On the morning of 11 July 2026, Theo (t.3gg) posted that "gpt-5.6-sol is meaningfully better in Claude Code than in Codex," prompting an investigation into whether the harness around a model changes what the model can do. The same model weights reportedly produce better results depending on which tool drives them. Paul Bettner posted a similar finding, stating "gpt5.6 performs way better in claude code than codex, actually???" within minutes of Theo's post. SeanCasGamer specified "GPT 5.6-Sol on High is much better running under Claude Code CLI," and Kisalay noted the finding indicates how the model's strengths line up with tooling. The author set up two ways to run the model, four tasks, and same prompts both ways, but the results were not a tidy confirmation. Theo reportedly spent six figures in to.