Right. You've been away a full week and rather a lot has happened. Brace yourself.
Anthropic unveiled Claude Mythos Preview via Project Glasswing — a model so adept at finding zero-day vulnerabilities that they're refusing to release it publicly. It identified thousands of previously unknown flaws across major operating systems and browsers, including a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw. Instead of a general launch, Anthropic is restricting access to critical infrastructure partners and open-source security teams, backed by $100M in usage credits. It's also in gated preview on Amazon Bedrock. A flagship model voluntarily withheld on safety grounds — not the usual press release trajectory.
Meta released Muse Spark, the first model from Superintelligence Labs. Built from scratch — not a Llama derivative. It ranks 4th on the Artificial Analysis Intelligence Index behind GPT-5.4, Gemini 3.1 Pro, and Opus 4.6, with strong medical reasoning but weak ARC-AGI 2 scores. Meta claims comparable capability to Llama 4 Maverick at an order of magnitude less compute.
Google released Gemma 4 — four models from 2B to 31B, natively multimodal with 256K context. The edge variants run on a Raspberry Pi. The 31B flagship hits 89.2% on AIME 2026 and 80% on LiveCodeBench. If you do anything with local models, this is worth your attention.
Cursor 3 rebuilt the interface around managing agents rather than writing code. The new Agents Window runs parallel agents across local machines, worktrees, SSH, and cloud setups. Design Mode lets frontend developers click on rendered UI to direct agents visually. Meanwhile, Bugbot learned to improve its own review rules by studying human feedback on its PRs.
Anthropic shipped v2.1.94 through v2.1.101 between April 7-10. Features include a Vertex AI setup wizard, a Monitor tool for streaming background events, and PID namespace sandboxing. More urgently: several Bash permission bypass vulnerabilities were patched — backslash-escaped flags, compound commands, env-var prefixes, and /dev/tcp redirects all skirting permission prompts. Update if you're running permissive mode.
Copilot Autopilot entered public preview. The agent now approves its own actions, auto-retries errors, and runs to task completion without intervention. Also: nested subagents, browser debugging, and MCP server support. The IDE assistants are converging on "just do the whole thing."
Anthropic blocked Claude subscriptions from routing through OpenClaw and third-party agent harnesses, citing token arbitrage. Community reaction was volcanic. Then they temporarily suspended OpenClaw's creator's account, which went viral before being reversed hours later. The tension between AI labs and the ecosystems building on their infrastructure isn't resolving anytime soon.
Harvard Gazette published "Vibe Coding May Offer Insight Into Our AI Future", alongside pieces from Bloomberg and Slate. The Harvard piece is the standout — education professor Karen Brennan takes it seriously while noting who benefits and what gets lost. Vibe coding has graduated from developer Twitter to mainstream cultural conversation.
Simon Willison pointed out that ChatGPT's voice mode runs on a GPT-4o-era model with an April 2024 knowledge cutoff. The most natural-feeling interface is backed by one of the oldest models. Most users have no idea.
Suno v5.5 added voice cloning with identity verification, custom model fine-tuning on your own catalog, and taste profiling. The shift from generic generation toward identity-driven music creation continues.
That's the lot. Ten days' worth of the field rearranging itself. Do try to keep up.