Codex + GPT-5.5 is actually the best model for STS2MCP now

[Gennadiyev]

Opus is getting a little bit unstable recently as always, so I opt for GPT-5.5.

I ask Codex to make API calls, bypassing the entire MCP. Using GPT-5.5 with high effort.

It's faster for most decisions than Opus, and learns to multi-play cards in a single PowerShell call without prompt; and sleeps before fetching the refreshed state.

I really hope that this could be the model that beats STS2 zero-shot.

[Gennadiyev]

It uses json mode, and still costs less due to the non-blocking sequence when playing cards.