GPT-6 Astra Shows a "Step Change" on a Robotics Spatial Reasoning Benchmark

According to The Decoder, OpenAI's GPT-6 Astra has demonstrated substantial progress in spatial understanding on a new robotics benchmark. On "StationeryBench," which uses a dual-arm robot, it completed 7 out of 100 tasks, while the competing MolmoAct2 completed none.

On its own, a 7% figure looks small, but researchers describe it as a "step change" — a qualitative leap — in spatial reasoning. Physical-world manipulation tasks demand a level of spatial grasp that goes a step beyond language tasks. The fact that comparisons between top-tier models are now producing a difference of "zero versus nonzero completed tasks" suggests the distance to real robotics applications is shifting.

OpenAI: "Keep Prompts for GPT-6 Astra Concise" — Lengthy Instructions and Rigid Approval Rules Get in the Way

The Decoder also reported on prompt design advice for GPT-6 Astra from OpenAI's Eric Provencher. Overly long skill descriptions, blanket documentation-reading requirements, and rigid approval rules actually get in the model's way.

The gist: highly capable models don't need excessive hand-holding; instructions should be tied to specific tasks, and task completion criteria should be stated explicitly. As models get smarter, operational design shifts from "piling things in" to "trimming things back." For agent implementers, this is a topic with direct practical implications.

AI's "Written Reasoning Steps" Can Be Identified in Internal States — New Findings for Safety Research

New research shows that steps an AI model outputs during reasoning — such as "calculation," "formula lookup," and "deduction" — are clearly separable in the model's internal states, particularly in its intermediate layers (The Decoder).

What makes this interesting are the safety implications. The model internally processes more information than the visible chain of thought reveals. In other words, the written reasoning alone does not allow us to fully know from the outside what a model is thinking. The research adds an internal-state analysis approach to the debate over reasoning transparency.

Tencent's "AuK" — A 1.5B Foundation Model That Handles Everything from Audio Generation to Editing and Separation via a Single Instruction

AuK, an audio foundation model released by Tencent's Hunyuan team, is generating buzz in the local LLM community. Trained on millions of hours of audio data, the 1.5B model covers zero-shot TTS, instruction-based TTS without a reference voice, acoustic editing such as speech rate, pitch, volume, and emotion, lyrics rewriting, noise removal, and even speaker separation — all controllable uniformly through natural-language instructions.

AuK-Flash, a variant distilled for fast inference, runs in just 4 inference steps. Beyond generating speech from text, its distinguishing feature is the ability to edit the content and acoustic properties of recorded audio after the fact — a design that appears aimed at consolidating audio workflows into a single model. The model is publicly available on Hugging Face.

Claude Code v2.1.269 — Fixes Permission Bypass via tee, Adds plugin eval Command

According to an article on Qiita, Anthropic released Claude Code v2.1.269 on September 11, 2026. The headline feature is the new "claude plugin eval" command, which enables quantitative verification of plugin quality, but even more noteworthy is a fix for a permission bypass via the tee command — an operation that slips past permission checks.

In agent-style tools, permission management for command execution is the cornerstone of safety. A hole that allows checks to be bypassed using basic shell features like pipes and redirection could be triggered unintentionally in everyday operations, not only by malicious input. The accumulation of fixes like these will be a precondition for trust in AI coding tools.