Member-only story
One of OpenAI’s launch customers ran the exact same agent on the exact same tools and burned 63.5% fewer tokens. The only thing that changed was that GPT-5.6 stopped calling its tools one at a time and started writing code to orchestrate them. That single shift is the most important thing in this release, and almost nobody is talking about it.
On Thursday, July 9, OpenAI took GPT-5.6 public — a three-tier family called Sol, Terra, and Luna — and shoved it into the ChatGPT Work agent and GitHub Copilot on the same day. The headlines went to the benchmarks and the pricing. But buried in the Responses API is a feature called Programmatic Tool Calling that quietly rewrites how agents use tools. And if you’ve ever watched an agent chew through 150,000 tokens to answer one question, this is the fix you’ve been waiting for.
I spent a day reading the docs, the customer numbers, and Anthropic’s research from last November — because, as we’ll get to, OpenAI didn’t invent this idea. But it may have just made it usable for everyone.
The problem: the tool-calling loop is a token furnace #
Here’s how tool calling has worked since function calling shipped. The model wants to use a tool, so it emits a structured request. Your code runs the tool, then feeds the entire result back into the model’s context. The model reads it, decides on the next tool, emits another request…