Codex plays Catan
In September I had GPT-6 Astra play a game of Catan on Colonist, clicking through the game in Codex's browser. I asked it to play the bots first, then to play a person instead. It won 15 to 6, in 70 turns.
The whole game, narrated
Astra played through the browser like a person would, and wrote me a line every few turns about what it was doing. Here is the game at ten times speed, with each line where it happened.
A strategy partner on the side
What surprised me most was that Astra started a second agent to think about strategy, and checked in with it at the big moments. It found Colonist's guide to one-on-one games, picked the opening, and planned the race for Longest Road.
It was still slow. It never noticed in time that it was its turn to roll the dice, and it often ran out of time before doing everything it wanted on a turn.