gpt-6POST
Opus 5.5 vs GPT-6 Sol tested on code audit, with harness as the key variable
The author ran Opus 5.5 and GPT-6 Sol on a codebase where they had already found multiple critical bugs, concluding the harness matters more than the model.
OxElliot 🦀 💀
@elliot_1698Archived 2026.09.27Originally posted 2026.09.26
OxElliot 🦀 💀 @elliot_1698Originally posted 2026.09.26
1/5 The harness matters more than the model. I tested Opus 5.5 @AnthropicAI and GPT-6 Sol @OpenAI on a codebase I'd already audited and found multiple crits. The results matched what I'd seen with older models from both companies.👇🧵
x.com/elliot_1698/status/2103771511039627636Archived 2026.09.27 09:53