Perplexity launches an AI 'computer,' Anthropic acquires a computer use startup, plus a browser rendering hack.
Analysis yesterday reveals AI-driven software development now costs less than minimum wage, a significant benchmark for engineering economics. Anthropic acquired Vercept to advance Claude's computer use, while Perplexity launched its own AI 'computer' product. For builders, one team showed how to build a video renderer by lying to the browser about time, and a new take argues for killing code reviews entirely.
Perplexity launched Perplexity Computer, an AI system that executes multi-step workflows for hours or months. It breaks down user outcomes into tasks, delegating them to specialized agents for web research, data processing, and API calls.
Read full story→Anthropic acquired Vercept, a company focused on AI software interaction, to expand Claude's ability to operate within live applications for complex tasks. This follows Claude Sonnet 4.6 scoring 72.5% on the OSWorld evaluation for computer use skills.
A new analysis compares Claude Opus 4.6 and GPT-5.3 Codex. Claude Opus 4.6 excels in reasoning depth and long-context analysis, scoring high on GPQA Diamond and MMLU Pro, with a 1-million-token context. GPT-5.3 Codex prioritizes agentic speed and coding throughput, offering 25% faster inference and superior terminal automation on Terminal-Bench 2.0.
Cursor launched cloud agents that control their own virtual machines, which can use dev environments and test code. These agents can iterate on software, build features, and test UIs autonomously. Over 30% of Cursor's merged pull requests are now agent-created, signaling a move towards self-driving codebases.
A new GitHub repository, 'Antigravity Awesome Skills,' collects over 978 agentic skills for AI coding assistants like Claude Code, Gemini CLI, and GitHub Copilot. The project uses a universal format, allowing skills to work across different AI tools. Builders can install it via npm or git clone, accessing curated bundles for roles and detailed usage guides, offering a library of pre-built actions that equip AI coding assistants with specific, repeatable tasks, reducing the need for manual prompt engineering for common development operations.
Replit built a web page video rendering engine by making browsers deterministic for capture. They injected a JavaScript shim that replaces browser timing APIs with a virtual clock, forcing frame rendering only when instructed.
Cloudflare used AI to rebuild Next.js from scratch in one week, creating 'vinext'. This new framework, built on Vite and deploying to Cloudflare Workers, delivers faster build times and smaller client bundles. The project, costing only $1,100 in tokens, illustrates how AI can drastically reduce the time and cost of replicating complex software, especially when paired with well-documented APIs and comprehensive test suites.
Geoffrey Huntley argues AI has driven software development costs below minimum wage, commoditizing traditional engineering skills. This shift redefines roles, allowing 'model-first' companies to operate leaner and faster.
The Pentagon is demanding "full and unfettered access" to Anthropic's Claude AI for all lawful purposes, escalating a conflict with the company. Anthropic resists, citing fears its AI could be used for autonomous weapons or mass surveillance, despite its safety policies. The Pentagon reportedly used Claude in a sensitive operation and now threatens actions like invoking the Defense Production Act if Anthropic doesn't comply. This dispute could set a precedent for government control over AI safety policies.
Traditional code reviews are failing under AI-generated code volume. The argument: shift human oversight to defining acceptance criteria and intent, letting AI agents handle code generation and verification. This approach uses layered AI trust, guardrails, and adversarial checks for faster iteration.
Apple released a Python SDK for its Foundation Models framework, giving developers access to on-device models powering Apple Intelligence on macOS. The SDK enables batch and on-device inference, real-time text streaming, and guided generation with structured, type-safe output. It requires macOS 26.0+ and Python 3.10+, and is currently in beta.