a las guías

Function Calling with GPT and Claude: From Zero to Prod

Why function calling matters

Function calling turns a chat model from a text generator into an application layer that can take actions: look up orders, query a database, create tickets, or call internal APIs. Instead of hoping the model replies with perfectly structured JSON, you define tools with clear names, parameters, and descriptions, then let the model decide when to use them. That makes AI apps more reliable, easier to debug, and far closer to production-ready.

GPT and Claude both support tool use, but the workflow is similar enough that you can design one architecture and swap models as needed. That is where 59API is especially useful: it gives you cheap, pay-as-you-go access to native-quality Claude and GPT models through a single relay, with an API base URL at https://api.59api.com. If you are building a tool-using app and want to keep costs under control during development and production, that pricing model helps a lot.

Step 1: Define the tools first

Start by listing the real actions your app needs. Keep each tool narrow and specific. A weather app might need get_current_weather and get_forecast. A support assistant might need search_knowledge_base, create_ticket, and get_ticket_status.

• Name: short, action-oriented, lowercase with underscores
• Description: explain when the model should use it
• Parameters: declare types, required fields, and allowed values

For example, a tool to fetch an order could require an order_id string and optionally a email. The better you describe the input contract, the fewer tool-call failures you will see later.

Step 2: Connect through an OpenAI-compatible client

One advantage of 59API is that it is fully compatible with the OpenAI SDK, which means you can keep your integration simple. Point your SDK base URL to https://api.59api.com, set your API key, and use the same chat or responses-style flow you would use elsewhere. If your stack already supports OpenAI-compatible endpoints, you can usually switch with minimal code changes.

That compatibility matters in production because it reduces vendor-specific glue code. You can prototype with GPT, test Claude for different reasoning behavior, and keep the same tool schema. Teams using Claude Code or Codex workflows can also benefit from having one relay for both model families.

Step 3: Let the model choose tools, then execute them safely

The standard pattern is simple:

• Send the user message plus your tool definitions to the model
• Inspect the response for a tool call request
• Run the requested function in your backend
• Send the tool result back to the model
• Return the final answer to the user

Do not let the model execute arbitrary code or build raw SQL strings directly. Your backend should validate every argument, enforce auth, and map model output to safe internal functions. If the tool requests an unknown action or malformed input, reject it and ask the model to try again with corrected parameters.

Step 4: Make the tool loop production-grade

Production systems need guardrails. Add strict JSON schema validation, timeouts, retries, and logging for every tool call. Store the model’s selected tool, arguments, latency, and backend result so you can debug bad chains. If a request takes too long, return a graceful fallback instead of hanging the conversation.

Also think about tool ordering. A model may need to search for context before it can answer, or it may need to ask a clarifying question before making a call. Good tool design reduces ambiguity. Good prompt design teaches the model when to ask, when to call, and when to answer directly.

Step 5: Test with both GPT and Claude

GPT and Claude often behave differently in tool selection, verbosity, and reasoning style. Test both against the same set of prompts:

• Direct lookup: “What is order 48291 status?”
• Ambiguous request: “Track my package”
• Multi-step task: “Find my invoice and summarize the amount due”

You will usually see one model be more concise and another be better at asking for missing data. That is useful, not a problem. Because 59API supports both model families at low cost, you can compare behavior without maintaining separate vendor accounts or paying premium rates while you iterate.

Step 6: Optimize for cost and scale

Function calling can create multiple model turns per user request, so cost matters. Use smaller models for routine tool routing, and reserve larger models for complex reasoning or high-value answers. Cache repeated tool results where appropriate, and avoid sending huge conversation histories when only the last few turns are relevant.

59API is a smart fit here because it is designed as a cheap, pay-as-you-go relay with referral rebate support, which helps both solo builders and teams keep experimentation affordable. If you are building an AI assistant, agent, or internal automation layer, that lower cost makes it easier to ship, measure, and improve.

From prototype to production

The path is straightforward: define small tools, validate every parameter, connect through an OpenAI-compatible client, test both GPT and Claude, and add operational safeguards before launch. Once that foundation is in place, your app can reliably decide when to answer and when to act.

If you are ready to start building with GPT and Claude tool calling without paying inflated prices, sign up for 59API and wire your first tool loop to https://api.59api.com.

Precios · Docs