AlphaCode 2 is a competitive programming system from Google DeepMind, described in a technical report released on December 6, 2023, by the AlphaCode Team. It replaced the bespoke models of the original AlphaCode with versions of Gemini fine-tuned for code, wrapped in a pipeline that generates many candidate solutions, filters them against the problem’s example tests, clusters the survivors, and selects a small set to submit.
On a held-out set of recent Codeforces contests, AlphaCode 2 solved about 43 percent of problems within ten attempts, nearly double the original AlphaCode’s 25 percent, and reached the 85th percentile on average among participants. The report also emphasizes efficiency: AlphaCode 2 needed roughly 100 samples to match what the first system achieved with a million, making it far more sample efficient even as the underlying model did most of the heavy lifting.
AlphaCode 2 is a concrete example of how a strong general-purpose foundation model, plus a sampling-and-filtering harness, can outperform a specialized earlier system. For a business reader it illustrates a recurring pattern in modern AI: capability gains often come from better base models combined with smart orchestration, not just from training a new model from scratch.