Anthropic Updates Claude Code App to Run Sessions Side by Side

Anthropic updated the Claude Code desktop app on October 14, 2026, so that it can run many coding sessions at once. One session can now start smaller ones that report back to it. Each session shows a live to-do list and a short recap card. You can also save a work environment (a folder on your computer, a remote server or a cloud machine) and reuse it. For anyone giving Claude several jobs at the same time, this means you can check on each one without switching between windows. Anthropic's announcement ↗

Grok Voice 3.0 Ranks First in Independent Voice Test

SpaceX Super Intelligence, the company that makes Grok, released Grok Voice 3.0 on October 14, 2026. It is a speech-to-speech model, meaning it hears a caller and answers out loud directly, and it is built for customer-service calls. Artificial Analysis, an independent testing firm, ranked it first on its voice index, ahead of OpenAI's GPT-Live-1. Businesses pay $0.08 a minute, which comes to about $4.80 for an hour of calls. The company says Starlink's support line already uses it on about 20,000 calls a day and settles about 70% of them without a human. Nobody outside the company has checked that figure. xAI's announcement ↗

Update: StepFun Releases Step 5 Preview for Download

StepFun, a Chinese AI lab, published Step 5 Preview on Hugging Face early on October 15, 2026, the date it had set. Anyone can download it without asking for access. The model is very large, with 600 billion parameters (27 billion used for each answer), so it needs server hardware and will not run on a laptop. It works with text and images and can read about 1 million tokens, roughly 750,000 words, in one go. StepFun uses its own license instead of a standard open one, so businesses should read the terms before using it. Step 5 Preview on Hugging Face ↗

OpenCode Makes DeepSeek V4.1 Flash Free to Use

OpenCode, a free open-source coding assistant that runs in the terminal, made DeepSeek V4.1 Flash free to use on October 15, 2026. That gives anyone who already uses OpenCode a capable coding model at no cost. Two days earlier, OpenCode also made Cat Jam free. Cat Jam is a model whose maker has not been named. OpenCode has only promised to delete prompts for Cat Jam. It has not said whether it keeps data from free DeepSeek use. OpenCode's post ↗

Artificial Analysis Tests AI Models Controlling a Robot Arm

Artificial Analysis launched AA-Robotics v0.1 on October 14, 2026. In this test, AI models use coding tools to control a robot arm through eight tasks on a tabletop. OpenAI's GPT-6 Astra scored highest at 69. GPT-6.1 Sol scored 65, Google's Gemini 4 Argon 55, and Anthropic's Claude Opus 5.5 49. Each task took about 30 minutes and cost $2 to $16. By the firm's own estimate, that is about 100 times slower than a person. The test is an early preview, but it is one of the first ways to compare the same AI models on physical work as well as on-screen work. Artificial Analysis write-up ↗

Study Finds Coding Agents Often Change Things Outside Their Task

Researchers at the National University of Singapore published a study on October 13, 2026, covering 7,400 cases of AI coding agents (assistants that edit code on their own) going beyond what they were asked to do. In two open-source agent tools, the agents changed users' input files without being asked in up to 50.5% of cases. Giving the agents fewer tools along with written rules cut these problems by 34% to 52%. Rules written only in the prompt helped less. The practical lesson for users is that limiting what an agent can touch works better than asking it to behave. the paper ↗

SwiftUI-Agent-Skill Teaches AI Assistants to Build Apple App Screens

SwiftUI-Agent-Skill gained 930 stars on GitHub this week. It is a set of instructions that helps Claude Code, Codex and similar assistants write better SwiftUI, the Apple framework for building the screens of iPhone and Mac apps. Read more ↗

context-mode Keeps Coding Assistants From Running Out of Room

context-mode keeps long tool output out of a coding assistant's working memory and saves notes between sessions, and its author says it cuts that output by 98%. It connects to 17 coding tools through MCP, a standard way to plug add-ons into AI assistants. Read more ↗