Meta Releases Muse Wonder, Its Most Capable AI Models

Meta released Muse Wonder and Muse Wonder Full on October 12, 2026. Both are built for agents, meaning AI that carries out long, multi-step jobs on its own. Wonder is available through Meta's developer API and Muse Code, Meta's coding app. It costs $4 per million tokens read and $20 per million written, and a million tokens is about 750,000 words. Wonder Full is only in paid Muse Code plans. Meta's own tests put Full ahead of Claude Opus 5.5 on two coding tests. Artificial Analysis, an independent testing firm, gives Full a 56 on its overall index. Opus 5.5 gets 58. Meta's announcement ↗

Grok, Cursor and X Premium Merge Into One Subscription

SpaceX Super Intelligence, the company behind Grok, merged its Grok and Cursor plans with X Premium into The X Subscription on October 12, 2026. The change matters most for people who use Claude or GPT models inside Cursor, a popular coding app. Those models move to API pricing, so you pay for each use instead of drawing on your plan's allowance. Current Cursor and SuperGrok subscribers can keep their plan or switch. Unused time is refunded. Cursor's switch page ↗

Anthropic Says New Defenses Sharply Cut Claude Copying Attempts

Anthropic said on October 12, 2026 that its new defenses cut the reasoning text that copiers pull from Claude by 10 to 100 times. The practice is called distillation: training a cheaper model on another model's answers. Its filter wrongly flags fewer than 1 in 1,000 normal requests, Anthropic says. But since September 24, flagged requests are still charged. Reuters reports that tens of thousands of accounts were banned and that Anthropic named Z.ai, a Chinese AI lab. Anthropic's research post ↗

New Coding Test Puts Claude Opus 5.5 First at 52%

Artificial Analysis opened a public leaderboard on October 12, 2026 for FrontierCode 1.1. The test was built by Cognition, maker of the Devin coding agent, and has 100 private tasks written by people who maintain real software projects. A task only counts if the code works and is clean enough to be accepted into the project. Nine models from six companies scored between 43% and 52%. Costs vary much more than scores. GPT-6.1 Sol scored 49% for about $0.81 a task. Opus 5.5 cost $10.15 a task. the leaderboard ↗

OpenBMB Releases MiniCPM-V 4.7 to Run AI on Phones

OpenBMB released MiniCPM-V 4.7 on October 12, 2026. The free models read both images and text, and they are built to run on phones and tablets without sending data to a server. The Apache 2.0 license allows commercial use. OpenBMB says the larger version writes more than 13 tokens a second on an iPhone 17 Pro Max, about ten words a second, using 3 GB of memory. No one has tested that claim independently yet. On-device models keep your photos and documents on the phone. the project page ↗

Z.ai Releases GLM-Legal for Contract Review

Z.ai introduced GLM-Legal on October 12, 2026, a model built for legal work such as contract review and evidence analysis, as well as compliance checks. Law firms and legal teams have to apply for API access. On Harvey's LAB legal test, Z.ai says GLM-Legal fully completed 19.2% of tasks. That is 12.5 points better than the model it is based on. Z.ai's blog ↗

rea Lets AI Coding Assistants Analyze Software Without Source Code

rea is on GitHub's weekly trending list. It is an MCP server and command-line tool that lets Claude Code and Cursor, plus Codex and Gemini CLI, study software without its source code, from Android APKs and firmware to websites, 163.com reported. It links its conclusions to evidence. Accuracy and sandboxing were not reported, nor was the license. the GitHub page ↗

agent-skills Collects Engineering Instructions for AI Coding Agents

agent-skills, from GitHub user addyosmani, is on GitHub's trending list this month. The repo describes itself as "production-grade engineering skills for AI coding agents." Skills are instruction files that a coding assistant loads so it follows a set way of working. the GitHub page ↗