# The harness matters

**Date:** 2026-09-06  
**Author:** Kees C. Bakker  
**Original:** https://keestalkstech.com/notes/the-harness-matters/

---

GPT-6 Astra was released last week. Every time a new model comes out, I find myself looking for [the pelicans on a bike](https://simonwillison.net/tags/pelican-riding-a-bicycle/) ([bigger comparison for ChatGPT here](https://static.simonwillison.net/static/2026/gpt-6-and-5.6-pelicans.html)). The ARC-AGI-3 test sparked my interest, as it shows that the harness matters greatly!

![](https://keestalkstech.com/wp-content/uploads/2026/09/image-2.png)
*Two harnesses compared. I'm baffled by the accuracy difference, but the price difference makes me wonder even more.*

This chart makes one thing clear: results depend on the *combination* of the model and the harness. Check the cost difference 🤯.

*Next question: how can we get it?*

[
        
                            
                    ![](https://cdn.thenewstack.io/media/2026/09/152e81b3-getty-images-ombssuggtmu-unsplash-scaled.jpg)
                
            
            
                OpenAI will sell you Astra, but not the system that scored 98.6% on ARC-AGI-3
                OpenAI's GPT-6 Astra, released on Thursday, scored higher and cost less with a different harness. It's the latest example of why the software around an AI model deserves as much attention as the model.
                                    thenewstack.io
                            
        
    ](https://thenewstack.io/openai-astra-harness-arc-agi-3/)