GPT-6 Astra vs Opus 5.5 head-to-head on the same bug
On the same bug, un-nerfed Astra found and fixed the issue on its own, while Opus 5.5 needed two or three corrections to diagnose it but proactively suggested more ways to prevent recurrence.
Put GPT-6-Astra and Opus-5.5 head-to-head on the exact same bug: GPT-6-Astra (un-nerfed) @ xhigh: - 🎯 Found the bug on its own with high precision. - Implemented a fix unprompted (likely due to the Codex harness). Opus-5.5 @ xhigh: - 👀 Needed 2–3 corrections before landing on the right diagnosis. - Saving grace: Proactively proposed more ways to help prevent the bug, which Astra didn't. This is just one example. For this particular one, GPT-6-Astra performed "smarter" than Opus-5.5, although it went in to implement a solution when I didn't ask for it. It could also be the harness that caused the difference.