Qwen 3.8 Max (2.4T) and 27B Open-Weight Models Released for Coding and CoworkUnpacking ChatGPT Work: How Memory, Proactivity, and Scheduling FunctionLLM 0.32 Released with Reasoning Traces, OpenAI Responses, and Server-Side ToolsMiniMax-H3 Now Available on MLX: Generates 15-Second Video with AudioFirst Production-Scale Study Reveals Copilot Agent Coding TracesUncertainty-Aware Inference Makes LLMs More Reliable for Operations ResearchClaude Code v2.1.222 Fixes Security Holes in Worktree Isolation and PreToolUse HooksGitHub Shows How to Turn One Giant AI PR into a Reviewable StackOn-Policy RL's Hidden Flaw: Verifier Rewards Can Reshape Support, Making Later Goals UnlearnableDLLM-TTS: Block Discrete Diffusion Model Speeds Up High-Quality Speech SynthesisVercel's skills.sh Now Lets You Bundle and Share Agent Skill PacksVercel Deploys Up to 33% Faster for Apps with Many ISR PagesGitHub Legal Team Uses Copilot CLI to Streamline Workflows Without CodingVercel's Eve Agent Now Gets a Full Browser: Click, Fill Forms, ScreenshotNova: An End-to-End MLIR Compiler for Deep LearningNew Executable Benchmark Evaluates AI Agents' Compositional Meta-Routing SkillsAI Agent Knows Everything About Me at 80% Lower CostOpenAI Responds to Third-Party Cybersecurity Evaluations, Announces New SafeguardsGoogle Announces July 2026 AI Updates: New Models and Featuresllm-anthropic 0.26 Adds Three Claude Models, WebSearch, and CodeExecution ToolsQwen 3.8 Max (2.4T) and 27B Open-Weight Models Released for Coding and Cowork

Latent Space published an external reconstruction of how ChatGPT Work functions, breaking down how Memory, Proactivity, Scheduling, Browser Use, Plugins, Skills, and Tools work together. This analysis provides an outside-in perspective for developers and AI enthusiasts to understand this agent product designed for a billion users.

Simon Willison released LLM 0.32 today, the most significant version since the project’s launch. It adds support for visible reasoning traces, server-side provider tools, redesigned content-addressable SQLite logs, and new features enabled by the OpenAI Responses API, along with new models. Developers can leverage these features to build smarter, more debuggable AI applications.

MiniMax released MiniMax-H3, a general-purpose, omni-modal generative system accepting text, images, audio, and video, generating up to 15-second video clips with audio. A Python package, minimax-h3-mlx, enables running on Apple Silicon. Users can input text or images to quickly produce short videos for prototyping or creative previews. Weights are open, so developers can use freely.

A new arXiv paper presents the first production-scale characterization of AI coding agents like GitHub Copilot, Claude Code, and Codex. It finds these agents interleave multi-step LLM inference with tool execution, creating a workload distinct from chatbots. Detailed analysis covers inference call frequency and tool sequences, providing data-driven insights for optimizing agent performance and system design.

A new arXiv paper introduces an uncertainty-aware simulation-based inference approach to improve LLM performance in operations research tasks. Standard autoregressive generation uses a myopic policy that fails to ensure coherent modeling. The method quantifies uncertainty to guide generation of more reasonable intermediate steps, boosting final answer accuracy. It could make LLMs more dependable for logistics, resource allocation, and other complex decision-making scenarios.

Claude Code released v2.1.222, fixing multiple security vulnerabilities: worktree-isolated sessions and their subagents could previously run destructive git commands against the main checkout; isolation now applies to file edits and Bash in every session type. It also fixed PreToolUse auto-allow hooks bypassing tool restrictions in background agent tasks. This update enhances security; users are advised to upgrade.

GitHub Blog explains how to handle huge AI-generated pull requests. Instead of one giant, un-reviewable PR, engineers can teach coding agents to decompose work into a clean, ordered stack of smaller PRs, each a logical unit. Reviewers can merge one at a time, dramatically reducing mental load and speeding up the merge process.

