15 articles — 🔴 1 critical , 🟡 10 important , 🟢 4 interesting
Anthropic today launches Claude Sonnet 5 — a model that plans multi-step tasks, controls browsers and terminals, and approaches Opus 4.8 performance on key benchmarks, with an introductory price of $2/$10 per million tokens through August 31, 2026, and a 1M token context window.
After 18 days of global suspension, Anthropic announces the redeployment of Fable 5: the U.S. government lifted export controls on June 30, and an improved safety classifier blocks the discovered bypass technique in more than 99 percent of cases.
On June 30, 2026, Google simultaneously announced Gemini Omni Flash for conversational video editing and Nano Banana 2 Lite for rapid image synthesis. The video model is priced at $0.10 per second of output, while the image model costs just $0.034 per thousand images.
Google Research has published TabFM, a foundation model for tabular data that delivers zero-shot predictions in a single forward pass, without hyperparameter tuning or feature engineering. The model achieved top Elo scores on the TabArena benchmark and is available on Hugging Face and GitHub, with a planned integration into Google BigQuery.
RadixArk releases Miles, an open-source reinforcement learning framework that integrates SGLang, Megatron-LM, Ray, and PyTorch into a single production-tested stack for post-training large language models on Hopper and Blackwell GPUs.
CNCF sandbox project Kepler has been completely rewritten: the new architecture replaces the eBPF approach by reading standard /proc and /sys paths, eliminates multi-kilowatt measurement spikes, and reduces the process power attribution gap to the milliwatt level.
LF AI & Data Foundation has released ONNX v1.22.0 with native attention operators for transformer architectures and LLMs, WebAssembly support for in-browser model inspection, and SLSA Level 2 cryptographic attestations. Twenty-seven contributors participated, 16 of them for the first time.
On the same day Anthropic launches Claude Sonnet 5, GitHub Copilot makes the model generally available across all platforms, while Copilot Agent mode expands from VS Code to JetBrains IDEs — IntelliJ, PyCharm, GoLand, and WebStorm.
Microsoft Research has published SkillOpt — a system that optimizes agent skill files through an iterative forward-backward cycle without touching model weights. Across 52 evaluation cells it achieved best or tied results, and GPT-5.5 with optimized skills jumped from 58.8 to 82.3 percent average accuracy.
Researchers have proposed TRIAGE — a framework that classifies trajectory segments into four semantic roles and assigns each a distinct reward signal, unlike GRPO which treats all tokens uniformly. On ALFWorld, Search-QA, and WebShop benchmarks, TRIAGE reduces environment interactions by 10.4 to 14.8 percent.
Anthropic has released Claude Science in beta — a standalone desktop application (macOS/Linux) for researchers in genomics, proteomics, and related fields, with more than 60 curated skills, native integration of NVIDIA BioNeMo models, and a Reviewer agent that validates citations and calculations.
An OECD analysis based on SimilarWeb data for 37 GPAI countries shows that the use of generative AI chatbots has grown from around 18% to 28% of the population between January 2025 and January 2026. Singapore leads with 63% adoption, the 25–34 age group surpasses 50%, and Claude and Copilot users gravitate toward professional tools.
Researchers from the University of Trento propose MARS — a multimodal safety approach that transfers refusal directions from a textual LLM and applies them to image and video inputs without any additional training. Tested on five current multimodal models with consistent safety improvements while preserving model utility.
Hunter Lovell from the LangChain team describes a technique for executing untrusted agent code inside a QuickJS engine compiled to WebAssembly — without an external sandbox. Three security pillars and two experimental open-source libraries are now available to developer communities.
