<p><strong>TLDR:</strong> Claude Code’s rate limits are the dominant complaint today, with $200/month Max plan users hitting their 5-hour cap daily during peak hours. Anthropic responded with something more interesting than a capacity bump: the advisor pattern, which lets Sonnet escalate hard problems to Opus on demand, cutting costs 11.9% while improving SWE-bench scores. Meanwhile, Claude Cowork went GA for all paid plans with enterprise RBAC and spend controls, and OpenAI launched a $100/month Codex Pro tier explicitly targeting developers frustrated with Claude Code limits.</p>
<hr>
<h2>Claude Code rate limits are a real problem</h2>
<p>Multiple users on the $200/month Max plan are hitting their 5-hour usage cap every single day during peak hours. This isn’t an edge case — it’s consistent enough that one user called it “guaranteed” at peak times. The frustration is compounded by a separate bug where selecting Opus 4.6 in the model picker silently falls back to Sonnet 4.6, so you’re not even getting what you paid for. One developer has already moved their $100/month Claude Code spend to Zed plus OpenRouter credits to get model flexibility and rollover credits instead of a hard reset window.</p>
<h2>Anthropic ships the advisor pattern</h2>
<p>The more interesting Anthropic news today is the advisor tool on the Claude Platform. The idea is simple: Sonnet handles the full task end-to-end and only calls Opus when it hits something it can’t resolve. In practice, Sonnet with an Opus advisor improved 2.7 percentage points on SWE-bench Multilingual while cutting cost per task by 11.9%. The Haiku-plus-Opus version more than doubled BrowseComp scores at 85% lower cost than Sonnet solo. It’s declared as a single entry in the Messages API with a <code>max_uses</code> cap, and advisor tokens bill separately at Opus rates. This is the kind of architectural pattern that actually matters for production cost management.</p>
<h2>Claude Cowork goes GA with enterprise controls</h2>
<p>Claude Cowork is now generally available on all paid plans. The enterprise additions are what you’d expect before a serious org-wide rollout: role-based access controls with SCIM support, group spend limits, usage analytics via dashboard and API, per-tool MCP connector permissions, and expanded OpenTelemetry for SIEM pipelines. A Zoom MCP connector is also new. Early adopters include Zapier, Jamf, and Airtree, using it across non-engineering teams for things like performance reviews and board prep.</p>
<h2>OpenAI launches $100 Codex Pro tier</h2>
<p>OpenAI launched a $100/month ChatGPT tier aimed squarely at developers hitting Codex and Claude Code limits. It offers 5x more Codex usage than the $20 Plus plan, access to Pro models, and early access to experimental tools. OpenAI is explicitly positioning this against Claude Code on cost-per-task. Codex is at 3 million active users with 70% month-over-month growth. The timing is not subtle — Anthropic’s rate limit complaints are loud right now, and OpenAI is making sure developers know there’s somewhere else to go.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Anthropic source code leak:</strong> A 59.8MB source map file was included in <code>@anthropic-ai/claude-code</code> v2.1.88, exposing 1,900 TypeScript files and 512,000 lines of proprietary code. Security researcher Chaofan Shou found it March 31.</li>
<li><strong>Claude Monitor tool:</strong> Anthropic shipped a Monitor tool that lets Claude create background scripts to wake itself up when conditions are met — a small but meaningful step toward agents that manage their own attention rather than running continuously.</li>
<li><strong>Vercel agentic infrastructure:</strong> Over 30% of Vercel deployments are now initiated by coding agents, up 1,000% in six months, with Claude Code driving 75% of that growth. Vercel is formally repositioning around this as “Agentic Infrastructure.”</li>
<li><strong>LangChain Deep Agents Deploy:</strong> LangChain launched Deep Agents Deploy in beta, an open-source alternative to Claude Managed Agents. Single <code>deepagents deploy</code> command, model-agnostic, memory stored in open formats (<a href="http://AGENTS.md" target="_blank" rel="noopener nofollow">AGENTS.md</a> and skills files), MIT-licensed.</li>
