From an initial testing of my personal benchmark it works better than Gemini 2.5 pro.

My use case is using Gemini to help me test a card game I'm developing. The model simulates the board state and when the player has to do something it asks me what card to play, discard... etc. The game is similar to something like Magic the Gathering or Slay the Spire with card play inspired by Marvel Champions (you discard cards from your hand to pay the cost of a card and play it)

The test is just feeding the model the game rules document (markdown) with a prompt asking it to simulate the game delegating the player decisions to me, nothing special here.

It seems like it forgets rules less than Gemini 2.5 Pro using thinking budget to max. It's not perfect but it helps a lot to test little changes to the game, rewind to a previous turn changing a card on the fly, etc...