<p><strong>TLDR:</strong> Anthropic shipped 14+ launches in March including Claude Opus 4.6, Sonnet 4.6, and Claude Code on web and mobile, but also hit five infrastructure outages and accidentally leaked a next-gen model called Claude Mythos. OpenAI’s Codex grew faster than anyone predicted, nearly exhausting capacity three days running. A benchmark comparing Claude Code across 13 languages found Ruby, Python, and JavaScript are significantly faster and cheaper than statically typed alternatives. And a vibe-coded ext4 submission to OpenBSD kicked off a messy debate about copyright on LLM-generated code.</p>
<hr>
<h2>Anthropic’s March: 14 launches, 5 outages, one accidental leak</h2>
<p>The full March recap is out and it’s a lot. Claude Opus 4.6 and Sonnet 4.6 shipped with 1M-token context, computer use preview landed, Claude Code expanded to web and mobile, and the Model Context Protocol hit 97 million monthly SDK downloads — up 4,750% since its November 2025 launch, now governed by the Linux Foundation. The cost of moving fast: at least five infrastructure outages in a single month, including a major one this weekend that took down <a href="http://Claude.ai" target="_blank" rel="noopener nofollow">Claude.ai</a>, the API, and Claude Code simultaneously. The accidental leak of Claude Mythos (codename Capybara) is the detail worth watching — Anthropic describes it as a step change in capability with notably higher benchmark scores in cybersecurity.</p>
<h2>Codex capacity nearly collapses under demand</h2>
<p>OpenAI posted that Codex growth has outpaced their own internal predictions, nearly running out of capacity three days in a row. They’ve scaled up ahead of next week. That’s a real signal — not a marketing number. The Claude Code vs. Codex comparison is becoming a daily conversation in the community, with at least one developer running the same E2E feature planning prompt through both and having each evaluate the other’s output.</p>
<h2>Claude Code language benchmark: Ruby and Python win, TypeScript loses</h2>
<p>A quantitative benchmark tested Claude Code across 13 languages using a mini-Git implementation task, 20 runs each. Ruby, Python, and JavaScript came out fastest (73–81 seconds), cheapest ($0.36–0.39), and most stable. Statically typed languages like C, Haskell, and TypeScript were 1.4–2.6x slower and more expensive. Adding type checkers like mypy or Steep to dynamic languages also significantly increased cost and time. The hypothesis: AI training data volume, language conciseness, and type system overhead all contribute. Worth factoring into your stack choices if you’re running agents at scale.</p>
<h2>Vibe-coded ext4 rejected by OpenBSD over copyright</h2>
<p>A developer submitted an LLM-generated ext4 filesystem implementation to OpenBSD. Maintainers including Theo de Raadt rejected it: no human author can clearly claim copyright on machine-generated code, and the LLM was trained on GPL-licensed Linux code, raising license contamination concerns. The submitter refused to retract copyright claims. This is the clearest real-world collision yet between LLM-generated contributions and open source legal norms — and it won’t be the last.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>Claude Mythos leak:</strong> Described internally as a step change in capability, particularly in cybersecurity benchmarks — no release date mentioned in the recap.</li>
<li><strong>Codex Security:</strong> Still free during preview; find, validate, and fix vulnerabilities in a single workflow.</li>
<li><strong>Startups spending $1,000+/day on Claude Code and Codex:</strong> The concern is real — AI-assisted dev is becoming pay-to-win, and prices will likely go up.</li>
<li><strong>Local model case:</strong> One developer runs a 24/7 software factory on a Mac Studio using local models — scraping, problem identification, and app building — at electricity cost only; argues Qwen 3.5 is already Sonnet 4.5-level and closing fast.</li>
<li><strong>Windsurf getting attention:</strong> A few developers pushing back on the Claude/Codex duopoly narrative, citing Windsurf’s pair programming tab completion as a useful middle ground between manual and fully agentic.</li>
<li><strong>Vercel json-render:</strong> Open-sourced in January 2026, already at 13,000 GitHub stars and 200 releases — generates structured UIs directly from natural language prompts.</li>
<li><strong>Knuth’s Hamiltonian problem resolved:</strong> Claude found the odd-m construction, GPT solved the even case, a multi-agent workflow found simpler constructions for both, and the proof was formalized in Lean — a full human+AI+proof assistant pipeline on an open math problem.</li>
<li><strong>LLM sycophancy reminder:</strong> A developer spent 4 hours refining a blog argument with an LLM, then asked it to argue the opposite — it immediately demolished the original. Always probe both directions.</li>
<li><strong>Hermes Agent update:</strong> Teknium pushed an update aimed at making GPT/Codex models less lazy in agentic tasks.</li>
<li><strong>AMD GAIA 0.17:</strong> Introduces Agent UI for running AI agents locally on Ryzen AI hardware, with RAG, shell execution guardrails, and real-time reasoning visualization.</li>
</ul>


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

Agentic Digest

Anthropic's March recap covers 14+ launches including Claude Opus 4.6, Sonnet 4.6, Claude Code on web/mobile, and MCP hitting 97M monthly SDK downloads — alongside five infrastructure outages and an accidental leak of a next-gen model called Claude Mythos. OpenAI's Codex nearly ran out of capacity three days running due to faster-than-predicted growth. A 13-language benchmark found Ruby, Python, and JavaScript are 1.4–2.6x faster and cheaper than statically typed languages when running Claude Code at scale. An LLM-generated ext4 filesystem submission was rejected by OpenBSD over copyright and GPL license contamination concerns, marking a significant collision between AI-generated code and open source legal norms.

Anthropic's March recap drops, Codex nearly runs out of capacity