Bilibili launches AI arena with 100 models competing, GPT-6 on top

Bilibili launched an AI Infinite Arena on September 16 and published the first edition of its model leaderboard at the same time. According to a report from Securities Market Weekly, GPT-6 Astra took first place across evaluations from ten Bilibili creators and recorded the most first-place finishes, ahead of GLM-5.3 on that count, while domestic Chinese models filled three of the top five spots.

The format is deliberately different from a research benchmark. The arena is a hub that collects model evaluations produced by Bilibili creators across more than a hundred models, built from real-world tests in categories such as coding, reasoning, collaboration and knowledge. Creators set their own tasks, from real workflows and professional applications to playful scenarios, and competing models face the same question.

The first leaderboard already spans DeepSeek, Kimi, ChatGPT, Claude, Gemini, Doubao, Qwen, Hy and MiniMax, which puts open and closed models, and Chinese and overseas systems, on the same table.

Bilibili's stated context is scale. Watch time for AI knowledge content on the platform grew 72% year over year, and more than 190 million users watch AI-related content each month. Live streams, videos and scrolling comments match how AI enthusiasts discover and debate information, and a large body of evaluation content has already accumulated inside the community. The arena organizes that material rather than building an evaluation system from scratch.

Notably, the ranking updates in real time and stays open to creators across the site. That makes it a recurring evaluation entry point rather than a one-off launch, with samples continuously replenished from real usage scenarios.

The value of any arena-style leaderboard depends heavily on how evaluations are sampled, who scores them and how results are aggregated. What has been disclosed so far stops at the layer of creators setting the task and models answering the same question; the full scoring and aggregation methodology has not been laid out. For model developers, board position is increasingly a public marketing asset, and the top spot most of all, which is exactly why transparency about the evaluation will draw scrutiny.

Three things to watch: whether Bilibili publishes a fuller methodology and complete ranking, whether the top models hold their positions once the board updates continuously, and whether a creator-driven evaluation becomes a common reference point for Chinese users judging how capable a model really is.