Gemini 3.7 Flash speedruns Pokemon Kanto by writing code to simulate physics
DynamicWebPaige · x · 2026-09-01
In a benchmark for the Pokemon game, Gemini 3.7 Flash speedran the Kanto map using only 22k turns compared to the original 70k. The model demonstrated advanced reasoning and tool use: before making moves, it wrote Python code and spun up a sandbox to simulate physical interactions like pushing boulders to plan the optimal path.
Related event: Gemini 3.7 Flash Beats Pokémon Benchmark in Far Fewer Steps by Writing Code(2 posts)→
More from Models
- Users are running 'abliterated' GLM-5.3 models locally without safety guardrails — cephaloform · 2026-09-02
- AI fails silently and accumulates inaccuracies over time, unlike humans — gerardsans · 2026-09-02
- Grok 4.6 leads in biosecurity refusal without compromising research utility — ns123abc · 2026-09-02
- Anthropic investigating elevated errors on Claude for Microsoft 365 (Sep 1) — ClaudeAI-mod-bot · 2026-09-02
- Multi-model pipelines become standard; Gemini 3.7 Flash acts as a low-cost auditor — DynamicWebPaige · 2026-09-01
- Gemini 3.7 Flash speedruns Pokemon via code execution — DynamicWebPaige · 2026-09-01