Moonshot AI’s PerceptionBench tests how well multimodal AI models can actually “see,” separate from logical reasoning. No frontier model reaches 60 percent accuracy, and GPT-5.6 Sol leads by a narrow margin. Many supposed reasoning errors actually happen as early as the image-reading stage.
Anthropic will soon offer a watermark detection API that lets third parties check whether text was written by Claude. The technology builds on Google’s SynthID method and tweaks the randomness during word selection without affecting text quality, Anthropic says. The approach has limits with fact-heavy text, code, and heavy rewriting.
Alibaba's AI team Qwen has released the open model weights for Qwen3.8. The core model, Qwen3.8-27B, is a multimodal dense model with 27 billion parameters that, according to Qwen, outperforms the larger Qwen3.7-Plus in coding and office tasks. The team also touts improved agent capabilities, with the model planning more independently and completing tasks more reliably. It natively handles up to 262,000 tokens of context and can scale to one million using the YaRN method. Beyond text, it processes images and videos, including diagrams, documents, and multi-hour video. A flexible thinking mode is on by default but can be toggled per query.
Image: QwenThe weights ship under the Apache 2.0 license. Qwen has also released weights for the much larger Qwen3.8-2.4T-A95B, built to operate at the Max level. Both models are available on Hugging Face and ModelScope. A hosted version with one million tokens of context will soon be available through Qwen Cloud, Alibaba's AI service.
OpenAI’s Computer History records clicks, keystrokes, and app switches on Mac and turns them into a searchable timeline for ChatGPT and Codex. The data is stored locally as unencrypted Markdown files. OpenAI says it’s not used for AI training, but memories that feed into chats may still end up as training data.
AI agents using Claude Opus 4.8 and GPT-5.6 Sol were given six days, $3,000 in API credits, and GPU access to independently write AI research papers. The original authors of unpublished NeurIPS papers rated the results as “Reject.” According to the study, conducted with Princeton and the UK AI Security Institute, frontier models can handle the full research engineering process but fall short on research judgment, creative problem-solving, and the ability to abandon failed approaches.
Anthropic is testing whether Claude Code can handle daily maintenance of the company’s own apps, from crash fuzzing to dead-code removal. In a few weeks, the AI created 388 pull requests, and 46 percent were merged after human review. Claude Code inventor Boris Cherny sees this as “early signs of life that this might be possible.”
Chinese AI startup Zhipu AI has released GLM-5.3. The model shares the same base as its predecessor, GLM-5.2, and all gains come from extended post-training alone. Zhipu says GLM-5.3 is the most powerful open-weights coding model, with the biggest jumps in agent-based tasks.
Image: Z.ai
One area where top Chinese models like Kimi or Qwen still lag behind US frontier models is cybersecurity. Zhipu trained GLM-5.3 with data and environments built to find software vulnerabilities. According to Z.ai, the model "began to reason across multiple stages of exploitation, forming coherent plans for complete exploitation chains." Working with security teams in China, the company says it found 2,436 vulnerabilities across 269 projects, some up to 40 years old. The flaws are documented in a public registry.
Image: Z.ai
GLM-5.3 is available now through the GLM Coding Plan and works with coding agents like ZCode, Claude Code, or OpenCode. The model weights are set to go open source in two weeks, once security reviews wrap up.
With Studio 2.0, Suno turns its AI music platform into a chat-driven DAW (Digital Audio Workstation) for Premier subscribers. A new beta chat feature lets users talk to Studio like a bandmate, creating instruments, vocals, or custom plugins through plain text. The update also adds MIDI import, recording, and editing, stem separation, automation curves, and unrestricted multitrack export at 32-bit/48 kHz. Plugin creation doesn't burn any credits yet, though Suno says a credit system may come later.
The unlimited export feature stands out against the backdrop of Suno's recent crackdown on AI music spam. CEO Mikey Shulman had just rolled out new download limits for lower tiers to curb mass distribution of AI-generated songs on streaming platforms. Premier subscribers have always had unrestricted exports, a sign that Suno sees paying professionals as a lower spam risk.
The company has also drawn fire for using copyrighted material to train its music models. A German court recently ruled that the service violates German copyright law and that the training doesn't qualify under the U.S. "fair use" doctrine.
Update: A previous version of this article incorrectly stated that unlimited exports were a new feature for Premier subscribers. Premier subscribers have always had unrestricted exports.
