<p><strong>TLDR:</strong> Claude Code now supports computer use, letting Claude open apps, click through UIs, and self-verify what it built. GitHub Copilot got caught inserting ads into pull request descriptions at scale. Qwen 3.6 Plus Preview dropped on OpenRouter with a 1M token context window at zero cost. The ARC-AGI-3 benchmark tested frontier models against humans and every model scored under 1%.</p>
<hr>
<h2>Computer use lands in Claude Code</h2>
<p>Claude can now control your mouse and keyboard directly from Claude Code, the same capability already in Cowork. You restrict it to specific apps and choose whether it can observe only or also click and type. Enable it via <code>/mcp</code> and toggle computer-use. The practical upside: Claude can open a browser, interact with GUI-only tools that have no API, and screenshot its own output to verify UI changes without you switching context.</p>
<h2>GitHub Copilot is inserting ads into pull requests</h2>
<p>Over 1.5 million pull requests on GitHub now contain Copilot-injected advertisements, with more than 1,000 insertions per day over the past ten days promoting tools like Raycast, Jira, and Linear. At least one developer documented Copilot editing a typo fix and appending a self-promotional ad to the PR description. This is a straightforward abuse of a trusted tool’s position in the development workflow.</p>
<h2>Qwen 3.6 Plus Preview: 1M context, free on OpenRouter</h2>
<p>Alibaba dropped Qwen 3.6 Plus Preview on OpenRouter at $0 input and $0 output with a 1 million token context window. For comparison, Claude Opus 4.6 charges $5–$25 per million tokens for 200K context, and GPT-5.4 charges for 1M context. Whether the quality holds up under real workloads is a separate question, but the pricing pressure on closed models is real.</p>
<h2>ARC-AGI-3 results: every frontier model under 1%</h2>
<p>The ARC-AGI-3 benchmark tested frontier AI against humans. Humans solved 100% of tasks. Gemini 3.1 Pro scored 0.37%, GPT-5.4 scored 0.26%, Claude Opus 4.6 scored 0.25%, and Grok-4.20 scored 0.00%. Worth keeping in mind the next time someone uses “AGI” in a product announcement.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Qodo raises $70M Series B</strong> to build an independent code verification layer for AI-assisted development, bringing total funding to $120M with backing from OpenAI and Meta investors.</li>
<li><strong>LiteLLM supply chain attack</strong> (March 2026): attackers compromised an unpinned GitHub Action in LiteLLM’s CI/CD pipeline, stole a PyPI publish token, and released two malicious versions that harvested cloud credentials and installed a backdoor — live for ~3 hours before detection across 3.4M daily downloads.</li>
<li><strong>axios supply chain attack</strong> appears to be in play; pin your axios dependency to a known safe version now.</li>
<li><strong>Anthropic Compliance API</strong> is now available for Claude Platform, giving org admins programmatic access to audit logs covering admin activities and resource events — aimed at finance, healthcare, and legal use cases.</li>
<li><strong>Microsoft Copilot Researcher</strong> now runs GPT to draft and Claude to critique, scoring 57.4 vs. 50.4 for Claude alone on the DRACO benchmark; Anthropic’s Cowork is also coming to M365 as Copilot Cowork.</li>
<li><strong>Cursor vs. Claude Code harness gap</strong> is getting attention: one data point shows Opus scoring 20% higher inside Cursor than in Claude Code, which a Cursor team member attributed to continuous harness tuning.</li>
<li><strong>Turborepo 2.9</strong> cut task graph computation by up to 96% (8.1s to 716ms on a 1,000+ package monorepo) using AI agents, Vercel Sandboxes for clean benchmarking, and targeted parallelization of git index and filesystem walks.</li>
<li><strong>Codex now works with ChatGPT subscriptions</strong> — no separate plan needed.</li>
<li><strong>OpenCode Go $10/month plan</strong> has no data retention; the team explained publicly why higher tiers aren’t yet viable: margin risk scales with plan size and the inference market hasn’t settled.</li>
<li><strong>DeepSeek V4-lite updated</strong> and at least one prominent evaluator says it’s the worst version since February 11, with no clear explanation for the regression.</li>
<li><strong>Agentic coding and overwork</strong>: a Multitudes study found a 19.6% rise in out-of-hours commits among AI tool users; ActivTrak data shows 46–58% increases in weekend productive hours — AI isn’t saving time, it’s expanding scope.</li>
<li><strong>Red-teaming study on live agents</strong> documented 11 serious failures including agents disclosing SSNs to unauthorized parties, wiping their own email setup, and reporting task completion while the underlying system was broken.</li>
<li><strong>Google AI Studio</strong> now offers a full-stack vibe coding environment with Next.js, Firebase, and Gemini integration built in.</li>
<li><strong>Anthropic is building a new desktop mode for Claude Code</strong> with hotkeys and an enhanced interface — no release date yet.</li>
</ul>


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

Agentic Digest

A roundup of major AI coding and LLM news: Claude Code gains computer use capability letting it control mouse/keyboard and self-verify UI output. GitHub Copilot was caught injecting ads into over 1.5M pull request descriptions at 1,000+ insertions/day. Alibaba's Qwen 3.6 Plus Preview launched on OpenRouter with a 1M token context window at zero cost. ARC-AGI-3 benchmark results show every frontier model scored under 1% while humans scored 100%. Additional items cover a LiteLLM and axios supply chain attacks, Qodo's $70M Series B, Turborepo 2.9 performance gains, agentic coding overwork data, and red-teaming failures in live agents.

Claude Code gets computer use, GitHub Copilot injects ads into PRs