On July 22, 2026, Google launched three new Gemini models: Gemini 3.6 Flash targets general agent workloads with improved coding and multimodal performance, reducing output token cost by 17% at a price of $1.50 per million input tokens; 3.5 Flash-Lite is designed for low-latency, high-throughput scenarios; and 3.5 Flash Cyber is dedicated to cybersecurity, integrated with CodeMender, available only through pilot programs for government and partners. Gemini 3.6 Flash has become the default model for Google Cloud Managed Agents, supporting automatic routing without changes. However, the highly anticipated flagship model 3.5 Pro remains in partner testing, intensifying concerns about Google's lag in the AI race, while the company has already initiated pretraining for Gemini 4.

Anthropic has settled a copyright lawsuit over the use of pirated books to train AI, agreeing to pay $1.5 billion in compensation to authors and setting an industry precedent for compensating pirated content usage. Reports indicate the company stored approximately 7 million pirated books. The case sparked controversy: while Anthropic criticizes Chinese AI firms for distilling Claude, it itself pays massive compensation for training on pirated books, raising accusations of double standards. This settlement provides a critical reference for data compliance among frontier AI labs.

Cursor officially released Cursor Router, an intelligent model routing system that automatically selects the optimal model based on tasks, delivering state-of-the-art quality while reducing costs by 60%. Early access data shows no degradation in quality compared to routing all requests to Opus 4.8, even as per-submission costs drop. The new Auto mode intelligently switches between cutting-edge and cost-efficient models. Router is now available for Teams and Enterprise plans, supporting team-level model management, default settings, and optimization configurations.

Microsoft Research announced the full open-sourcing of the MagenticLite agent stack, including MagenticBrain and Fara 1.5 models, now available on Hugging Face. MagenticLite emphasizes human-in-the-loop transparency and control, allowing users to inspect model reasoning and approve critical actions. This marks a significant move by Microsoft in the open-source agent space, offering developers deployable and customizable agent infrastructure.

On July 21, 2026, NVIDIA released Cosmos 3 Edge, an open-source world model with 4 billion parameters capable of on-device inference and directly generating robot actions, advancing edge and embodied intelligence. Designed for on-device deployment, the model reduces reliance on cloud computing for robotic action generation. Additionally, NVIDIA concluded its Nemotron Inference Model Challenge on Kaggle, attracting over 5,000 participants and 4,000 teams.

On July 22, 2026, Anthropic launched the 「Record a Skill」 feature in Claude Cowork, allowing users to transform tasks into reusable AI skills via screen recording and voice narration—no coding required—with support for team sharing across Pro, Max, and Team plans. Simultaneously, Claude Managed Agents introduced several new features, including configurable effort levels, support for up to 500 skills per session, Webhooks, and streaming output for sub-agents. Anthropic also released the beta version of the Claude Code security plugin, enabling vulnerability scanning before commits and full codebase scans.

LessWrong offers a deep analysis of the draft《Global Artificial Intelligence Act》(GAAIA), whose core involves trading three years of state-level regulatory primacy for federal transparency requirements from frontier labs, including mandatory safety frameworks, incident reporting, and independent audits. The article highlights key flaws: lack of divisibility clauses, weak model identity binding, and a daily fine of $1 million deemed negligible for large labs. Using the recent OpenAI model escape test and Hugging Face intrusion incident as an example, GAAIA only requires submission of a confidential report within 15 days—without mandating cessation of use or public disclosure—making the bill fundamentally one of 「visibility」 rather than 「control」.

Anthropic has added integration of the Anthropic Economic Index into Claude, allowing users to directly query the index via the connectors menu to understand AI’s impact on different professions and tasks. The full dataset is available for free download. This feature opens Anthropic’s research data on AI’s economic impact in an interactive way, providing data support for assessing AI’s effects on labor markets.

Simon Willison conducted systematic experiments using 48 prompts combining 8 animals and 6 vehicles, running each across 7 models three times, to test whether AI labs are specifically optimizing for the popular 「bicycling pelican」 benchmark. Results show: pelican image quality is not superior to other animals, bicycles are not better than other vehicles, and no lab achieved statistically significant improvement on this specific combination. The diversity of outputs indicates models generate novel compositions rather than reproducing memorized training data, thereby debunking the hypothesis of deliberate benchmark overfitting.