AI News: Huge Updates From Anthropic, OpenAI and Google

TL;DR

OpenAI, Anthropic, and Google all launched major app updates this week, alongside new models and agentic features transforming how AI tools work.

Key Points

• 1.OpenAI's Codex app is evolving into a super app. New features include background computer use with its own cursor, parallel agent workflows, in-app image generation via GPT Image 1.5, an in-app browser with comment-mode UI editing, and the ability to build, run, and self-test local desktop apps.
• 2.Anthropic updated Claude Code with parallel sessions and a redesigned UI. Users can now run multiple coding sessions simultaneously across different repos, with added features like an integrated terminal, in-app file editor, faster diff viewer, and expanded HTML/PDF preview pane.
• 3.Google launched the Gemini desktop app for Windows and Mac. It mirrors browser Gemini functionality — including image generation via Imagen, video via Veo, and music — and also added slash-command 'skills' to Chrome so saved prompts run on any webpage.
• 4.Google released Gemini 3.1 Flash TTS, a highly controllable text-to-speech model. It supports inline emotion tags like 'whisper,' 'panic,' and 'laughs,' enables two-speaker podcast-style audio, and is available in Vertex AI, Google Vids, and AI Studio.
• 5.Claude Opus 4.7 is Anthropic's new top coding model. On SWE-Bench Pro it scored 64.3% vs Opus 4.6's 53.4%, with the withheld Mythos preview scoring 77.8%; the biggest gains are in agentic coding, instruction following, and multimodal support.
• 6.Perplexity launched Personal Computer, an on-device agentic system. Unlike their cloud-based Perplexity Computer, this runs on your local machine — accessing local files, native apps, iMessage, and email — while still using Perplexity's remote inference servers; Matt Wolf plans a dedicated Mac Mini review.
• 7.OpenAI introduced GPT Rosalind, a reasoning model for life sciences. It targets biology, drug discovery, protein engineering, and genomics; outperforms Gemini and Grok on chemistry and experimental design benchmarks but is restricted to approved scientists and researchers only.
• 8.Several open-source models and quirky news rounded out the week. MiniMax M2.7 hit 56.2% on SWE-Bench Pro (non-commercial license); Alibaba's Qwen 3.6 35B is open-source and locally runnable; Allbirds rebranded as New Bird AI to sell GPUs after dropping from a $4B IPO to a $39M sale; Boston Dynamics showed a robot completing to-do list tasks read from a whiteboard.