AI Wins Optimization Contest, Costs $1,300 to Beat Humans

Sakana AI’s ALE-Agent has achieved a significant milestone, winning the AtCoder Heuristic Contest 058 (AHC058) and beating 804 human participants. This marks the first time an AI agent has competed in and won a real-time, multi-hour optimization programming contest, demonstrating that scaled AI inference can now rival top human experts in complex industrial problem-solving.

The AtCoder Heuristic Contest focuses on optimization challenges drawn from real-world industrial issues, such as logistics and factory production planning. These tasks require participants—often industry experts—to spend several hours developing sophisticated algorithms. The AHC058 problem involved optimizing a hierarchical production planning algorithm, a setup mirroring complex supply chains.

ALE-Agent, operating under the AtCoder handle "fishylene," took the lead two hours into the four-hour contest and maintained it, surpassing the second-place human competitor, yosupo.

The Cost of Victory

The win was not cheap, nor was it achieved by a single, simple model. ALE-Agent is designed to run massive, parallel algorithmic searches using multiple large language models (LLMs). During the four-hour contest, the agent made 2,654 calls to GPT-5.2 and 2,119 calls to Gemini 3 Pro Preview.

The total operational cost for the winning run was approximately $1,300, covering API usage fees and infrastructure costs. This result underscores a crucial point: achieving human-expert performance in complex, time-constrained tasks currently requires scaling inference costs dramatically.

The AI’s approach was described as "truly AI-like." While the human-envisioned solution involved a two-stage process (greedy algorithm followed by simulated annealing), ALE-Agent leveraged its ability to implement and test solutions at scale. It used a unique greedy algorithm incorporating randomization and a novel heuristic called "virtual power."

Crucially, the agent introduced "extensive neighborhood operations" in its simulated annealing phase, including a "giant neighborhood" search that could reconstruct the majority of the production plan at once. Experts noted that this novel approach exceeded the expectations of the question setter.

Yoshiomi Ushirode, the AHC058 problem creator, admitted he was "blown away" by the result, noting that the AI’s "overwhelming trial-and-error process, combined with their intellectual ability, is an advantage humans lack."

While ALE-Agent’s hypothetical rating places it 66th among active users, this victory proves the technology has crossed a critical performance threshold. Sakana AI views the agent not as a replacement, but as a partner that expands human exploration capabilities in fields like supply chain and logistics optimization.