⚙️ API & Deprecations
OpenAI is sunsetting its AI-powered browser, Atlas, after less than a year. The company is moving some agentic browsing features to its desktop app and a Chrome extension. (https://techcrunch.com/2026/07/09/openai-is-shutting-down-atlas-but-its-ai-browser-ambitions-are-still-growing/)
Anthropic has added the ability to set an expiration date when creating new API or Admin API keys. Existing keys are unaffected, and creators are emailed before keys with a lifetime of at least 7 days expire. (https://docs.anthropic.com/en/release-notes/api#july-8-2026)
Anthropic introduced a beta header (agent-memory-2026-07-22) that changes the behavior of listing memories. It enforces a stable, server-defined order and restricts the depth parameter to 0, 1, or omitted. (https://docs.anthropic.com/en/release-notes/api#july-2-2026)
🚀 Models & Releases
OpenAI has released several new models in its GPT-5.6 series on OpenRouter. These include the fast, cost-efficient GPT-5.6 Luna, the reasoning-enhanced GPT-5.6 Luna Pro, and the reasoning-enhanced GPT-5.6 Terra Pro. (https://openrouter.ai/models/openai/gpt-5.6-luna, https://openrouter.ai/models/openai/gpt-5.6-luna-pro, https://openrouter.ai/models/openai/gpt-5.6-terra-pro)
xAI has released Grok 4.5, described as its smartest model with frontier performance on coding, knowledge work, and STEM. A grok-latest model alias has also been added. (https://openrouter.ai/models/x-ai/grok-4.5, https://openrouter.ai/models/~x-ai/grok-latest)
AionLabs released Aion-3.0-Mini, a multi-model roleplaying and storytelling system built on the DeepSeek family. (https://openrouter.ai/models/aion-labs/aion-3.0-mini)
NVIDIA published several new models on Hugging Face: Kimi-K2.7-Code-DFlash (a Model Optimizer for Moonshot's Kimi model), nvDock, and Nemotron-Labs-Audex-2B, a 2B parameter audio-language model. (https://huggingface.co/nvidia/Kimi-K2.7-Code-DFlash, https://huggingface.co/nvidia/nvDock, https://huggingface.co/nvidia/Nemotron-Labs-Audex-2B)
🏢 Industry & Business
OpenAI's second-in-command, Fidji Simo, is stepping down from her full-time role after an extended medical leave. This creates a leadership vacuum as the company eyes a possible IPO and races to catch Anthropic in the enterprise market. (https://techcrunch.com/2026/07/09/fidji-simo-steps-down-from-openais-no-2-role/)
OpenAI stated that its new GPT-5.6 family of models will continue to be the "preferred model" for Microsoft Copilot 365, amid ongoing chatter about a potential breakup between the two companies. (https://techcrunch.com/2026/07/09/openai-says-gpt-5-6-is-the-preferred-model-for-microsoft-copilot-amid-breakup-chatter/)
🛠️ Tools & Products
Weaviate released maintenance versions v1.37.12 and v1.36.21. Both contain fixes for issues like HNSW maintenance flag leaks, BM25 correctness, queue resilience, and RBAC gRPC service mismatches. Neither release contains breaking changes. (https://github.com/weaviate/weaviate/releases/tag/v1.37.12, https://github.com/weaviate/weaviate/releases/tag/v1.36.21)
Ollama v0.31.2 was released, enabling flash attention on older NVIDIA GPUs (compute capability 6.x), improving iGPU vision model offloading, and fixing structured output for thinking models. (https://github.com/ollama/ollama/releases/tag/v0.31.2)
OpenHands released version 1.11.0, adding features like passing repository metadata to observability traces and Agent Profiles for its cloud/SaaS backend. (https://github.com/OpenHands/OpenHands/releases/tag/1.11.0)
Pydantic-AI released v2.7.0, adding support for Azure response shorthands and the xAI grok-4.5 model. (https://github.com/pydantic/pydantic-ai/releases/tag/v2.7.0)
OpenAI's Codex (rust) released v0.144.0, adding features like usage-limit reset credit management, a writes app-approval mode, and interactive authentication for MCP tools. (https://github.com/openai/codex/releases/tag/rust-v0.144.0)
Llama.cpp commit b9946 introduced Hexagon optimizations, including tiling and tracing for unary ops to avoid VTCM overflow. (https://github.com/ggml-org/llama.cpp/releases/tag/b9946)
🔬 Research
Anthropic researchers developed a technique providing a clearer glimpse inside LLMs as they reason, revealing processes ranging from mundane to unnerving. (https://www.technologyreview.com/2026/07/09/1140293/anthropic-found-a-hidden-space-where-claude-puzzles-over-concepts/)
A new paper proposes "Agentic Regeneration" to bypass database driver bottlenecks (like JDBC/ODBC) for analytical workloads, enabling direct, high-performance columnar reads. (https://arxiv.org/abs/2607.07696v1)
UniClawBench, a new universal benchmark for evaluating proactive agents on real-world tasks, was introduced. (https://arxiv.org/abs/2607.08768v1)
OpenCoF, a novel method for learning to reason through video generation, was proposed. It uses temporally connected frames as a distinct reasoning path from traditional Chain-of-Thought. (https://arxiv.org/abs/2607.08763v1)
⚖️ Regulation & Society
Hawaii students will help draft national AI policy at a fellowship in Boston. (https://news.google.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?oc=5)
An AI Policy Framework event was held with Senator McCormick. (https://news.google.com/rss/articles/CBMiigFBVV95cUxQTkZOeHVjU0k3eUx4WDVxV25zVWRsS3poZUE3dFM5U2FLZkVuUG5MWmhGUmVOemNsSF9udjY5WEhYNWtEQlhEbDNFR01HMkNSVTh2dlZKbXFUT1NFSEprdDNUTWNsZFlWWlRraGNCdE1mN3lsa2I3MTczOU9wbHBNREFuTUVNV2ZKLUE?oc=5)
India is signaling a shift in its approach to AI regulation, according to notes from the Asia-Pacific region. (https://news.google.com/rss/articles/CBMimwFBVV95cUxPMUZPS0VBUVFWaVVwTG41V1pPWFg1WjZvd1NNZnRfWnBTMjFWTEZjSnVfUFdvU1Y0VGNMS3R6MW1wQlBsM1dkRURneElEbldZdEo5NER0b3Q2UmVMQVBEeEk4T056aGZwLW9HcjI0VmJmOS1xNEtWTXU3OWJTV1pkMU9WOUthdVktd2FQS3pJdzh1WVJ4VjFGaFpsaw?oc=5)
💡 Key Takeaway
OpenAI is consolidating its agentic browsing strategy by shutting down the standalone Atlas browser, while Anthropic is tightening API security with expiring keys. The GPT-5.6 series expansion offers more granular model choices for cost vs. reasoning quality trade-offs.