AlphaCode 2

AlphaCode 2, unveiled by Google DeepMind alongside Gemini in December 2023, rebuilt the original AlphaCode on fine-tuned Gemini Pro models. Rather than a single model, it is a pipeline: policy models sample up to a million candidate solutions per problem, then filtering, behavioral clustering, and a learned scoring model narrow them to about ten final submissions.

On 77 problems across 12 recent Codeforces contests with over 8,000 participants, it solved 43% of problems within ten attempts, nearly double the original AlphaCode's 25%, which DeepMind estimated at the 85th percentile of human competitors, between Expert and Candidate Master rank. It was research-only, never shipped as a product or API, and its techniques fed directly into DeepMind's later Gemini and AlphaProof reasoning work.