Cursor punches up with Composer 2
🤔 “How do I explain to my mom that AI is involved in global warfare policy?” Share Handy AI with your family and friends to help them understand the crazy world of modern artificial intelligence (and save you some time).
last week’s top stories
🚀 Cursor’s Composer 2 coding model matches Claude Opus 4.6 at 86% lower cost. Composer 2 scores 61.3 on Cursor’s internal CursorBench, beating Claude Opus 4.6 at 58.2, while input pricing starts at $0.50 per million tokens versus Opus’s $15 and GPT-5.4’s $2.50. Trained exclusively on code data and optimized for long-horizon tasks requiring hundreds of individual actions, the model’s faster variant ships as the default experience.  Built on Moonshot AI’s open-source Kimi K2.5 base, Composer 2 proves that code-specialized fine-tuning on a strong open foundation can reach frontier performance at commodity pricing. Read more
🎬 Val Kilmer gets a posthumous AI lead role in “As Deep as the Grave.” First Line Films announced Kilmer has joined the cast of “As Deep as the Grave,” with producers noting he signed on to the role before his death but was unable to film due to his health.  His estate approved the generative AI performance, and the production compensated it per SAG guidelines.  Kilmer’s daughter Mercedes stated he always viewed emerging technologies with openness. Read more
📲 Claude Dispatch lets phones command desktop agents as persistent worker threads. Dispatch enables a persistent conversation with Claude that runs on your computer, letting users assign work remotely and retrieve completed results at their convenience, available as a research preview via Claude Cowork on Max plans.  Instead of starting a new session per task, the thread doesn’t reset, retaining context from previous tasks across phone and laptop. Read more
💬 Claude Code Channels turns Telegram and Discord into agent control panels. Claude Code’s new channels feature lets messages, notifications, and webhooks flow directly into a running session through MCP servers, supporting two-way communication via Telegram and Discord with the option to build custom channels. The feature requires Claude Code version 2.1.80 or later and a claude.ai login, with Team and Enterprise organizations needing to explicitly enable channels. Read more
🎯 OpenAI declares internal “code red,” kills side quests to chase Claude. The “do everything all at once” approach made it harder for the company to compete and react to rivals, with CEO of Applications Fidji Simo telling employees the company is “very much acting as if it’s a code red.” OpenAI credits pressure from Anthropic’s Claude Code and Cowork as a direct driver of the pivot. Abandoning Sora-style sprawl to double down on coding and enterprise is the correct call (about 18 months late). Read more
📱 OpenAI to merge ChatGPT, Codex, and Atlas into one desktop superapp. Simo told employees in an internal note: “We realized we were spreading our efforts across too many apps and stacks, and that we need to simplify our efforts.That fragmentation has been slowing us down.” Anthropic now captures 73% of first-time enterprise AI spending as Claude overtook ChatGPT as the most downloaded US app in March 2026.  OpenAI president Greg Brockman is overseeing the product overhaul. Read more
🪟 Cursor Glass debuts as a unified agentic coding workspace in alpha. Glass introduces a unified interface for managing agents, repositories, and cloud tasks in one place, with a Cloud Handoff feature that lets agents switch between local machines and cloud environments mid-task.  Shipping alongside Composer 2, it positions Cursor directly against Claude Code and OpenAI Codex as an agent-first IDE replacement. Read more
⚡ OpenAI drops GPT-5.4 mini and nano for subagent-era high-volume coding. GPT-5.4 mini runs more than 2x faster than its predecessor while approaching GPT-5.4 performance on SWE-Bench Pro and OSWorld-Verified, designed for coding assistants delegating narrower subtasks in parallel.  GPT-5.4 nano targets classification, data extraction, and subagent tasks at $0.20 per million input tokens. Read more
🎨 Google Stitch becomes a voice-driven “vibe design” platform. Google redesigned Stitch with an infinite canvas, voice interaction, and an SDK and MCP server connecting it to coding assistants including Claude Code, Gemini CLI, and Cursor, sending Figma shares down 8%. The tool now features a design agent that tracks progress across the full project’s evolution, available free with 350 standard generations per month. Read more
🌍 Anthropic’s 80,508-person global survey exposes AI’s “light and shade” paradox. People in Sub-Saharan Africa and Asia expressed more optimism about AI than those in Western Europe and North America, with economic gains from AI forming the main aspiration for most respondents.  Unreliability ranked as the top concern at 26.7%, with job displacement second at 22.3% and loss of human agency third at 21.9%.  The tension Anthropic calls “light and shade” (where the features users love most are identical to their deepest fears) is the most honest framing of the adoption problem anyone has published. Read more
🧪 AI Research of the Week
What 81,000 People Want from AI
From Anthropic
Jake's Take: Anthropic turned Claude into a conversational interviewer to chat with over 80,000 people across 159 countries. From this interviews, they’ve coined a paradox called "light and shade": the realization that the exact AI features we love the most are also the ones that terrify us the deepest. While users are chasing massive productivity and economic gains, they are simultaneously kept up at night by the tech's unreliability (the #1 concern at 26.7%), job displacement (22.3%), and the loss of human agency (21.9%).
What’s uniquely fascinating is the geographical divide: folks in Sub-Saharan Africa and Asia are embracing the economic upside with profound optimism, while Western Europe and North America remain much more anxious and skeptical.
This report provides the a very honest framing of the AI adoption problem, proving that the world isn't neatly divided into AI optimists and doomers, but rather filled with people experiencing both hope and fear at the exact same time.
and then, even more news…
🖼️ Microsoft’s MAI-Image-2 debuts at #3 on Arena.ai, rolling into Copilot now. MAI-Image-2 placed directly behind Google’s Gemini 3.1 Flash and OpenAI’s GPT Image 1.5 on launch day, released by the Microsoft AI Superintelligence team led by Mustafa Suleyman.  The model targets photographers and designers, built for natural light, accurate skin tones, and environments that reduce post-production work.  A diffusion-based architecture with 10B-50B parameters is solid, but the 1:1-only output restriction and 15-image daily cap will frustrate anyone outside the playground. Read more