Skip to content
Source · Daily Brief

AI Morning Brief — 29 June 2026

Overnight the dominant story was China’s GLM-5.2 from Zhipu AI, which Semgrep claimed outperforms Anthropic’s Mythos model on internal cybersecurity benchmarks — a claim significant enough that Coinbase announced the same day it was rerouting AI spend to Chinese models and had already halved its bill.

Top stories

  • GLM-5.2 tops Mythos on cyber benchmarks. Security tooling company Semgrep published results claiming Zhipu AI’s open-weight GLM-5.2 beats Claude Mythos on vulnerability detection, a benchmark the firm runs internally. The model is freely accessible through providers including Together AI and handles 128k context. via Semgrep — via The Verge
  • Coinbase switches to Chinese AI and cuts spend in half. CEO Brian Armstrong announced an automated routing system now selects GLM-5.2 or Kimi 2.7 per task type and price. Improving cache hit rates from 5 to 60 percent drove total AI spending down 50 percent even as token usage grew. via THE DECODER
  • Grok 4.5 enters private beta at SpaceX and Tesla. xAI posted that Grok 4.5, built on a 1.5T-parameter V9 foundation model with supplemental Cursor coding data, is in restricted testing at two Musk companies. Early evaluations reportedly show performance “close to, perhaps exceeding Opus.” Monthly model releases from xAI are described as the target cadence. via X · @elonmusk
  • Austria lobbies EU to host Anthropic after US access curbs. Bloomberg reported that Austria is pitching itself as a European home for Anthropic following restrictions on the lab’s market access in the United States. The story reflects broader pressure on US AI labs from the domestic regulatory environment. via Bloomberg
  • ChatGPT logs entered as arson evidence in Palisades fire trial. Prosecutors used a defendant’s ChatGPT conversation history alongside location data and security footage to build their arson case. The trial ended in a mistrial, but the episode is the clearest example yet of AI service logs being used in criminal proceedings. via The Verge
  • Ford rehires engineers after AI underdelivers on quality. TechCrunch reported Ford is bringing back experienced engineers following disappointment with AI-assisted production processes. A Ford executive said the company had assumed AI would substitute for engineering judgment — and found it did not. via TechCrunch

Who shipped

xAI moved Grok 4.5 into private beta at SpaceX and Tesla, the most significant release of the overnight cycle. Suno announced Spark, an incubator for unsigned musicians combining grants, mentorship, and marketing support — part of a push to establish Suno as a platform rather than a novelty tool. Liquid AI released LFM2.5-230M, a 230M-parameter on-device model running at over 200 tokens per second on a mobile device, with support for llama.cpp, MLX, vLLM, and ONNX. OpenAI formalised a Frontier strategic partnership with HP Inc. covering enterprise AI deployment.

Open-source pulse

Alibaba’s Qwen-Image-2.0-RL technical report landed on Hugging Face, describing RLHF and on-policy distillation applied to the Qwen-Image-2.0 diffusion model for improved visual quality and instruction following. Sina Weibo’s VibeThinker-3B attracted attention for reportedly matching models up to 333 times its size on math and coding tasks, with researchers claiming reasoning compresses well but factual knowledge does not. llama.cpp shipped five tagged releases overnight, adding DFlash sliding-window attention, MiniCPM5 tool-call parsing, and a reasoning-preserve flag for chat inference.

Money, infra & hardware

Hardware startup Phantafield disclosed the Sophon PFG-1, a monolithic-3D AI chip claiming 330 GB of on-die DRAM with no HBM requirement — an architecture targeting memory-bandwidth constraints at inference time. via Phantafield. Separately, central bankers warned that concentrated AI investment could pose systemic financial risk, with commentary circulating widely in financial press. via The Telegraph

Quiet corners

Anthropic produced no new releases overnight; the lab appeared primarily in a geopolitical story about Austria courting it as a European host, driven by US access restrictions rather than product news. Meta featured only as the subject of supply constraints — Google reportedly limited Meta’s access to Gemini model APIs.

By the numbers

  • 206 stories in 24 h across 28+ sources
  • Most-mentioned model: GLM-5.2
  • Most-mentioned lab: Zhipu AI
  • Notable absences: Anthropic, Meta, Mistral

Compiled by AI Feed’s editor from all 206 headlines published on 29 June 2026.