GPT | ChatGPT | Midjourney — News

Every Claude Code user should install /stack. This mod gives you a full overview of everything installed in Claude Code: skills, subagents, commands, hooks, MCP servers, plugins, settings, and output styles. Everything is grouped by configuration level — Enterprise, User, Project, Plugin, and CLI flags. The module runs a full scan and shows results in an interactive panel, letting you see which components are active, disabled, or being overridden by others. This finally helps explain why your Claude Code sessions eat up so many tokens and why Claude behaves the way it does. You can also explore all built-in commands, skills, and subagents shipped with Claude Code — including hidden ones not mentioned in the official docs. Install it with: npx claude-code-templates@latest --mod ui/stack-viewer

🤖 Anthropic launches Claude Duet — a personal agent that messages you first. Unlike earlier Claude agents that required you to launch them and wait, Duet works in a messenger-style thread: you drop a task, it runs in the background on its own cloud computer, and pings you when it's done or needs a decision. Each morning it sends a summary and otherwise stays quiet unless something important comes up. Duet taps into your memory, skills, and apps from your Claude account, connecting to 6,000+ services like Gmail, Calendar, Slack, Stripe, and HubSpot. It always asks permission before sending emails or making payments. Currently a research preview with a waitlist, available only to Pro and Max users in the US, with apps for iOS, Android, and web. 🔗 Claude Duet

OpenAI and Google announce partnership OpenAI's models will soon be available in Google Antigravity and Gemini Agent. Google already offers access to Gemini and Claude models, and now GPT models will join the lineup. This means Antigravity users will be able to work with all frontier AI models in one place — a notable move as former rivals now collaborate. 🤟 Read more

A guy used Claude to reverse-engineer his TV. He recreated a minimalist, Apple TV-style home screen, stripped out all pre-installed ads and bloatware, and cut RAM usage by 70%, noticeably speeding up performance. An open-source version is available for anyone who wants to try it. 😸

Mistral unveils Voxtral 3 — a new generation of voice AI models for building voice agents. Key features: - 50+ languages for speech recognition and speaker identification - Sub-second latency for real-time conversations - 35+ languages for voice generation with control over emotion, intonation, and pace - Voice cloning using just 30 minutes of audio - AI agents capable of handling calls, working with APIs, and completing tasks - Local deployment on your own infrastructure API pricing: speech recognition from $0.0035/min, speech generation at $13 per million characters. According to Mistral's benchmarks, Voxtral 3 outperforms Whisper, Gemini 3.5 Transcribe, and Scribe v2 in speech recognition. The model is designed to power voice AI agents that don't just talk, but actually complete real tasks. mistral.ai/news/voxtral-3

DeepSeek open-sources DeepInfer, its own inference engine for running AI models, already powering DeepSeek-V4.1-Flash with high performance and minimal latency. Key highlights: - Faster inference — handles more requests in less time - Efficient GPU utilization — lower computing costs - Parallel processing — high throughput under heavy loads - Reduced AI costs — fewer resources per response Released under the Apache 2.0 license, DeepInfer lets developers integrate DeepSeek's technology into their own projects. GitHub: https://github.com/deepseek-ai/DeepInfer

Job hunting? Try ai-job-search in Claude Code (50,000 stars on GitHub). Paste a job posting link, and the tool drafts a tailored resume and cover letter. A second AI agent then reviews both, fixes weaknesses, and returns only the corrected versions. It can also scan multiple job postings and rank the best matches. Before an interview, it predicts likely questions based on the job description and your submitted resume, and can even run a mock interview with you. A public fork adds a /setup command to save your contact info and salary expectations. To set it up privately, send Claude Code this instruction: "Copy https://github.com/MadsLorentzen/ai-job-search into a new private repository of mine. Don't create a fork. Install the tool following the setup instructions from the repository, then walk me through the /setup command step by step." 🔗 GitHub repo

Three plugins to boost your Claude Code workflow: • Image Preview — shows a preview of images on hover. • Sub Meter — continuously tracks your subscription limit usage. • Commit — helps you safely push changes to your repository using Haiku. 🔗 https://betterclaude.dev

