June 28, 2026

AI News Brief

ROT DECAY REBUILD
◀ All Briefs
// STORY_01

1. GLM-5.2: Zhipu AI's open-weight model triggers Five Eyes security alarm

Zhipu AI (Z.ai) released GLM-5.2 on June 13, a 753B-parameter open-weight model under MIT license that beats GPT-5.5 on coding benchmarks and rivals Claude Opus 4.8 at roughly 25% of the cost ($1.40/M input, $4.40/M output). The Five Eyes intelligence alliance issued a warning calling the model a "severe, immediate cyber threat" as it is already being exploited by hostile actors.

// STORY_02

2. OpenAI launches GPT-5.6 family (Sol, Terra, Luna) under government restrictions

OpenAI unveiled three new frontier models - GPT-5.6 Sol, Terra, and Luna - but the Trump administration blocked wide release, limiting initial deployment to government-approved partners. GPT-5.6 Sol reportedly beats Anthropic's Claude Mythos 5 in coding benchmarks, though OpenAI called the restricted rollout "unsustainable."

// STORY_03

3. US government allows limited release of Anthropic's Mythos 5 to 100+ organizations

After a two-week ban, the US Commerce Department approved Anthropic to restore access to its Mythos 5 cybersecurity AI model to over 100 "trusted" US organizations. The closely held Fable 5 model remains offline as negotiations with the Trump administration continue.

// STORY_04

4. JD.com open-sources JoyAI-VL-Interaction: first proactive vision-language AI

JD.com released JoyAI-VL-Interaction under Apache 2.0, an open-source vision-language model that watches live video streams and speaks without being prompted - the first proactive AI of its kind shipped with full weights, training data, and a deployable system. It was evaluated across 58 event-driven scenarios.

// STORY_05

5. MirrorCode benchmark: Claude Opus 4.7 solves 56% of weeks-long coding projects

The new MirrorCode autonomous coding benchmark shows Claude Opus 4.7 completing 56% of software projects estimated to take human engineers weeks - including rebuilding a 60,000-line interpreter in under 14 hours. Epoch AI and METR open-sourced the scaffold, paper, and 22 tasks alongside the results.

// STORY_06

6. OpenAI unveils "Jalapeno" - its first custom AI inference chip, built with Broadcom

OpenAI and Broadcom unveiled Jalapeno, OpenAI's first custom AI inference chip, reaching tape-out in roughly nine months. It is the first product in a planned multi-generation AI compute platform aimed at reducing inference costs and dependence on Nvidia, with deployment set for late 2026.

// STORY_07

7. AI startup Lindy ditches Claude entirely for DeepSeek, saving millions

AI startup Lindy replaced Claude entirely with DeepSeek across its stack after AI costs exceeded personnel costs, with CEO Flo Crivello calling the switch "a matter of survival for the business." The move signals growing cost pressure on closed-model providers as open-weight alternatives mature.