OpenAI’s Codec App Evolves into a Computer Agent
The artificial intelligence world saw a flurry of updates this week, with major players like OpenAI, Anthropic, and Google releasing new features and models. While many focused on new AI models, OpenAI’s significant update to its Codec app signals a move towards AI agents that can operate your computer.
Codec is no longer just for coding. The updated app can now generate images, remember user preferences, learn from past actions, and handle repetitive tasks. This evolution makes it seem like a step towards OpenAI’s rumored “super app,” a single platform integrating chat, image generation, and code generation.
Codec Takes Control of Your Computer
The most striking new capability of Codec is its ability to operate your computer alongside you. It can see, click, and type with its own cursor, and multiple AI agents can work on your Mac simultaneously without interrupting your own tasks.
This is a big change from earlier AI tools, which often took over your entire computer. Codec allows you to work in one application while the AI handles tasks in another. The app also includes a built-in browser, letting you comment on web pages to give context to Codec for specific tasks.
New Features in Action
During a demonstration, Codec successfully generated a website mockup for surfboards and tacos. It then built the site based on that mockup, a process that took just over six minutes. Afterward, the AI was instructed to add a surfboard and taco truck image as a background, which it did in under two minutes, ensuring it didn’t cover the text.
Codec also proved capable of creating a desktop application. A request to build a local Connect 4 game for Mac resulted in a functional app. The AI even tested the game by playing against itself, demonstrating its ability to build, run, and test applications.
Anthropic and Google Also Update Their Tools
Anthropic updated its Claude Code application, introducing parallel processing. This allows users to run coding tasks across multiple projects or different parts of the same project simultaneously. The interface has been slightly redesigned, and new features include an integrated terminal, an in-app file editor, and an expanded preview pane.
Google also expanded access to its Gemini app for desktop users on Windows. The app combines Google Search with AI capabilities, similar to the browser version.
Users can generate images, videos, and music, and access features like canvas and deep research. Google also introduced slash commands to Chrome, allowing users to save AI prompts as one-click tools for quick access.
Google’s New Text-to-Speech Model
A new text-to-speech model, Gemini 3.1 Flash TTS, was released by Google. Available through Vertex AI and Google AI Studio, this model allows for fine-tuned voice control, enabling specific tones like whispers, panic, or laughter. It can also handle dramatic pauses, making generated audio more dynamic.
Google also updated its image generation model, Nano Banana, to better use personal data from a user’s Google account. If connected to your email, calendar, and drive, Nano Banana can generate more personalized images. However, the effectiveness depends heavily on how much personal data is linked to the account.
Perplexity’s Personal Computer and New AI Models
Perplexity launched “Personal Computer,” a feature that brings its agentic capabilities to a user’s local machine. This allows the AI to work with local files, native applications, and the web to complete complex workflows. While the AI models still run on Perplexity’s servers, this new feature enables deeper integration with a user’s own device, especially useful for tasks requiring persistent local access.
New open-source AI models were also announced. MiniMax M2.7 claims state-of-the-art performance on coding benchmarks, though its license prohibits commercial use. Alibaba released the open-source Quinn 3.6 35B A3B model, which offers decent benchmark scores and could potentially run locally on powerful hardware.
Anthropic’s Claude Opus 4.7 and Specialized Models
Anthropic released Claude Opus 4.7, a new model that shows significant improvements in agentic coding tasks. While benchmarks show a substantial leap over previous versions and GPT 4.6, the biggest gains are for coders. The company also hinted at an even more powerful model, Mythos, which is not yet available to the public.
OpenAI introduced GPT Rosalind for life science research, a specialized reasoning model for biology and drug discovery. This model is available through trusted access only, highlighting the growing trend of AI being developed for specific scientific breakthroughs.
Why This Matters
The rapid evolution of AI tools like OpenAI’s Codec and Perplexity’s Personal Computer shows a clear trend towards AI agents that can actively operate within our digital environments. These advancements aim to automate complex tasks, streamline workflows, and reduce the need for manual intervention across various applications.
For professionals, especially coders and researchers, the new capabilities mean increased productivity and the potential to tackle more complex problems. The development of specialized models for fields like life sciences suggests AI’s growing role in scientific discovery and solving critical global challenges.
The week’s updates indicate a future where AI is more integrated into our daily computing, capable of understanding context, performing actions across multiple applications, and even contributing to scientific progress. These developments are enabling more sophisticated AI assistants that can handle increasingly complex and personalized tasks.
Google’s Gemini 3.1 Flash TTS model is available in Vertex AI and Google AI Studio. Perplexity’s Personal Computer feature can be experienced on a Mac Mini. Claude Opus 4.7 is now available.
Source: AI News: Huge Updates From Anthropic, OpenAI and Google (YouTube)
