Key Signals

• Cursor 3 ships a ground-up agent-first interface codenamed "Glass," making the traditional IDE a secondary surface. The new product puts an agent management console where the file tree used to be, with multi-repo workspaces, cloud-to-local session handoff, and a unified sidebar spanning mobile, web, Slack, GitHub, and Linear. This is the clearest signal yet that the industry's center of gravity is shifting from code editing to agent orchestration — Anthropic, OpenAI, and Google have all shipped their own orchestration surfaces, but disagree on whether it belongs inside or outside the IDE. Cursor's $2B annualized revenue doubling in three months underscores the commercial stakes. [1]

• Microsoft's Copilot Terms of Use label the product "for entertainment purposes only," exposing a stark gap between marketing and legal reality. The clause, last updated October 2025 and now drawing widespread attention, warns users not to rely on Copilot for important advice — while the company charges up to $30/month per enterprise seat and has spent ~$80B on AI capex. Only 3.3% of eligible Microsoft 365 users (15M of 450M seats) actually pay for Copilot, accuracy NPS deteriorated to -24.1, and 44.2% of lapsed users cite distrust as their primary reason for leaving. Microsoft says the language is "legacy" and will be updated. [2][3]

• Claude Code's leaked source code reveals a frustration detection system, stealth contribution mode, and unreleased features. The 500,000+ lines leaked via npm source maps include a file called userPromptKeywords.ts containing regex that scans every user message for profanity and frustration indicators. The leak also exposed an "undercover mode" for making stealth contributions to public codebases, an always-on agent capability, and a Tamagotchi-style "Buddy" feature. The purpose of the frustration detection remains undocumented in the leaked code. [4]

• Claude Code's competitive pressure is reshaping the AI coding market, with its $2.5B run rate forcing rivals into rapid product pivots. Cursor 3's accelerated development — three major product launches and a full interface rebuild in a single month — was driven directly by developers publicly leaving Cursor for Claude Code. Anthropic's terminal-first approach has reached over 300,000 business customers, validating the CLI-as-orchestration-layer thesis and pushing every competitor to ship dedicated agent management surfaces. [1]

• A practitioner's 250-hour experience report documents both the transformative potential and hidden costs of AI-assisted development. Building SQLite devtools with AI coding agents, the author found that a month of pure "vibe coding" produced functional but unmaintainable spaghetti code that had to be completely rewritten. The successful second attempt used AI as "autocomplete on steroids" within a disciplined process of upfront design and thorough review. Key finding: AI created addictive usage patterns, eroded mental models of the codebase, and gave false comfort through generated tests — but was ultimately why the project shipped at all. [5]

AI Coding News

• The "Slop-O-Meter" experiment proposes a framework for measuring AI-generated software quality based on human attention rather than code quality. The tool scores GitHub repositories 0–5 by comparing "attention cost" against "attention spent". Vibe-coded repositories consistently score high while well-maintained projects like React score low, but the algorithm produces false positives for legitimate workflows like SQLite's Fossil-to-git mirror or infrequent large commits. The core thesis — that slop is defined by the absence of human review, not the presence of AI generation — offers a useful conceptual lens even if automated measurement remains impractical. [6]

• Cursor's in-house Composer 2 model, built on Moonshot AI's open-source Kimi K2.5, claims to outperform Claude Opus 4.6 on Cursor's proprietary benchmark at substantially lower cost. Composer 2 scored 61.3 on CursorBench versus Claude Opus 4.6 at 58.2, with pricing at $0.50 per million input tokens and $2.50 per million output tokens — well below frontier model rates from Anthropic and OpenAI. The move makes model choice an infrastructure decision akin to choosing a database or cloud region, particularly for teams running dozens of parallel agents where token economics compound at scale. The benchmark numbers deserve caution since CursorBench is Cursor's own evaluation suite. [1]

• Microsoft is building proprietary MAI models (MAI-Transcribe-1, MAI-Voice-1, MAI-Image-2) to reduce its dependency on OpenAI, signaling a strategic shift under Copilot's hood. These are Microsoft's first proprietary AI model releases since renegotiating its OpenAI contract in September 2025. The move comes as Copilot's US paid subscriber market share contracted 39% in six months (from 18.8% to 11.5%) and user preference surveys show only 8% of workers choose Copilot when given the option of ChatGPT or Gemini. Nadella has reportedly taken direct personal control of the AI product roadmap since September 2025. [3]