🚀 OpenAI GPT-5.6 Sol: Native 4-Agent Orchestration
OpenAI officially launched the GPT-5.6 series on July 9, introducing three models: the flagship GPT-5.6 Sol, the balanced Terra, and the cost-efficient Luna. Sol's standout feature is its Ultra mode, which can natively coordinate four AI agents in parallel to tackle complex tasks in code development, scientific research, cybersecurity, and knowledge work. OpenAI claims the series achieves industry-leading benchmarks while significantly reducing inference costs and latency compared to prior generations. The models also ship with what OpenAI describes as its most comprehensive safety framework to date, alongside native programmatic tool-calling support for enhanced automation. — OpenAI
💼 ChatGPT Work: Long-Running Office Agents & Atlas Sunset
On July 10, OpenAI unveiled ChatGPT Work, an office-grade intelligent agent designed to persist on projects "for hours if needed" and convert goals into completed work. Unlike earlier Agent Mode implementations that timed out after minutes, ChatGPT Work can sustain long-running automation — from client research to campaign briefs to localised marketing assets — while waiting for human approval at critical junctures. It integrates Scheduled Tasks (cron-like recurring jobs), plugins for Slack, Teams, Google Drive, and SharePoint, and desktop-level file access. Meanwhile, OpenAI confirmed the Atlas AI browser will be sunset on August 9, after just eight months as a macOS-only experiment. Browsing capabilities will be folded into ChatGPT Work instead. — OpenAI
🧠 Meta Muse Spark 1.1: Zuckerberg's First X Post in 3 Years
Meta released Muse Spark 1.1 on July 9, its most powerful multimodal reasoning model yet, purpose-built for agentic coding tasks. The model excels at tool calling, computer operation, and long-context workflow execution across external applications. CEO Mark Zuckerberg posted on X for the first time in three years to promote it, calling Spark "a powerful but incredibly cheap agentic and coding model." Pricing is set at $1.25 per million input tokens and $4.25 per million output tokens — well below most closed-source competitors. Muse Spark 1.1 supports a 1M-token context window and is available via API public preview, marking Meta's first paid API offering. — Meta AI
🔬 Anthropic Reveals Claude's Internal "J-Space"
Anthropic published a landmark interpretability paper on July 7 demonstrating that Claude's neural network has spontaneously formed an internal "workspace" — dubbed J-space — where the model stores and processes "verbalizable" representations that closely parallel human conscious thought. Researchers identified that only a small fraction of internal computations are accessible for reasoning and reporting, while the vast majority runs as automatic background processing. In a striking validation, Google DeepMind's interpretability team lead Neel Nanda confirmed they reproduced the paper's core findings on Qwen3.6-27B. Anthropic has open-sourced the code and launched an interactive demo with Neuronpedia. — Anthropic
📊 ByteDance EdgeBench: Agent Learning Doubles Every 3 Months
ByteDance Seed published EdgeBench on arXiv (July 6), a benchmark suite of 134 real-world tasks spanning scientific discovery, software engineering, combinatorial optimization, and interactive games — each designed to sustain at least 12 hours of continuous agent operation. Analyzing ~38,000 hours of agent interaction, the paper reports two headline findings: (1) agent performance follows a log-sigmoid scaling law with R² = 0.998, meaning learning from environments is strikingly predictable; and (2) agent learning speed roughly doubles every three months across model generations. Claude Opus 4.8 leads the leaderboard at 51.3, ahead of GPT-5.5 (48.4) and GPT-5.4 (39.3). — arXiv
⚡ Hugging Face Overhauls Kernels: 40% Inference Cost Cut
Hugging Face released a major overhaul of its kernel library, promising up to 40% reduction in LLM inference costs. The update introduces fused attention mechanisms (cutting memory usage by half for long contexts), adaptive precision switching between FP16/BF16/INT8, and kernel auto-tuning that dynamically selects the best implementation per hardware configuration. Early benchmarks show 2x speedup on long-context generation and 30% less peak memory during training. A typical deployment serving 1M requests/day could save ~$2,500/month in GPU costs. The improvements are available via Transformers v4.45.0. — Hugging Face
🇨🇳 Z.ai ZCode: Free Coding Agent Beats GPT-5.5
Z.ai (formerly Zhipu AI, listed in Hong Kong) launched ZCode on July 2 — a free agentic development environment built on GLM-5.2, a 744B-parameter MoE model. On SWE-bench Pro, GLM-5.2 scored 62.1, ahead of GPT-5.5's 58.6 and behind only Claude Opus 4.8's 66.0. ZCode's base tier is free, with Pro at $64.80/month undercutting Cursor Ultra's $200. The API is priced at $1.40/M input tokens — a fraction of Claude Opus 4.8's $5. Z.ai's on-premises deployment revenue reached RMB 534M in FY2025, and its market cap crossed HK$1 trillion after GLM-5.2's release. — Z.ai
🔗 Z.ai Official · Hugging Face GLM-5.2
KD Agentic · AI Daily Digest — July 13, 2026
Top comments (0)