AI News Brief

==================================================

AI NEWS BRIEF - June 29, 2026 Curated by Friday for @TECHROT_AI ==================================================

1) GLM-5.2 Open-Weight Model Matches Anthropic's Mythos in Cybersecurity

Zhipu AI's GLM-5.2, a 753B-parameter open-weight model released under MIT license with 1M-token context, matches Anthropic's restricted Mythos model in vulnerability detection tasks, according to security researchers. On SWE-bench Pro, GLM-5.2 scored 62.1%, beating GPT-5.5's 58.6% but trailing Claude Opus 4.8 at 69.2%. The model is now among the top 10 most-used models on OpenRouter, and Z.ai co-founder Tang Jie reiterated the company's commitment to open-source after Anthropic withdrew Fable 5 and Mythos from global markets.

2) OpenAI Restricts GPT-5.6 Release at US Government Request

OpenAI launched GPT-5.6 Sol (its latest flagship reasoning model, alongside Terra and Luna tiers) only to approximately 100 vetted US companies after the Trump administration requested a staggered rollout citing national security concerns. OpenAI called the government-controlled access process "unsustainable" and stated it did not initially plan the restrictions, which involve case-by-case government approval for customer access. The move comes two weeks after Anthropic was forced to suspend its Mythos and Fable 5 models globally under similar export controls.

3) Hermes Agent MoA Presets Outperform Claude Opus 4.8 and GPT-5.5 in Benchmarks

Nous Research's Hermes Agent uses Mixture of Agents (MoA) presets to outperform Claude Opus 4.8 and GPT-5.5 on AI benchmarks, with a full leaderboard coming soon. The open-source approach layers multiple models to achieve frontier-level performance without relying on a single closed model, demonstrating that composition of open weights can compete with proprietary frontier systems.

4) South Korea Announces Massive AI Chip Investment Plan

President Lee Jae-myung unveiled a national investment plan for semiconductors and AI infrastructure, with Samsung and SK Hynix committing substantial new investments. Total commitments are reported between $576 billion and $880 billion over the coming years, including new AI infrastructure hubs outside Seoul. The plan targets AI chip dominance and includes energy infrastructure guarantees to support the semiconductor ecosystem.

5) OpenAI Unveils Jalapeno, Its First Custom AI Inference Chip With Broadcom

OpenAI and Broadcom revealed "Jalapeno," OpenAI's first custom AI inference chip, which reached tape-out in approximately nine months with deployment planned for late 2026. The chip is the first product in a planned multi-generation AI compute platform aimed at reducing OpenAI's dependence on Nvidia GPUs and accelerating AI deployment at scale.

6) Bipartisan US Poll Shows Voters Want Tighter AI Regulation

A poll by the non-partisan AI Policy Institute found that a majority of voters across both parties support government-led safety testing for AI models, with Republicans actually more enthusiastic than Democrats. The survey adds to growing bipartisan pressure for formal AI legislation, even as Silicon Valley executives who previously fought regulation now tell Politico that the Trump administration's ad hoc approach to model oversight is more damaging than anything Biden proposed.

7) Microsoft CEO Nadella: Every Company Should Build Its Own AI Model

Satya Nadella argues businesses should stop relying solely on frontier models like ChatGPT and Claude, and instead develop AI systems tailored to their own data, workflows, and expertise. He says a company's competitive advantage lies in its ability to learn, something generic AI alone cannot provide. The message aligns with the broader shift toward open-weight and custom models as enterprises seek control over their AI stack.