AI News Brief

🚀 Z.ai Releases GLM-5.2 Open-Weight Model, Matching Claude Opus at 1/6 the Cost

Z.ai (formerly Zhipu AI) released GLM-5.2 under MIT license, scoring 62.1% on SWE-bench Pro and 74.4 on FrontierSWE - just behind Claude Opus 4.8's 75.1 but ahead of GPT-5.5's 72.6. Priced at $1.40/$4.40 per million tokens vs Opus's $5/$25 (roughly 82% cheaper on output), GLM-5.2 is now the strongest open-weights coding model available. Z.ai also launched ZCode, a free agentic IDE built around the model.

🧮 Mistral's Leanstral 1.5 Solves 587 of 672 Putnam Math Problems - Apache 2.0 Licensed

Mistral AI released Leanstral 1.5, an open-source code agent model for Lean 4 automated theorem proving and proof engineering. The model solved 587 of 672 Putnam math problems, with weights released under Apache 2.0 license and a free API endpoint available. The release marks a real shift for formal verification tooling.

🤖 Tencent Releases Hy3 Open-Source Model, Punching Above Its Weight Class

Tencent unveiled Hy3, the official release of its latest Hunyuan LLM, as a smaller open-source model designed to rival larger competitors in coding, reasoning, and agent tasks. Hy3 approaches flagship-model performance while keeping parameter count down - a notable entry in the open-weights space.

⚡ OpenAI Previews GPT-5.6: Sol Hits 88.8% on TerminalBench 2.1, Outscoring Claude Opus

OpenAI previewed GPT-5.6 in limited release with three new model variants; "Sol" scored 88.8% on TerminalBench 2.1 versus Claude Opus 4.8's 78.9%. "Ultra Mode" can spawn subagents autonomously, while the lower-cost "Terra" variant cuts inference pricing. METR flagged risk concerns, and the White House's June 2 executive order requests 30-day pre-release access to frontier models from trusted partners.

🔧 Nvidia's Kyber AI Rack System Delayed to 2028 on Manufacturing Snags

Nvidia's next-gen Kyber rack system for its 2027 Rubin Ultra chips has slipped to 2028 due to a hard-to-build circuit board, per SemiAnalysis. The delay breaks Nvidia's breakneck annual release cadence and gives rivals an opening as manufacturing limits collide with chip-design ambition.

💰 Samsung Posts 1,800% Profit Surge, Q2 Operating Profit Tops Nvidia's Record

Samsung Electronics forecast a record Q2 operating profit of 89.4 trillion won on AI memory chip demand, with profit surging 1,800% year-over-year. Samsung and SK Hynix also announced a massive southwestern South Korea chip fab buildout following a June 29 government directive to anchor global HBM, DRAM, and NAND capacity.

⚖️ Illinois Signs Nation's Strongest Frontier AI Model Law

Illinois Gov. J.B. Pritzker signed into law new transparency and safety reporting obligations for companies developing the most advanced AI models - the strongest state-level frontier AI law in the U.S. The move comes as the Trump administration's June 2 executive order establishes a voluntary federal pre-release framework with an August 1, 2026 deadline, and Beijing reportedly considers curbing overseas access to China's top AI models via a tiered export system.