Recently, I conducted a multi-model competition, and I made an interesting discovery.

For the same problem, I asked Claude and Codex to independently come up with solutions.
Codex missed one thing: my strategy is distributed across more than 20 independent processes, and its default assumption that "all components run in the same place" made the solution invalid. Claude immediately spotted this issue.
On the other hand, Claude missed another thing: it was fixated on creating new independent modules, but I already had a complete framework ready to be integrated—just needed to add one field. Codex picked up on this.
The most interesting part was the next step: I asked both models to review each other's final solutions. Both subconsciously anchored on "the other side's existing framework," which caused them to overlook a boundary case—only discovered when I manually ran production data.
So now, the habit is: for each round of competition, I set independent success criteria and prohibit one model from seeing the other's draft. The result is two separate drafts with non-overlapping blind spots, which together form a more complete picture.