Bilibili Launches AI Infinite Arena: Real-World Model Evaluation Puts GPT-6 Astra on Top

Background

Bilibili (B站) has launched the "AI Infinite Arena" (AI 无限竞技场), a large-model evaluation platform where content creators (UP主) design real-world scenario tests. The first round results show GPT-6 Astra ranking first across 10 UP主 evaluations, beating GLM-5.3 most frequently. Chinese models occupy three of the top five positions.

AI Infinite Arena: Not Fixed Answers

Unlike traditional benchmarks with fixed questions and scoring, the arena lets UP主 from different domains create tests based on actual workflows — code, reasoning, collaboration, knowledge, and professional applications. Models face the same user-provided problems, enabling side-by-side comparison in realistic settings.

This approach better reflects user experience. For example, code ability can be tested by asking a model to understand an unfamiliar repository, fix a real bug, add missing tests, explain performance bottlenecks, or complete cross-file refactoring after multi-turn feedback. A single score cannot capture these nuances, while scenario-based evaluation reveals strengths and weaknesses directly.

Limitations are clear: question difficulty, prompts, context, judging criteria, and runtime environments may differ across tests. Winning a creative task only shows better performance on that specific task, not a comprehensive capability ranking.

Why GPT-6 Astra Topped the First Round

GPT-6 Astra won the most first-place finishes in the 10 evaluations and against GLM-5.3. The article analyzes a typical AI programming task as four stages: understanding the goal, generating a plan, invoking tools, and verifying results. A model suitable for daily development must sustain progress through failures and correct its plan based on test feedback, not just produce the prettiest code once.

Developers should focus on which task types a model wins. The article provides a breakdown of evaluation dimensions and key questions:

Code : Can it modify a real project and pass tests?

Reasoning : Can it maintain consistency across long chains?

Collaboration : Does it converge correctly after multi-turn feedback?

Knowledge : Can it distinguish facts, uncertainty, and hallucinations?

Agent : Can it stably call tools and meet acceptance criteria?

The leaderboard serves as a first-round candidate filter, but production selection must return to the team's own task suite.

Chinese Models Occupy Three of Top Five Spots

Three Chinese models in the top five indicates domestic models are competing beyond single Chinese Q&A scenarios, entering code, reasoning, and workflow comparisons. However, rankings reflect performance on the current question set only — not a permanent technical conclusion. Model versions, inference tiers, context lengths, tool capabilities, and pricing change; the live-updating leaderboard means today's ranking may shift quickly.

For teams, the real value is converting public evaluations into internal benchmarks: pick candidate models from the list, then run them on the same repositories, tests, and tool permissions. Record completion rate, failure recovery, human interventions, token consumption, and total cost to determine fit for your workflow.

The article includes a configuration snippet for developers using Claude-compatible agent tools who need to switch model services via Code80 endpoint, keeping model and service address separate from business code:

env:ANTHROPIC_AUTH_TOKEN = "<你的 API Key>" claude

Bilibili's AI Content Scale Changing Evaluation

Public data shows Bilibili's AI knowledge content consumption time grew 72% year-over-year, with over 190 million monthly users watching AI content. A complete content loop has formed around model releases, discussions, evaluations, usage, help-seeking, and co-creation.

This pushes model evaluation from labs into broader user scenarios. Models must now withstand repeated real-world trials and public comparisons across domains. For AI product teams, rising volumes of authentic feedback demand greater attention to version tracking, task reproducibility, and explanation of disputed results.

FAQ

How does AI Infinite Arena differ from traditional benchmarks?

Traditional benchmarks use fixed questions and scoring; AI Infinite Arena lets UP主 create tests from real workflows, emphasizing scenario experience and same-question comparison.

Does GPT-6 Astra topping the list mean it is comprehensively strongest?

No. First-place only means it performed best in the current 10 evaluations and their specific tasks; it cannot replace a team's own code, tool, and cost assessment.

How should development teams use such leaderboards?

Treat the leaderboard as a candidate screening entry point, then re-test with fixed repositories, fixed tests, and unified permissions, recording success rate, latency, tokens, human corrections, and failure recovery.

Why is real-workflow evaluation important?

Models ultimately enter codebases, documents, data, and business processes. Real tasks expose actual weaknesses in long-context handling, tool calling, collaboration, and acceptance stages.