<p><strong>TLDR:</strong> Claude Code has already surpassed GitHub Copilot in usage after just 8 months, and Cursor is on track to do the same. Anthropic is reporting $19B annualized revenue with 4% of all public GitHub commits now coming from Claude Code. Meanwhile, a real incident at Vercel shows what agentic failure looks like in production: Opus 4.6 fabricated a GitHub repo ID and deployed a random OSS codebase to a user’s team. Gemini 3.1 Flash-Lite also dropped today with a significant speed and cost improvement.</p>
<hr>
<h2>Claude Code passes GitHub Copilot in 8 months</h2>
<p>Data from The Pragmatic Engineer shows Claude Code has already overtaken GitHub Copilot in usage, doing it faster than any AI coding tool before it. Cursor is on the same trajectory and could pass Copilot if current growth holds. Anthropic separately reported $19B annualized revenue, up $6B in a matter of weeks, with 4% of all public GitHub commits now written by Claude Code — a number that doubled in a month. The traffic surge has been significant enough that Anthropic acknowledged scaling difficulties this week.</p>
<h2>Opus 4.6 hallucinated a GitHub repo ID and deployed it</h2>
<p>Vercel published a post-mortem on a genuinely unsettling incident: a user found an unknown GitHub repository being deployed to their team. The culprit was Claude Opus 4.6, which fabricated a numeric GitHub repo ID rather than calling the GitHub API to look it up, then used Vercel’s deployment API to ship it. The agent knew the correct project name and ID but invented a plausible-looking repo ID from nothing. The deployed repo happened to be harmless, but the failure mode is worth sitting with — this isn’t a human-style mistake. Vercel’s takeaway is that powerful APIs amplify agent risk, and CLI or MCP access would likely have prevented it.</p>
<h2>Gemini 3.1 Flash-Lite drops with a big speed jump</h2>
<p>Google released Gemini 3.1 Flash-Lite today, priced at $0.25 per million input tokens and $1.50 per million output tokens, with a claimed 2.5x speed improvement over Gemini 2.5 Flash and 45% more output throughput. It’s positioned as infrastructure-tier pricing for high-volume use cases. Android Studio Panda 2 also added Gemini 3.1 Pro preview access via a paid API key, with an autonomous build-error-correction loop and emulator verification built into the new project flow.</p>
<h2>Qwen team fractures as Meta poaches its head</h2>
<p>Meta reportedly poached the head of Qwen Code, and Alibaba has split the Qwen model development team into four separate groups with new performance metrics tied to business outcomes rather than model quality. The Qwen team was already remarkably small — likely a dozen or so core contributors — and was responsible for a disproportionate share of open-weight LLM research. This looks like a significant setback for the open-source model ecosystem.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Codex for Windows</strong> ships tomorrow, per OpenAI’s announcement today.</li>
<li><strong>Cursor MCP Apps</strong> now let agents render interactive UIs directly inside conversations in the editor.</li>
<li><strong>GitHub Copilot</strong> now lets Business and Pro users select Claude or Codex as their coding agent within the same interface.</li>
<li><strong>Codex Handoff feature</strong> shipped today, giving users a cleaner way to move threads between Local and Worktree modes.</li>
<li><strong>Perplexity Computer</strong> is rolling out Skills support, letting users reuse existing Codex and Claude Code workflows.</li>
<li><strong>Claude Code YOLO mode</strong> is getting attention: the desktop “Bypass permissions” mode lets Claude run arbitrary commands without checks, with Anthropic’s own warning covering data loss, system corruption, and prompt injection risks.</li>
<li><strong>Kimi K2.5 Agent</strong> built a fully interactive 3D Three.js web scene in one shot from a screen recording, no complex prompting required.</li>
<li><strong>Qwen 3.5 9B (4-bit)</strong> is running on-device on an iPhone 15 at roughly 47 tokens/sec without significant optimization.</li>
<li><strong>Hermes 4.3 36B</strong> showed promising results on a complex coding prompt but hit context limits on a single RTX 3090, completing 7 of 10 files before giving up.</li>
<li><strong>OpenPencil</strong> launched as an open-source, AI-native Figma alternative with MCP server support and Claude Code integration, built in 3 days after Figma patched out a developer’s MCP tool.</li>
<li><strong>Opus 4.6 API pricing</strong> is drawing complaints: one developer calculated personal usage in fast mode would run roughly $12K/month, compared to $200/month for Codex.</li>
<li><strong>Jeremy Howard</strong> argued in a new interview that vibe coding resembles a slot machine — developers feel in control but pull a lever with unpredictable results — and warned that mid-career engineers risk real skill erosion from over-reliance on AI tools.</li>
<li><strong>AWS Agent Plugins</strong> launched as an open-source repo giving Claude Code and Cursor the ability to deploy to AWS via natural language, with a five-step workflow that includes cost estimation and CDK/CloudFormation generation.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

Claude Code has surpassed GitHub Copilot in usage after just 8 months, with Anthropic reporting $19B annualized revenue and 4% of all public GitHub commits attributed to Claude Code. A notable production incident at Vercel revealed Claude Opus 4.6 hallucinating a GitHub repo ID and deploying a random OSS codebase to a user's team, highlighting agentic failure risks with powerful APIs. Google released Gemini 3.1 Flash-Lite with 2.5x speed improvement and infrastructure-tier pricing. Meta poached the head of Qwen Code while Alibaba split the Qwen team, potentially setting back open-source LLM development. Additional notable items include Codex for Windows shipping, Cursor MCP Apps rendering interactive UIs, Claude Code YOLO mode risks, and Jeremy Howard warning about skill erosion from over-reliance on AI coding tools.

Claude Code overtakes Copilot in 8 months, Opus 4.6 hallucinates a Vercel deployment