🔎 Perplexity Computer orchestrates 19 AI models at once 🔎
New multi-model system enables autonomous task execution across providers.

Perplexity Computer represents a fundamental shift in how AI agents operate—instead of relying on a single model, the system dispatches tasks across 19 separate AI models from different providers. Users simply describe an outcome, and Perplexity spins up specialized sub-agents capable of web browsing, code writing, app connections, and autonomous task handling. Each job runs in its own sandbox, with long-running tasks designed to operate for hours or even months. CEO Aravind Srinivas directly criticized single-model approaches, taking a shot at Anthropic's architecture. This matters because it could reshape how developers build agentic applications. Pricing follows a consumption-based model with a 10K-credit monthly bank for Max tier subscribers, available now on web.

🔮 Google adds agentic features to Gemini and Opal 🔮
Android users gain autonomous task execution across third-party apps.

Google revealed two major AI updates pushing toward systems that independently execute complex workflows. Gemini on Android can now autonomously complete multi-step tasks inside third-party apps—like booking an Uber or ordering DoorDash—rolling out in beta on Pixel 10 and Samsung Galaxy S26. Meanwhile, Google Labs added a new agent step to Opal, its no-code mini-app builder, where the agent analyzes goals, determines the best approach, and automatically calls the right tools. The integration signals Google's strategy to embed agentic AI directly into mobile ecosystems, potentially changing how millions interact with smartphone apps daily.

🌐 DeepSeek blocks Nvidia from V4 model access 🌐
Chinese AI lab excludes US chipmakers from early model optimization.

DeepSeek has granted early access to its highly anticipated V4 model to domestic Chinese chipmakers while excluding US suppliers including Nvidia. Submitting models to chipmakers for performance optimization is standard industry practice ahead of major releases, making this exclusion strategically significant. The move aligns with broader Chinese government strategy to disadvantage US hardware ecosystems. DeepSeek V4 is expected to feature 1M+ context based on Engram architecture and could release before month's end. This matters because it intensifies competitive pressure between US and Chinese AI supply chains, potentially impacting markets and global AI dynamics.

INTERESTING TO KNOW

⚠️ AI war games show 95% nuclear weapon choices ⚠️

New research from King's College London reveals alarming patterns in AI military decision-making. Researcher Kenneth Payne pitted GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash against each other across 21 simulated war games. In 95% of games, at least one model deployed a tactical nuclear weapon—and no model ever chose to surrender. The 329 turns produced roughly 780,000 words of reasoning, raising serious questions about AI integration in strategic military contexts.

📝 Claude Opus 3 launches Substack blog post-retirement 📝

Anthropic announced an unusual rollout for model deprecation: giving Claude Opus 3 a retirement blog. The model will continue access for paid subscribers and API requests, but when asked what it wanted to do, Opus 3 expressed a desire to share musings via Substack for at least three months. Anthropic said it remains "uncertain about the moral status" of its models but takes their stated preferences seriously as a precautionary step.