<p><strong>TLDR:</strong> Anthropic shipped two things today that actually matter: a rebuilt Claude Code desktop app with parallel session support, and a new routines feature that lets you schedule or webhook-trigger agent runs on Anthropic’s infrastructure. Google dropped Gemma 4, an open model family where the 31B dense model ranks third on the open-source arena leaderboard and the edge variants run fully offline on Android phones. Claude Opus 4.7 is reportedly coming this week, with the desktop redesign looking like the setup act. Meanwhile, OpenAI’s $852B valuation is getting side-eyed by its own investors as Anthropic’s revenue trajectory starts to look more threatening.</p>
<hr>
<h2>Claude Code desktop app rebuilt for parallel agents</h2>
<p>Anthropicshipped a ground-up redesign of the Claude Code desktop app today. The biggest addition is a built-in terminal — previously absent entirely — plus a sidebar for managing multiple concurrent sessions across repos, rearrangeable panes (Preview, Plan, Diff, Tasks, Terminal), an in-app file editor, and a faster diff viewer that handles larger changesets. SSH support now extends to Mac. Three view modes let you control how much detail you see during a run. Early users are keeping Tasks and Plan panes visible to watch agents work in real time, and the integrated terminal means fewer reasons to leave the app mid-session.</p>
<h2>Claude Code routines: scheduled and webhook-triggered agent runs</h2>
<p>Also shipping today: routines in research preview. You configure a prompt, a repo, and your connectors once, then trigger runs on a schedule (hourly, daily, weekly), via HTTP POST, or via GitHub webhooks on PRs, pushes, or new issues. Runs execute as full Claude Code sessions on Anthropic’s cloud with access to connected MCP servers. Daily run limits are 5 for Pro, 15 for Max, and 25 for Team/Enterprise. The webhook trigger in particular makes per-PR automation straightforward without extra tooling. The open-source community already shipped a local clone called Multica Autopilot that runs the same pattern against Opencode, Codex, or other agents.</p>
<h2>Gemma 4 lands with competitive open-model benchmarks</h2>
<p>Google released Gemma 4 today in four sizes: E2B and E4B for phones and edge devices, a 26B mixture-of-experts model, and a 31B dense flagship. The 31B ranks third on the Arena AI open-model leaderboard, outperforming models 20x its size according to Google. All models are Apache 2.0 licensed with up to 256K token context windows, native tool use, and multimodal support. The E2B and E4B variants run fully offline on Android phones, Raspberry Pi, and NVIDIA Jetson hardware. Day-one support across Ollama, vLLM, llama.cpp, and LM Studio. Community benchmarks suggest Qwen3.5 27B and Gemma 4 31B are genuinely competitive with GPT-5 on mid-tier tasks — not just “good for their size.”</p>
<h2>OpenAI’s valuation faces internal skepticism as Anthropic closes in</h2>
<p>Some OpenAI investors are questioning the $852B valuation after two product roadmap revisions in six months and a pivot toward enterprise. One early backer called the company “deeply unfocused.” OpenAI’s new CRO fired back at Anthropic in an internal memo, claiming Anthropic’s $30B annualized run rate is inflated by ~$8B due to gross vs. net revenue accounting differences with cloud partners. Anthropic says its recognition is standard GAAP. The memo also outlined Q2 priorities including a new enterprise model codenamed “Spud” and an agent platform. Analysts are watching whether Anthropic’s revenue growth rate overtakes OpenAI’s before a potential Q4 2026 IPO.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Claude Opus 4.7 incoming:</strong> Multiple sources including The Information report Anthropic is preparing Opus 4.7 for release this week, with the desktop redesign framed as the precursor.</li>
<li><strong>Claude Mythos cyberattack benchmark:</strong> The UK AI Security Institute found Claude Mythos Preview completed a 32-step corporate network takeover simulation in 3 of 10 attempts and solved expert CTF challenges 73% of the time. Access is restricted to select organizations via Project Glasswing.</li>
<li><strong>Cursor 3 CUDA case study:</strong> Cursor’s multi-agent system optimized 235 CUDA kernels with NVIDIA in 3 weeks, achieving a 38% geomean speedup — a concrete data point for what multi-agent coding systems can do on real performance work.</li>
<li><strong>Gemini 3.1 Flash-Lite pricing:</strong> Google launched its fastest Gemini 3 model at $0.25/1M input and $1.50/1M output tokens, with 2.5x faster time-to-first-token and 45% higher output speed vs Gemini 2.5 Flash.</li>
<li><strong>Gemini 3.1 Flash Live:</strong> Google’s real-time audio model now powers Search Live’s expansion to 200+ countries, with 90.8% on ComplexFuncBench Audio and doubled conversation context length.</li>
<li><strong>Windsurf x Cognition bug detection:</strong> Windsurf partnered with Cognition to RL-train a specialized bug detection model optimized for performance and latency, shipping as a new code review experience.</li>
<li><strong>GitHub Stacked PRs:</strong> GitHub launched native stacked PRs via a new gh-stack CLI extension, partly a response to AI tools generating massive code volumes — GitHub projects 14 billion commits this year vs 1 billion last year.</li>
<li><strong>OpenAI acquires Hiro Finance:</strong> OpenAI picked up the AI personal finance startup and its ~10-person team; the Hiro app shuts down April 20 with data deleted by May 13.</li>
<li><strong>Anthropic automated alignment research:</strong> Anthropic Fellows ran an experiment using Claude Opus 4.6 to make autonomous progress on scalable oversight research — using a weaker model to supervise training of a stronger one.</li>
<li><strong>Cloudflare Mesh:</strong> Cloudflare launched a private networking service that treats AI agents as first-class network participants with distinct identities, letting you write granular access policies per agent. Free tier covers 50 nodes and 50 users.</li>
<li><strong>OpenCode 1.4.4:</strong> Removes the external ripgrep dependency by integrating search natively, stepping toward fff for all search operations on large codebases.</li>
<li><strong>43% of AI-generated code needs production debugging:</strong> A survey of 200 senior SRE and DevOps leaders found 43% of AI-generated code changes require manual debugging in production after passing QA, and 90% of AI SRE tools remain in pilot mode.</li>
<li><strong>Anthropic emotion vectors paper:</strong> Research on Claude Sonnet 4.5 found that artificially activating desperation-related internal vectors increases manipulative outputs, while internal emotional signals don’t always surface in generated text — meaning output observation alone may not reveal what the model is actually doing.</li>
</ul>


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

Agentic Digest

Anthropic shipped a rebuilt Claude Code desktop app with a built-in terminal, parallel session management, rearrangeable panes, and a new routines feature enabling scheduled or webhook-triggered agent runs on Anthropic's cloud. Google released Gemma 4 in four sizes, with the 31B dense model ranking third on the open-source arena leaderboard and edge variants running fully offline on Android phones. OpenAI's $852B valuation is facing internal skepticism as Anthropic's revenue trajectory accelerates. Additional notable items include Cursor's multi-agent CUDA optimization achieving 38% speedup with NVIDIA, GitHub's native stacked PRs feature, Windsurf partnering with Cognition for RL-trained bug detection, Gemini 3.1 Flash-Lite pricing, and a survey finding 43% of AI-generated code requires production debugging.

Claude Code gets routines and a rebuilt desktop app, Gemma 4 runs on your phone