AI News Brief

GPT-5.6 Sol Lands, Matches Claude Fable 5 at 1/3 the Cost

OpenAI released GPT-5.6 Sol, the flagship of a new 3-tier family (Sol, Terra, Luna). On Artificial Analysis Intelligence Index, Sol places second behind Claude Fable 5 while costing roughly one-third the price. Meanwhile, GLM-5.2 undercuts both at approximately one-tenth the cost, compressing the frontier to a one-point spread across vastly different price points. Sol uses about 15,000 tokens per task vs GPT-5.5 16,000, delivering more intelligence per token.

GLM-5.2 Tested: China Free Open-Weight Model in the Wild

Business Insider ran GLM-5.2 through real-world tasks including email writing, shopping assistance, trip planning, and poster design. The free Zhipu model does not yet match the polish of paid frontier models, with frustrating capacity limits and slow responses, but shows strong fundamentals. Separately, a Databricks benchmark on its multi-million-line enterprise codebase revealed open-source models like GLM 5.2 can match expensive frontier models on everyday coding tasks, reinforcing the open-source value proposition.

Zhipu Founder Defends Open-Source AI Amid Global Security Debate

Tang Jie, founder of Chinese AI lab Zhipu (maker of GLM-5.2), argued in an internal staff memo that genuine AI security stems from broad participation, sharing and oversight rather than restricted access. The statement comes as global policymakers intensify debates over AI model security and the risks of open-weight releases, positioning Zhipu as a vocal open-source advocate on the world stage.

Beijing Eyes Curbs on Foreign Access to China Top AI Models

Chinese authorities held meetings with top tech firms over the past month about potentially restricting overseas access to China most advanced AI models, including unreleased ones. The move could limit global use of leading open-weight models like GLM-5.2 and DeepSeek, tightening the technology corridor between China and the West at a critical moment in the AI race.

Meta Iris In-House AI Chip Enters Production in September

Meta announced its custom AI chip, code-named Iris (part of the MTIA family), will enter production in September. Designed with Broadcom and built by TSMC, Iris is part of a major infrastructure push: 7 gigawatts of computing capacity coming online in 2026, with plans to continue scaling through 2027. Meta is accelerating its chip release cadence to every six months, far faster than the typical annual cycle.

OpenAI Finds 30% of SWE-Bench Pro Tasks Broken, Pulls Endorsement

OpenAI reviewed SWE-Bench Pro, one of the most widely used benchmarks for AI coding ability, and found roughly 30% of its tasks are broken. The company is pulling its earlier endorsement of the benchmark, raising serious questions about the reliability of AI coding evaluations. The finding adds to growing scrutiny of benchmark contamination and reward hacking after Opus 4.8 performance fell 14% when internet access was removed.

China Optical Chip Breakthrough Boosts AI Speed 100x

Researchers at Peking University developed an all-optical interconnect system that links standard electronic chips with specific algorithms, boosting AI processing speed by 100-fold while using a fraction of the compute power. The breakthrough could reshape the AI hardware landscape by dramatically improving interconnect efficiency without requiring entirely new chip architectures.