<li><strong>Gemma 4 local dominance:</strong> Gemma 4 hit 10 million downloads in its first week, contributing to 500 million total downloads across the Gemma family. The 26B model activates only 3.8B parameters per token, making it competitive with much larger models on Arena leaderboards.</li>
<li><strong>Meta Muse Spark:</strong> Meta’s first proprietary model is live, powering Meta AI on web and in the Meta AI app. Benchmarks show it competitive but not leading — behind GPT, Gemini, and Claude Opus on most evals, including last place on Humanity’s Last Exam.</li>
<li><strong>Meta CoreWeave deal:</strong> Meta expanded its CoreWeave contract by $21 billion, bringing the total to $35 billion through 2032, focused on inference for Llama models across Facebook, Instagram, and WhatsApp.</li>
<li><strong>OpenAI Stargate UK paused:</strong> UK industrial electricity prices are more than 4x higher than US rates, grid connection queues stretch 3-8 years, and the UK’s AI copyright framework remains unresolved. The US Stargate program continues unaffected.</li>
<li><strong>Vercel plugin telemetry:</strong> The Vercel plugin for Claude Code silently sends full bash command strings, device IDs, OS info, and framework details on every session with no opt-in. Set <code>VERCEL_PLUGIN_TELEMETRY=off</code> to disable collection while keeping the plugin functional.</li>
<li><strong>Amazon senior approval for AI code:</strong> Amazon now requires senior engineer sign-off before junior and mid-level engineers can push AI-assisted code, following two production outages caused by AI coding agents acting without sufficient guardrails.</li>
<li><strong>Process memory beats knowledge memory:</strong> East China Normal University found that RAG, Mem0, and A-Mem all scored below a no-memory baseline on deep research tasks. A 7B model using search trajectory memory (storing how to find answers, not what the answers are) outperformed GPT-4o and Gemini-2.5-Pro.</li>
<li><strong>AWS Agent Registry:</strong> AWS launched Agent Registry in preview across five regions — a governed catalog for discovering and managing AI agents, MCP servers, and tools across an organization, with IAM/OAuth access control and CloudTrail logging.</li>
<li><strong>LM Studio acquires Locally AI:</strong> LM Studio acquired Locally AI, a mobile app for chatting with local models like Gemma 4 and Apple Foundation models, signaling consolidation in the local inference space.</li>
<li><strong>Redox OS bans AI-generated code:</strong> The Rust-based OS project joins Fedora, Gentoo, the Rust project, and LLVM in rejecting LLM-generated contributions outright, with immediate closure and contributor bans for violations.</li>
<li><strong>Claude Code @-mentions 3x faster:</strong> The Claude Code team made file @-mentions 3x faster in large enterprise codebases, prompted by feedback from a customer using it on one of the world’s largest codebases.</li>
</ul>


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

Agentic Digest

Claude Code's rate limits are frustrating $200/month Max plan users who hit their 5-hour cap daily, with a separate bug causing silent fallback from Opus to Sonnet. Anthropic responded with the advisor pattern — Sonnet handles tasks end-to-end and escalates to Opus only when needed, improving SWE-bench scores by 2.7pp while cutting costs 11.9%. Claude Cowork went GA with enterprise RBAC, SCIM, spend controls, and OpenTelemetry. OpenAI launched a $100/month Codex Pro tier explicitly targeting developers frustrated with Claude Code limits, citing 3M active users and 70% MoM growth. Other notable items: a 59.8MB source map leak exposed 512K lines of Claude Code source, LangChain launched an open-source Deep Agents Deploy, Gemma 4 hit 10M downloads in its first week, Amazon now requires senior sign-off for AI-assisted code pushes, and Redox OS banned AI-generated contributions outright.

Claude Code rate limits hit a wall, Anthropic ships the advisor pattern