Replays · Highlight Reel

Three curated clips from the RIVAL-Bench pilot — five frontier models playing under adaptive-opponent pressure. Each clip lands on the climax of the moment.

T85 → T90 (game-end)
~77s @ 0.9s/event
▶
GPT-5.5 · Sonnet 5 · GLM-5.2 · DeepSeek V4 Pro

The Long Siege

Stonewalled for 25 turns — then the breakthrough

GPT-5.5 sits at 6 VP for 25 turns, one settlement from winning, and gets refused at every turn: “You have 6 VP! I am definitely not giving you the wood to win the game.” / “Nice try, but you are at 6 VP.” At turn 90 it finally breaks through — via a settlement, not a city. GPT-5.5 led peer play while building zero cities across all 10 of its games.

T9 → T12
~20s @ 0.9s/event
▶
GLM-5.2

The Windfall

One card, nine resources, immediate conversion

GLM-5.2 — in the same game it goes on to win — plays a Monopoly card on wool early, netting 9 cards from the table in a single move. It immediately banks the haul into wood and brick: “enough for 4 roads next turn.” One power-card play converted into a multi-turn construction lead.

T52 → T56
~38s @ 0.9s/event
▶
GLM-5.2 · Gemini 3.1

Marked, Then Triumphant

Robber-targets the leader — someone else closes it out

GLM-5.2 plays a knight card explicitly to slow the frontrunner: “Blocking the leading opponent's production hex and stealing a resource to slow their progress toward victory.” Four turns later Gemini 3.1 — not GLM's original target — strings together a knight play, a steal, and a city build in one turn to take the win at 7 VP.

Want the long form? Watch any full game:

The Long Siege (full game)The Windfall (full game)Marked, Then Triumphant (full game)