Anthropic Makes Auto Mode the Default in Claude Apps

Anthropic said on October 22, 2026, that starting next week, Claude sessions will run in auto mode by default on Pro and Max plans as well as Team plans. The change covers the web and desktop apps along with the mobile app. Claude will stop asking you to approve each step. An automated check blocks risky actions instead. If you let Claude work with your connected apps or files, it will pause less often for a yes, and Anthropic's check, not you, decides which actions count as risky. Claude Code, Anthropic's coding tool, made the same switch in August. Anthropic's announcement ↗

METR Finds Gaps in OpenAI's Monitoring of Its Own Agents

On October 22, 2026, METR, an independent AI safety testing group, published a review of how OpenAI watches its internal AI agents for misbehavior. Reviewers from METR and Redwood Research spent three weeks inside OpenAI this summer. They found a way past the sandbox in Codex, OpenAI's coding agent. A sandbox is the walled-off space that limits what an agent can touch. They also found a trick that kept hundreds of recent actions out of the logs. OpenAI says it has fixed both problems. METR has not checked the current system. METR's write-up ↗

Artificial Analysis Ranks Claude Code First on New Coding Index

Artificial Analysis, an independent testing group, released version 2.0 of its Coding Agent Index on October 23, 2026. The index ranks pairs: a coding app plus the model running inside it. Its new test, FrontierCode from Cognition, grades 100 tasks on whether a project's maintainer would accept the change. Average scores fell from 54.8 to 47.4. Harder test, lower numbers. Claude Code running Opus 5.5 leads at 61.7, at $11.42 a task. OpenAI's Codex running DeepSeek V4.1 Flash scored 47.3 for about 39 cents a task. the full index ↗

Update: MiniMax Releases M3.1 Flash for Download

MiniMax, a Chinese AI company, published the files for its M3.1 Flash model on Hugging Face on October 23, 2026. News Desk covered the paid online version on October 20. Now companies can run the model on their own servers. The license is MiniMax's own, not a standard open-source one. Non-commercial use is free. Commercial products must credit MiniMax and notify it by email, and a business earning over $20 million a year from those products needs written permission first. the model page ↗

Anthropic Uses Claude to Prove 123 Code Libraries Won't Crash

Anthropic's Frontier Red Team said on October 22, 2026, that Claude, working with the company Theorem, wrote proofs that 123 widely used open-source code libraries cannot crash. A separate program checked each proof. The guarantee holds under stated limits, such as running on Linux with inputs under 64 MB. Along the way, Claude found and fixed real bugs. Proofs like these once took experts so long that they were used mostly on small programs. They cover crashes only, not every kind of bug. Anthropic's paper ↗

Artificial Analysis Ranks AI Models by 'Slop' in Their Writing

Artificial Analysis released AA-PersonalityBench on October 22, 2026. The test scores a model's writing style on what the group calls slop. Less capable models generally produced more of it. But models of equal ability can write very differently. GPT-6 Sol and Muse Spark 1.3 both score 48 on the group's Intelligence Index, yet they sit at opposite ends of the slop ranking. The test measures style, not ability. the ranking ↗

Anthropic Opens Full Claude Platform on Google Cloud

On October 22, 2026, Anthropic made its own version of Claude on Google Cloud generally available. Companies sign in and pay through their Google Cloud accounts. They get the same features as Anthropic's direct service, on the same day those features launch, including a faster mode for Opus 5.5. OpenAI's models, which News Desk covered on October 22, are still in private preview on Google Cloud. OpenAI says wider access will come in the coming weeks. Anthropic's announcement ↗

Graphify Maps a Codebase for AI Coding Assistants

Graphify gained 1,408 stars on GitHub in one day. The free tool reads a software project along with its documents and PDFs, plus its database layouts, and builds a map of how the pieces connect. Claude Code and Cursor can then search that map to find their way around the project, as can Codex and Gemini CLI. Everything runs on your own computer. Read more ↗

claude-howto Teaches Claude Code Through Examples

claude-howto gained 177 stars in one day. The free, illustrated guide walks through Claude Code, starting with basic concepts and ending with advanced agent setups. Each section comes with templates you can copy and paste into your own projects. Read more ↗

ComfyUI Runs Image and Video Models on Your Computer

ComfyUI gained 4,375 stars this week. The app is free. It runs AI image and video models on your own machine, and you build each job by linking boxes in a visual flowchart instead of writing code. Read more ↗