AI News Brief

1. GLM-5.2 Tops Open-Weight AI Rankings, Powered by Huawei Silicon

Z.ai (formerly Zhipu AI) released GLM-5.2 under the MIT open-source license on June 17, and it now leads all open-weight models on SWE-bench Pro at 62.1%, ahead of GPT-5.5 (52.2 on Humanity's Last Exam) and closing the gap with Claude Opus 4.8 (69.2%). The model was trained entirely on Chinese-made Huawei chips, making it the first frontier-class open model built without Nvidia, AMD, or Western silicon.

Sources:

2. Meituan Open-Sources LongCat-2.0: 1.6T Parameter Model Trained on Chinese Chips

Chinese food-delivery giant Meituan released LongCat-2.0, a 1.6 trillion-parameter Mixture-of-Experts model with ~48B parameters activated per token, trained on 50,000 domestic Chinese chips (no Nvidia, no AMD). The model has a 1-million-token context window, was trained on over 30 trillion tokens, and is released under the MIT open-source license. It has been leading OpenRouter usage charts.

Sources:

3. US Lifts Export Controls on Anthropic's Fable 5 and Mythos 5

The US Commerce Department lifted restrictions on Anthropic's most powerful AI models, Claude Fable 5 and Claude Mythos 5, less than three weeks after ordering their suspension. Commerce Secretary Howard Lutnick sent a letter confirming Anthropic agreed to proactively detect security risks, work with the government on release protocols, and report malicious activity. Meanwhile, the White House separately asked OpenAI to delay public rollout of GPT-5.6, sparking confusion over US AI policy.

Sources:

4. Claude Sonnet 5 Launches as Anthropic's New Default Model

Anthropic released Claude Sonnet 5 on June 30 as the new default model for Free and Pro plans, replacing Sonnet 4.6. It beats Sonnet 4.6 across all benchmarks and edges past the larger Opus 4.8 on the GDPval-AA v2 knowledge work test (score: 1,618). Introductory API pricing is $2 per million input tokens through August 31, rising to $3 after. Available via Claude Code, API (claude-sonnet-5), and all subscription tiers.

Sources:

5. Portugal Open-Sources National Amalia AI Model

Portugal launched Amalia, its first national large language AI model, as a fully open-source platform. The model, training datasets, and source code are all released under an open-source license, enabling governments, researchers, and businesses to build sovereign European Portuguese AI applications without dependence on foreign proprietary models.

Sources:

6. South Korea Unveils $576B AI Chip Investment With Samsung and SK Hynix

South Korea announced a $576 billion AI and semiconductor investment plan, with Samsung Electronics and SK Hynix committing $517.87 billion to build two new chip fabrication plants. President Lee Jae Myung framed the initiative as a "national survival strategy" centered on semiconductors, physical AI, and data centers. The plan accelerates fab construction in the Yongin cluster by a decade to meet surging HBM demand.

Sources:

7. OpenAI Previews GPT-5.6 (Sol, Terra, Luna) But Delays Public Release

OpenAI previewed GPT-5.6 in three variants on June 26: Sol (top cybersecurity performer), Terra, and Luna. Public release has been delayed to a "small group of trusted partners" following a White House request to stagger deployment. The models introduce Ultra mode, enabling multiple specialized sub-agents to work in parallel rather than sequential reasoning. OpenAI is working with the government on a formal review process for model releases.

Sources: