GPT-5 New Developer Features — Tutorial [Jupyter Notebook]

The GPT-5 API does A LOT more than the web platform.

Let me show you.

Check out the video on YouTube

Course Update: My voice is gone…

… after many many hours of filming presentations for these slides.

Excited to bring these to you in September for anyone who has purchased the course on zazencodes.com

Dropping this on zazencodes.com in September for anyone who has purchased the course.

Topics from this week’s video

• Introduction

  • Hands-on demo of new developer features in GPT-5

  • Features covered: reasoning strength, verbosity, freeform function calling, context-free grammar

  • Target audience: developers, engineers, AI practitioners

• Responses API vs Chat Completions

  • Old way: Chat Completions endpoint

  • New way: Responses API

    • Output accessed through output_text

    • Benefits:

      • Supports passing chain-of-thought between turns

      • Reduces reasoning tokens → cost savings

• Reasoning Strength

  • Levels: minimal, low, medium (default), high

  • Minimal reasoning (new in GPT-5)

    • Useful for simple questions

    • Example: quantum mechanics test problem

• Verbosity Flag

  • Options: low, medium, high

  • Controls response length/detail

  • New feature in GPT-5

• Freeform Function Calling

  • Allows executing tools without strict argument schemas

  • Example: arbitrary bash command tool

  • Compared with structured function calling:

    • Structured requires parameter schema (e.g., horoscope function)

    • Freeform simplifies tool definitions

• Context-Free Grammar (CFG)

  • Forces model outputs to match specified grammar

  • Example 1: Regex grammar for phone numbers

    • Validates output as phone number or “unknown”

    • Verified output extracted from tool call

  • Example 2: Emoji-only grammar

    • Summarize quantum computing in emoji

    • Output validated as emoji sequence

  • Example 3: Lark grammar

    • Generates valid Postgres SQL queries

    • More expressive than regex, ensures valid database syntax

  • Use case: contract-like enforcement of acceptable outputs

---

“Which would have advanced the most at the end of a month,” he posited, “the boy who had made his own jackknife from the ore which he had dug and smelted, reading as much as would be necessary for this—or the boy who had ... received a Rodgers' penknife from his father?”

Henry David Thoreau, Walden (1854)