Early Reactions to GPT-6 Astra

It’s been about two days since GPT-6 Astra launched, and enough people have access now that useful early reports are starting to show up.

The strongest feedback I’m seeing is that Astra improved more on agency than intelligence.

Artificial Analysis currently gives Astra the same Intelligence Index score as GPT-5.6 Sol: 61. Its Coding Agent Index moves from 65 for Sol to 67 for Astra. That’s an improvement, but hardly the generational jump the GPT-6 name suggests. (Hacker News)

Strip away the launch branding, and the independent intelligence numbers look similar to the previous generation. (Reddit)

Matt Shumer describes letting Astra work through email, newsletter publishing, and advertising tools. Some tasks now run unattended, without him watching every click.

His larger experiments use what he calls a “Manager Loop”: one Astra instance coordinates the project while another handles implementation, spawning more agents where useful. He reports longer useful runs than with previous models. But he also says these runs can still get stuck in details and need careful coordination. (Something Big Is Happening)

Claire Vo reports something similar from a software development angle. She tested Astra on tasks that previous models had failed repeatedly, including a product feature, 3D applications, and a hardware project. Astra finally completed them. She also found browser use useful for QA against real applications, not just code generation. (Lenny’s Newsletter)

OpenAI reports 99.9% on ARC-AGI, which looks strong at first. But that number comes from OpenAI’s Provider Adapter harness, which preserves reasoning state between requests and compacts the conversation. Run the standard ARC-AGI harness instead, and the result drops. (Simon Willison’s Weblog)

We normally talk about model intelligence as though it belongs entirely to the model. Astra is another example of why that idea is misleading.

The same model can produce very different results depending on the harness around it. That harness decides how state is preserved, how context gets compacted, which tools are available, and how long the model can keep working.

There are plenty of complaints too.

Astra is expensive. (Reddit)

It’s also not the best model at everything. Early reviewers still prefer Claude for some visual and design work, and Artificial Analysis currently puts Claude Fable 5.1 ahead of Astra on its general Intelligence Index. (Something Big Is Happening)

So I’m not convinced this is “AGI”.