A new planning benchmark shows frontier AI models outperforming humans at Wikipedia-link navigation — right up until the routes get complicated.
Researchers released LLM-WikiRace, a benchmark where models must travel from one Wikipedia article to a target by following hyperlinks one step at a time, with no shortcuts or backtracking. On easy difficulty, Gemini-3, GPT-5, and Claude Opus 4.5 all reached superhuman performance levels. Hard mode told a different story: Gemini-3, the best performer, completed only 23% of hard games. The researchers found that broad world knowledge carries models through simpler routes, but once paths grow long and indirect, planning and long-horizon reasoning become the deciding factors — and that is where current models fall short.
The specific failure mode matters more than the headline number. When models hit a dead end, they do not regroup: they loop, revisiting the same dead-end pages rather than reconsidering their approach. For anyone building agentic systems that are supposed to handle obstacles without a human stepping in, that is a concrete data point about where the autonomy actually breaks.
WikiRace is a deliberately simple domain, which is precisely its value as a test. If the best available model can solve roughly one in four hard games, the gap between "superhuman on easy" and "reliable planner" is still large enough to drive a truck through.