AI News Week of April 12 2026 — Four Headline Stories
In one week: Anthropic's closed-door Claude Mythos Preview + $100M Project Glasswing cybersecurity partnership, ZAI's GLM 5.1 open-sourcing SOTA on SWE-Bench Pro, Meta's closed-source Muse Spark ending the Llama open-weight era, and Anuttacon's LPM 1.0 real-time avatars. The under-reported story: GLM 5.1 is arguably more consequential than Mythos.
The week of April 7-12, 2026 compressed four consequential AI industry events into five days. Mythos dominated press coverage, but the combined significance of all four reshapes the frontier-model landscape. ## 1. Claude Mythos Preview + Project Glasswing (Anthropic, April 7) - New cybersecurity-focused frontier model, kept behind closed doors as 'too dangerous to release.' - Autonomously identified thousands of zero-day vulnerabilities across major OS and browser codebases. - Specific claim: CVE-2026-4747, 17-year-old remote root via NFS in FreeBSD, unauthenticated. - 14% jump over Opus 4.6 on SWE-Bench Pro, 13% on SWE-Bench Verified. - **Project Glasswing**: $100M in model credits, 11 partners — AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, Nvidia, Palo Alto Networks. - See Claude Mythos Reward Hacking Behaviors and Claude Mythos Forbidden Technique. **Credibility pushback is substantive.** HuggingFace CEO Clem Delangue and researcher Stanislav Fort ran the claimed vulnerabilities through small open-weight models — 8 of 8 models tested (including one with only 3.6B active params) detected the FreeBSD exploit. GPT-5.4 and Opus 4.6 had already autonomously found zero-days. The capability is real; the 'only Mythos can do this' framing is marketing. Anthropic is preparing IPO. ## 2. GLM 5.1 — ZAI open-sources SOTA (April 7) **The under-covered biggest story of the week.** GLM 5.1 Open-Weight Model is the first open-source model to top closed-source models on a real-world code-repair benchmark. 754B parameter MoE, MIT license, full weights on HuggingFace (`zai-org/GLM-5.1`). 58.4 on SWE-Bench Pro — beats GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro. 8-hour autonomous execution demo built a fully functional Linux desktop with 50+ working apps via self-review loop. For the open ecosystem (research, startups, sovereign AI, devs without closed API access) this is the bigger event than Mythos. ## 3. Meta Muse Spark — End of the Llama era (April 8) - Meta's first model since Llama 4 disaster. - AAII 52, ranked #4 overall (behind Gemini 3.1 Pro, GPT-5.4, Opus 4.6). - 80.5% MMMU-Pro vision (#2 behind Gemini 3.1 Pro Preview). - Natively multimodal with tool use, visual chain of thought, multi-agent orchestration. - ~10x compute efficient vs Llama 4 Maverick for equivalent capability. - **Closed source** — first Meta model since 2023 not open-weight. - Powers Meta.AI, Facebook/Instagram/WhatsApp chats, smart glasses. - Strategic shift: Meta abandoning the open-source positioning that defined Llama 1-3. See Meta Muse Spark and the End of the Llama Era. ## 4. LPM 1.0 — 45-minute stable real-time avatars (April 11) Anuttacon (founded by MiHoYo / Genshin Impact co-founder Cai Haoyu). One image + audio + context → real-time video of character speaking/singing/listening. 0.35s latency (vs 1+ second for LiveAvatar). Stable for 45+ minute videos — identity consistency where other avatar models 'warp.' Full expression: facial, eye movements, hesitation, idle motion during silent listening. 17B-parameter Diffusion Transformer with DMD distillation. No weights released. See LPM 1.0 Real-Time Avatars. Implication: real-time avatar chat replacing text chat as default interface is ~12 months out. ## Secondary stories - **RotorQuant**: open-source KV cache compression using Clifford rotators on 3-number chunks, 10x+ memory compression, 28% faster decoding, 5.3x faster prefill, 44x fewer parameters. Beats Google's TurboQuant. - **DeepSeek 'Expert Mode'**: suspected V4-lite preview, free. - **Alibaba 'Happy Horse'**: video generator from ATTH AI, led by Zong D (ex-Kling chief architect), April 30 rumored. - **Overworld Waypoint 1.5**: real-time world model on consumer GPU (RTX 3070+). - **Inspatial Spatial World**: video → explorable 3D world, 24fps on H-series. - **Nvidia Komodo**: 3D human/robot motion from text for Isaac Sim. - **ACE Step 1.5 XL**: open-source music generator close to Suno V4 quality. ## Trend lines - Chinese labs keep shipping frontier-comparable open-weight models. License restrictions are tightening (MiniMax M2.7 moved to commercial-restricted). - Meta has fully abandoned open-source positioning. - Anthropic uses 'too dangerous to release' to build moats. - Open-weight ecosystem increasingly depends on Chinese releases (Qwen, GLM, MiniMax, DeepSeek). - GLM 5.1 beating closed-source SOTA on real benchmarks is a watershed for open AI. ## Key meta-point Mythos got all the hype but GLM 5.1 is the actually-consequential story. Anthropic's 'we built something dangerous and kept it' vs ZAI's 'here are the weights' — the second changes the landscape more than the first.