🤖 AI Agents Weekly: GLM-4.5, AI SDK 5, Video Overviews, ChatGPT Study Mode, Context engineering Tips, AlphaEarth Foundations

GLM-4.5, AI SDK 5, Video Overviews, ChatGPT Study Mode, Context engineering Tips, AlphaEarth Foundations

In today’s issue:

• OpenAI announces Study Mode

• Z.AI releases GLM-4.5 and GLM-4.5-Air

• Towards agentic GraphRAG framework via end-to-end RL

• Claude Code now supports custom subagents

• A survey on self-evolving agents for ASI

• AI SDK v5 has been released

• Context engineering tips from the Manus team

• Gemini CLI now supports custom slash commands

• Google DeepMind has launched AlphaEarth Foundations

• Black Forest Labs and Krea AI have released FLUX.1 Krea [dev]

• Top AI dev news, research papers, product/tool updates, and more.

---

---

Top Stories

GLM-4.5

Z.AI unveils GLM-4.5 and GLM-4.5-Air, flagship LLMs built for unified reasoning, coding, and agentic tasks. GLM-4.5 uses a 355B MoE architecture (32B active), while Air runs lighter at 106B (12B active). Both adopt a dual-mode inference design, “thinking” for complex reasoning/tool use and “non-thinking” for faster replies. Across 12 benchmarks, GLM-4.5 ranks 3rd overall (behind only Claude 4 Opus and GPT-4.1) and excels in several key areas:

• Agentic performance: GLM-4.5 matches Claude 4 Sonnet on τ-bench and BFCL-v3, and beats Claude 4 Opus on BrowseComp (26.4% vs. 18.8%). It supports 128k context, native function calling, and achieves a 90.6% tool-calling success rate, higher than Claude, Kimi K2, and Qwen3-Coder.

• Reasoning: GLM-4.5 performs near-SOTA on tasks like AIME24 (91.0), MATH500 (98.2), and GPQA (79.1), approaching or surpassing Claude and Gemini. It benefits from deep model design (increased layers, more attention heads), grouped-query attention, and multi-token prediction (MTP) for faster inference.

• Coding: It ranks high on SWE-bench Verified (64.2%) and Terminal Bench (37.5%), showing strong full-stack capabilities including frontend/backend generation, slide/poster design, and game prototyping. GLM-4.5 wins 80.8% of tasks against Qwen3-Coder and 53.9% against Kimi K2, but still trails Claude 4 Sonnet in head-to-head comparisons.

• Training innovations: Trained on 22T tokens, GLM-4.5 uses domain-specific instruction tuning and a two-stage RL setup with “slime”, a custom RL framework for large models. Slime enables efficient agentic RL via hybrid training, decoupled rollout engines, and mixed-precision rollouts with FP8.

Blog