This week, every major AI company shipped a new app redesign on the same day. The twist? The real story is not Opus 4.7, it’s the quiet super app race happening in the background.
Matt Wolfe, the creator behind Future Tools, rounded up the whole dump and tested most of it live. His breakdown surfaced the signals worth keeping, and I pulled the juiciest bits for you.
🖥️ The Codex shift
The original poster walked through the new Codex desktop app from OpenAI, and it’s starting to look like their long-rumored super app. A few things he tried that stood out:
- Generate images directly inside Codex (no more jumping to ChatGPT)
- Background computer use, so multiple agents run in parallel while you keep working in other apps
- An in-app browser with comment mode, where you highlight a section of a webpage and Codex edits it
- He had it build a Connect 4 desktop app, then asked it to play the game itself to test the UX
🧪 The mini-workflow he demoed
- Open Codex, start a fresh project
- Prompt a mockup image (it uses GPT Image 1.5 in-app)
- Say “build this site from the mockup”
- Use comment mode to tweak elements visually
- Let it run and test its own output
🔧 Other drops the author flagged
- Claude Code desktop now runs parallel sessions across repos, with an integrated terminal and file editor. CLI is becoming optional.
- Gemini launched a native Mac app plus a Windows desktop app rollout.
- Chrome added Skills, basically slash commands for saved prompts. This mirrors what Perplexity Comet already does.
- Gemini 3.1 Flash TTS lets you tag emotion inline, like “whispers,” “panic,” “laughs,” and it actually follows them.
- Nano Banana now pulls personal context from your Gmail, Calendar, Drive, and Photos to generate images about your actual life.
- Perplexity Personal Computer runs agentic workflows on your own Mac, reading local files, iMessage, and email.
- Canva AI 2.0 (coming soon) connects to Slack, Notion, Gmail, and learns your design style over time.
🧠 The model news
The expert called Opus 4.7 the new coding king, with a big leap on SWE-Bench Pro. Most non-coders won’t feel it much, but instruction-following is noticeably better. On the open source side, MiniMax M2.7 and Qwen 3.6 35B A3B both dropped with solid benchmarks. OpenAI also released GPT-Rosalind for life sciences, restricted to vetted researchers, aimed at biology, chemistry, and drug discovery.
⚡ Pro tips from the roundup
- If you use Claude for coding, switch to Opus 4.7 in Cursor or Claude Code today.
- For cheaper, faster images with short text, try Microsoft’s MAI-Image-2-Efficient. Keep the original MAI-Image-2 for longer text.
- Editors on DaVinci Resolve 21 should try AI IntelliSearch to hunt faces and keywords across B-roll.
🤖 The weirdest bit
This savvy professional closed with two wild stories. Allbirds, the shoe company, pivoted to AI and their stock jumped 600 percent. And Boston Dynamics trained a robot to read tasks off a whiteboard, then go crush cans, sort laundry, and walk the dog. Prompt engineering in the physical world.
Watch the full video from Matt Wolfe for the live demos, the Connect 4 build, and the Gemini TTS podcast test. Worth the 35 minutes.