🔥 MiniMax M3.1 Flash: three times faster and three times cheaper than M3, with weights releasing October 23. The model known as "Space Bunny," which topped global rankings after processing over 200 trillion tokens, turns out to have been an early version of M3.1 Flash. MiniMax reports strong benchmark results: 62% on Automation-Bench, 30% on ALE, 74% on DeepSWE, and 91% on BabyVision — all while costing a third of M3 and running three times faster. API pricing: $0.15 per 1M input tokens, $0.60 per 1M output tokens, and $0.015 for cache reads. Notably, just days after DeepSeek 4.1 Flash launched, MiniMax's new Flash model is already beating it on several key benchmarks. You can try M3.1 Flash via the API, the M Plan subscription, or the desktop MiniMax Code app. Source: platform.minimax.io #MiniMax

shot — a mod for Claude Code and an MCP tool that gives an AI agent access to screenshot-taking tools. Now you can ask Claude to explain step by step how to perform an action in an app, and it will show each step right on the screen. 👀 In the example, Claude finds unused code on Amazon's homepage using Chrome DevTools (menu ⋮ → More tools → Coverage). Out of 9.2 MB of JavaScript and CSS, 5.3 MB was unused. Coverage lets you determine which parts of JavaScript and CSS are not used on a page. Try running it the first time your site loads. Installation: /plugin marketplace add davekiss/mods /plugin install shot@davekiss

xAI releases Grok Voice 3.0 xAI has launched its new speech model, Grok Voice 3.0, which operates in speech-to-speech format — listening and responding with voice directly, without an intermediate text step. The company claims response times under one second. On the Artificial Analysis Speech-to-Speech Index, it scored 86.9 points, beating GPT-Live-1 (83.2) and Gemini 3.8 Live (82.6). On Full Duplex Bench, which measures pauses, turn-taking, and interruptions, it hit 97%. However, on τ-Voice, a test simulating support calls, it trailed GPT-Live-1 with 71.6% versus 74.5%. The model currently does not support Russian. Grok Voice 3.0 is already deployed in Tesla vehicles and Starlink support, handling 20,000 calls daily and resolving about 70% of them without human help — from customer verification to arranging equipment replacements. Via the Voice Agent API, pricing remains unchanged from the previous version at 8 cents per minute. #news #ai #ml

New open-source testing framework: e2e A next-gen framework for testing web and mobile apps has been released, with fully open source code. ✓ Combines AI with deterministic tests ✓ Works with OpenAI, Claude, and any other AI models ✓ Offers a unified API for web, iOS, and Android 🔗 GitHub repository

🔧 Google has introduced Harness², an AI agent that improves its own working environment. An agent's performance depends not just on the model, but on its "harness" — system prompt, tools, memory, skills, and subagents. Harness² lets this setup be adjusted for new tasks without retraining the model. The agent proposes changes, tests several variants on the same task, compares run behavior, and assembles a better configuration — no predefined correct answer or external evaluator needed. After each task, the system saves the chosen setup and confirmed insights, using them as a starting point for the next improvement cycle. In tests with Gemini 3.5 Flash and GLM-5, a single improvement step boosted results by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points in partial pass rate on LAB. Further iterations improved results even more. Configurations can be tested in parallel or refined step by step. OpenCode and Codex are supported. GitHub

Developers previously had to manually copy code from AI chatbots and paste it into their projects themselves, without any direct integration between the chat interface and the codebase.

🪰 "The Matrix" for a fly: brain simulation meets AI-generated world Odyssey announced it has connected a software model of a fruit fly's brain to its Odyssey-3 system. According to the team, the simulated fly now navigates AI-generated virtual worlds entirely on its own. Odyssey-3 builds interactive environments that react and change in response to actions taken within them, allowing researchers to study agent behavior in worlds that have never existed in reality. Source #Odyssey

Claude Code gets 3x faster — thanks to a '90s trick Developers have sped up Claude Code threefold by rediscovering an old-school solution: gzip compression. Instead of chasing complex modern optimizations, the team simply applied this decades-old technology to cut down data size and speed up processing significantly. The result shows that sometimes the best performance fixes aren't new innovations, but proven tools from the past that got overlooked in the rush toward newer tech.

A developer compared Astra and DeepSeek V4.1-1017 Flash using the same prompt and maximum reasoning level, with surprising results. Astra: finished in 20 minutes, cost $7, nailed the details, and produced something more original than typical GPT outputs. DeepSeek: took 1 hour 40 minutes, cost just $0.90, and focused more on the environment, which looked absolutely stunning. However, its koi fish didn't turn out too well.

✔️ ChatGPT for Chrome now reads tabs and controls the browser OpenAI updated its ChatGPT browser extension, adding the ability to read open tabs via the @chrome command and control browser actions directly. In chat mode, the assistant can parse data from several pages at once — e.g. comparing prices across suppliers or summarizing long texts. Results and full chat history now appear in a redesigned sidebar. For more advanced tasks, OpenAI introduced a new Work mode, powered by Computer Use technology. It allows ChatGPT to interact with web apps directly: filling forms, downloading files, sending messages, and organizing tabs. The update is rolling out to all ChatGPT plans, and the refreshed extension is already available. #news #ai #ml