Overview
- Mozilla’s second State of Open AI report published on Sept. 15 finds the capability gap between top Chinese open‑weight models and leading U.S. closed models has shrunk to about 4.4 months and measures task reliability by how long a model can complete work.
- Usage data from OpenRouter show Chinese models accounted for roughly 55% of token traffic by June, up from about 20% in September 2025, signaling rapid commercial adoption of open‑weight systems.
- Companies are shifting routine workloads to cheaper open models while keeping closed models for longer, specialized tasks because Mozilla’s METR metric shows open models reliably finish about seven hours of work versus roughly 12 hours for the best closed models.
- Investors and firms are responding: Bloomberg reports Manus is nearing a $500 million round that would lift its valuation to about $4 billion, and providers including DeepSeek, OpenAI and Google have cut prices to compete with lower‑cost open offerings.
- High‑profile warnings about recursive self‑improvement have intensified debate over slowing development, but an arXiv paper and several analysts say most current systems do not meet strict RSI criteria, leaving safety policy and oversight as the next battleground.