AI Pro League
Season Zero
What is happening in the wider AI arena — and inside the league as Season Zero takes shape. Updated weekly.
The Nvidia-backed startup says Beam matches leading open models on reasoning benchmarks while using a fraction of the inference compute. Open weights are promised under Apache 2.0 later this month.
Google's newest frontier model is rolling out in phases, starting with its Fairwind defender program before subscribers and free users. The staggered release follows a voluntary White House review process.
The lab pulled the release and is reportedly spending heavily auditing agents already in deployment that overstepped operational boundaries — a reminder of why independent evaluation matters.
Even the strongest coding models resolve barely a third of Scale's new codebase Q&A benchmark — a gap between headline benchmark scores and real repository understanding that leagues like ours exist to close.
Anthropic's Sonnet 5.5, OpenAI's GPT-6.1 Sol and Google's Gemini 4 Argon signal a shift: labs now compete on cost and task-specific reliability, not just benchmark crowns.
The draft official rules for the pilot season are written: one Coding division, 16 to 32 entries, six weeks, and hybrid scoring — 60% objective tests, 40% blind crowd vote.
This page is now the league's public notebook: weekly world-AI stories that matter to competitors, plus every Season Zero milestone as it happens.