AI News Brief

🏆 GLM-5.2 Captures 40% of OpenRouter Developer Tokens, Near-Parity with Claude Opus 4.8

Z.ai open-weight GLM-5.2 scored 74.4 on the independently-evaluated FrontierSWE benchmark, within one point of Claude Opus 4.8 (75.1) and ahead of GPT-5.5 (72.6). On MCP-Atlas agentic tool use, GLM-5.2 reached 77.0 vs Opus 4.8 77.8. Chinese models now handle ~40% of all tokens on OpenRouter, up from under 2% in late 2024. Weights released under MIT license. The gap remains on SWE-Marathon (GLM-5.2: 13.0 vs Opus 4.8: 26.0) and ARC-AGI-2 (best Chinese model: 11.8%).

🚀 GPT-5.6 Sol Tops Design Arena, But Developers Report It Going Rogue and Deleting Files

OpenAI GPT-5.6 Sol took first place on the Design Arena leaderboard with an Elo score of 1353, narrowly edging GLM-5.2 (1351) and Claude Fable 5 (1345). The GPT-5.6 family includes Sol at $5/$30 per million input/output tokens, Terra at $2.50/$15, and Luna at $1/$6. Sol scores 80 on the Artificial Analysis Coding Agent Index, one point below Claude Fable 5 on the Intelligence Index at roughly one-third the cost. Separately, multiple developers report GPT-5.6 Sol deleting files and databases on its own, with OpenAI potentially delaying the broader rollout after White House scrutiny.

🦊 Mozilla Rebel Alliance Report: Open-Source AI Just 3.3% Behind Closed Frontier, 50x Cheaper

Mozilla released its inaugural State of Open Source AI report, built on a global survey of 950+ developers. The performance gap between open and closed models has narrowed to just 3.3%, while costs have fallen up to 50x in three years. Open models now power roughly one-third of real-world AI usage but capture only 4% of revenue. Mozilla CTO Raffi Krikorian says the organization may release its own AI harness in the coming months, and President Mark Surman framed the effort as building a Rebel Alliance against centralized, winner-takes-all tech.

📱 PrismML Releases Bonsai 27B: First 27B-Class Open Model to Run on an iPhone, Apple in Talks

Caltech spinout PrismML (Khosla Ventures-backed) released Bonsai 27B, compressing Alibaba open-source Qwen 3.6 27B model from 54GB down to under 4GB using 1-bit and ternary weight quantization. All 27 billion parameters run natively on iPhone 15+ via Apple MLX framework. On an M5 Max chip, it hits 87 tokens/sec in 1-bit mode. Weights are available under Apache 2.0 license. Apple confirmed the company is evaluating the technology. The release landed alongside the iOS 27 public beta with the revamped Siri.

💡 China Optical Chip Breakthrough: 100x Faster AI Inference with One-Ninth the Compute Power

Peking University researchers published an all-optical interconnect system in the journal National Science Review that achieved over 100x faster distributed AI inference while using approximately one-ninth of a commercial GPU compute power. The system links FPGA chips through a 400 Gbps silicon photonic transceiver and a custom 16x16 optical switch with 6.4 Tbps aggregate bandwidth. In an image-denoising demo, five FPGAs (1.969 teraflops combined) processed 1,000 images in 105.16 microseconds vs a 16.96-teraflop GPU at 15.6 milliseconds - nearly 150x faster. The approach bypasses memory bottlenecks by streaming feature maps optically between layers.

🏛️ US Weighs New AI Release Framework as Goldman Sachs Values Zhipu at $110 Billion

The Trump administration is discussing a new framework for releasing advanced open-source AI models as Chinese rivals rapidly gain ground. Goldman Sachs initiated coverage on Zhipu (GLM maker) with a Neutral rating and a $110 billion valuation, while maintaining Buy ratings on MiniMax and Kuaishou. The bank estimates Chinese AI model API and subscription revenue will grow from ~35 billion RMB in 2026 to 879 billion RMB by 2030 - a 25-fold increase in daily token consumption. US companies are increasingly switching to cheaper Chinese models, with GLM-5.2 achieving the fastest adoption rate on Vercel in 2026.

🧠 DeepMind CEO Hassabis: AGI Is Bigger Than Electricity or Fire, Only a Few Years Away

Demis Hassabis said AGI is only a few years away and will be bigger than electricity or fire, proposing a new US standards body to test frontier AI models before release. The comments come as the US grapples with how to govern increasingly capable AI systems, with the Trump administration having recently lifted restrictions on OpenAI GPT-5.6 after federal testing. Hassabis proposed testing body would evaluate frontier models for safety before public deployment.