Claude Opus 4.7 Enhances Long‑Task Handling & Qwen 3.6‑35B‑A3B Open‑Source Release

Model Releases

Claude Opus 4.7 Improves Long‑Task Processing

Anthropic announced the release of Claude Opus 4.7, branding it as the most powerful Opus model to date. The update focuses on handling long‑running tasks, employing stricter execution flows and more precise instruction following, and it can verify its own output before reporting results. Engineer Boris Cherny reported that internal testing showed "incredibly efficient" performance. Because Opus 4.7 consumes more thinking tokens, Anthropic raised the rate‑limit quotas for all subscription tiers as compensation.

Alibaba Open‑Sources Qwen 3.6‑35B‑A3B

Alibaba’s Tongyi Qianwen team released Qwen 3.6‑35B‑A3B, a sparse‑Mixture‑of‑Experts model with 35 billion total parameters but only 3 billion active parameters, under the Apache 2.0 license. According to independent researcher Simon Willison, the 21 GB model generates pelican images that appear superior to those from Opus 4.7, and its agent‑programming capability rivals models ten times larger in parameter count. The model is available on HuggingFace and ModelScope.

Development Ecosystem

Anthropic Engineer Shares Opus 4.7 Usage Tips

Boris Cherny, a core member of the Claude Code team, outlined practical tips for developers: leverage the model’s deep‑thinking ability for complex multi‑step tasks and monitor the output‑verification stage. He also noted the rate‑limit increase to offset higher token consumption.

OpenAI Codex Extends Functionality

The new Codex version adds computer‑use capabilities, an in‑app browser, image generation and editing, over 90 new plugins for various services, multi‑device operation, SSH access to development boxes, automatic thread automation, and rich document editing. Co‑founder Greg Brockman observed that Codex can automatically gather dispersed information across Slack, Google Docs, and Notion, performing cross‑platform integration that feels "like real magic."

GLM‑5.1 Tool‑Call Fix for vLLM/SGLang

The Z.ai team identified a mismatch where vLLM and SGLang automatically convert tool messages to an array format [{"type": "text", "text": "..."}], while the original chat template only accepts a plain string, causing the model to loop on tool calls. The fix is to replace the chat_template.jinja file from the HuggingFace repository with the updated version.

Technical Insights

Ternary Bonsai Achieves 1.58‑bit Efficiency

PrismML introduced the Ternary Bonsai series, which replaces binary weights with ternary values (‑1, 0, +1). This yields an extreme parameter efficiency of 1.58 bits per weight, shrinking model size nine‑fold compared with 16‑bit counterparts while surpassing most same‑size models on standard benchmarks. The team open‑sourced 8 B (1.75 GB), 4 B (0.86 GB), and 1.7 B (0.37 GB) variants under Apache 2.0, making high‑performance AI feasible on edge devices.

Tencent Releases HY‑World 2.0

Tencent’s Hunyuan team open‑sourced HY‑World 2.0, a multimodal world model that can generate, reconstruct, and simulate interactive 3D environments from text, images, or video. Key features include one‑click world generation, pipelines for Unity and Unreal Engine output (mesh, 3DGS, point‑cloud formats), and a physics‑aware real‑time exploration mode, targeting game‑development pipelines and embodied simulation workflows.

Sim2Reason Trains LLMs in a Virtual Physics World

A research team proposed Sim2Reason, which trains large language models inside a virtual environment governed by real physical laws, eliminating the need for manual labeling. By letting the AI observe cause‑and‑effect relationships, the method achieves a 5‑10 % zero‑shot improvement on International Physics Olympiad questions. This approach bridges deep learning with physics simulation for better physical reasoning.

Product Updates

Claude 4.7 Rate‑Limit Increase

Anthropic confirmed that, due to higher thinking‑token usage in Opus 4.7, all subscription tiers have received a substantial rate‑limit boost, with both 5‑hour and weekly limits reset. The ClaudeDevs account also fixed a previous rate‑limit calculation bug.

Google DeepMind + Boston Dynamics: Gemini on Spot

Google DeepMind announced a partnership with Boston Dynamics to deploy the Gemini Robotics embodied‑reasoning model on the Spot robot. Gemini enables Spot to better perceive its surroundings, recognize objects, and follow simple commands such as tidying a room, marking a significant step toward commercializing embodied intelligence.

Physical Intelligence Releases π0.7 Robot Model

Physical Intelligence unveiled π0.7, a controllable general‑purpose robot model that demonstrates emerging embodied‑intelligence capabilities. While detailed technical specifications remain forthcoming, the release is expected to advance robot‑control research.