Google twins beat Pokémon Blue (with little help)

Google’s The most expensive AI model It seems to have passed a major stage: beat a 29-year video game.

Last night Google Sundar Pichai CEO Posted triumphantly on x“What a finale! Gemini 2.5 Pro just finished Pokémon Blue!”

In order to be clear, Twins play a live pokemon was created by (in his words) “30 -year -old software engineer who is not related to Google” that passes Joel Z.S But Google leaders have cheered the effort.

For example, Logan Kilpatrick, the leading product for Google Ai Studio, Posted last month That twins “made a lot of progress in completing Pokémon” and “won their 5 badge (the next the best model has only 3 so far, though with a different harvest agent),” leading Pichai to joke“We work on API, artificial Pokémon Intelligence :)”

Why Pokémon? Back in February, The anthropian emphasized progress The fact that his Claude AI models are made at Pokémon Red, writing that “Extended Thinking and Claude Training” gives him a “main impetus” to “more unexplained” tasks, such as a classic game. (“Pokémon Red” and “Blue” are different versions of Gameboy title First published in 1996 and bound by the long -time Pokémon franchise). There is even Claude plays channel Pokemon Twitch This is Joel Z cited as an inspiration.

Despite his progress, Claude seems to have not yet defeated Pokémon Red. Does this mean that the twins are objectively better in the game? On your Twitch Joel Z page calls on viewers: “Please don’t consider this a benchmark for how well LLM can play. You can’t make direct comparisons – Gemini and Claude have different tools and receive different information.”

Both AI models need help to play the game – here’s where The aforementioned agent is confronted Sign in by providing screenshots of the game, covered with additional information, which allows the model to decide how to react (which may include a call to specialized agents) and then press the button that corresponds to the AI ​​instruction.

Joel Z admitted that there were other “DEV interventions” to help twins finish the game, but insisted that this was not cheating.

“My interventions are improving the overall decision -making and reflection abilities,” he says. “I do not give specific hints – there are no instructions or direct instructions for specific challenges such as MT. Moon. The only thing that approaches even nearby is to inform twins that he has to talk with a rocket to get a lifting key that was later fixed in a yellow pokemon.”

Plus, he said, “Gemini plays Pokémon is still actively developing and the frame continues to develop.”