A new arXiv paper reveals that on-policy reinforcement learning with verifiable rewards (RLVR) can, while improving the current objective, make successful behaviors for subsequent objectives too rare to sample and reinforce. The researchers call this ‘verifier-induced support reshaping’. This implies that optimizing for short-term goals during AI training could harm its ability to learn long-term, complex tasks, offering crucial insights for RL training strategy design.

A new paper introduces DLLM-TTS, a text-to-speech system based on a block discrete diffusion language model. Current systems face a trade-off between quality and speed: autoregressive models offer high intelligibility but are slow and data-hungry, while non-autoregressive models are fast but sacrifice quality. DLLM-TTS aims to get the best of both worlds, using a novel block-level discrete diffusion approach to increase synthesis speed while maintaining high fidelity, offering new possibilities for real-time or low-latency speech applications.

Vercel’s skills.sh now supports skill packs, letting users bundle multiple agent skills into a shareable collection via a single URL or GitHub organization. Previously, sharing skills was piecemeal; now teams can standardize on one skill set across projects with a single link. This reduces overhead for sharing and managing agent skills, especially useful for team collaboration.

Vercel announces deploys are now up to 33% faster for apps using Incremental Static Regeneration (ISR). The more prerendered pages, the larger the savings. ISR combines static speed with dynamic flexibility, updating content without full rebuilds. Large ISR-heavy sites can now push updates faster, shortening release cycles.

GitHub Blog shares how its legal team used Copilot CLI to streamline daily workflows. Without writing a single line of code, the team built tools through natural language commands to automate processes. This case shows how non-technical teams can use AI coding assistants to solve real business problems and boost efficiency, offering a reference for legal, operations, and other non-engineering departments.

Vercel released agent-browser, an extension giving eve agents a full set of browser tools: navigate pages, read content, click, fill forms, take screenshots, inspect console and network activity. Everything runs inside the agent’s environment with no extra setup. Developers can now have agents simulate human browsing for interactive, complex tasks.

A new arXiv paper presents Nova, an end-to-end MLIR compiler for deep learning. It argues that performance at scale depends on how well high-level math ops map to underlying hardware. Nova optimizes this mapping via a full MLIR compilation pipeline, cutting intermediate conversion overhead. The compiler could improve training and inference efficiency for AI models.

An arXiv paper presents a new executable benchmark designed to evaluate the ‘compositional meta-routing’ capability of agentic systems. This capability involves deciding not only what answer to produce, but which reasoning and execution operations (like answering directly, decomposing a request, retrieving evidence, executing code, delegating to a specialist, or verifying) should precede it. The benchmark provides a standardized method for evaluating the workflow decision-making abilities of complex AI agents.

A report from Ben’s Bites highlights a new AI agent that can deeply understand user information while cutting operational costs by 80%. Through more efficient architecture, the agent reduces API call expenses while delivering personalized services. Developers and businesses relying on AI-driven personalization can now achieve higher-performance user profiling at a lower cost.

OpenAI has addressed recent third-party cybersecurity evaluation incidents involving its models and outlined new safeguards to strengthen AI model testing and evaluation. These measures aim to mitigate risks during evaluations and prevent misuse of models without proper authorization. OpenAI says it will continue collaborating with the security community to improve transparency and safety.

Google AI Blog has published a roundup of its latest AI updates from July 2026, covering new model releases, product feature improvements, and developer tool upgrades. While specific details have not been fully disclosed, Google says these updates will continue advancing practical AI applications, providing developers and enterprise users with more powerful tools. Full announcements are available on the blog.

Simon Willison’s llm-anthropic 0.26, enabled by LLM 0.32, adds three models—claude-fable-5, claude-sonnet-5, claude-opus-5—plus server-side tools including WebSearch, WebFetch, CodeExecution, and AnthropicMCP. These are accessible via the -T flag or Python tools= parameter. Developers can now leverage these new models and tools directly within the LLM framework.