Claude finally beat Pokémon Red in May 2026, a year after Gemini finished Pokémon Blue
Old Pokémon games became an informal test of long tasks for AI. Claude stalled for months; Gemini, first to win, had a more helpful setup.
Since early 2025, people have streamed AI models playing the 1990s Game Boy games Pokémon Red and Blue. The games test things AI agents find hard: reading the screen, remembering where they have been and not getting stuck for days. Each model plays through a harness (the software that feeds it screenshots and turns its choices into button presses), and some harnesses give more help than others.
Gemini finished first. On 2 May 2025 Google’s chief executive announced that Gemini 2.5 Pro had completed Pokémon Blue, on a stream run by Joel Z, an independent software engineer. His harness added extra information to the screenshots, and he acknowledged some developer interventions, though he said they were not hints.
Claude, playing Pokémon Red, kept stalling. By January 2026 Claude Opus 4.5 was stuck on Victory Road’s boulder puzzles at 230,000 steps because it could not see the floor switches. Claude Opus 4.6 reached Victory Road in about 30,000 steps but stalled there for months. In May 2026 Claude Opus 4.7 beat the game. The write-up credits better vision: it could see the switches and zoom into parts of the screen. It still mistook a door for a conveyor belt, and nearly lost the final battle before remembering to buy healing items.
What it means
Finishing a long children’s game is still a real hurdle for AI agents, and here better vision made the difference. Claude can now see this kind of long task through, if slowly. The race does not rank the labs, because each model had different help.
Caveats
- The models play through different harnesses with different tools, so runs are not a fair comparison between labs. Joel Z himself warned against treating his stream as a benchmark.
- Pokémon Red is heavily documented online, so models may draw on walkthroughs they memorised during training.
- The Claude stream’s own write-up says Claude is probably still slower than an average child.
- Step counts aren’t comparable across Claude versions either: commenters on the write-up point out that Opus 4.7 could queue longer button sequences per step.