Daily AI & LLM Trends Report — April 12, 2026
The week of April 7, 2026 rewrote the story of AI's most capable models. Here's what you need to know.
🔒 Claude Mythos: Anthropic Builds Its Best — and Locks It Away
Anthropic confirmed on April 7 that Claude Mythos is the most capable model it has ever built — and then immediately gated it behind Project Glasswing, a closed early-access program limited to ~50 critical infrastructure partners (AWS, Apple, Microsoft, Google, NVIDIA, Cisco, CrowdStrike, JPMorgan, and others). No public API. No general availability date.
Why? Anthropic explicitly views Mythos as a cybersecurity risk. Internal drafts leaked (via Roy Paz and Alexandre Pauwels, March 28) warn that Mythos "presages an upcoming wave of models that can exploit vulnerabilities in ways that far outpace the efforts of defenders." The model can reportedly scan entire OS kernels and large codebases for exploitable flaws — including bugs undetected for decades.
Preview pricing: $25/M input tokens, $125/M output tokens — but only for Glasswing partners.
Context: Anthropic spent March in a standoff with the Pentagon after refusing to allow Claude in autonomous weapons systems. Multiple U.S. agencies began a six-month phase-out of Claude models, labeling the company a "supply-chain risk."
🟢 GLM-5.1: The Model That Beat the Frontier — and Costs Nothing
The same day Anthropic was locking things down, Zhipu AI dropped GLM-5.1 under the MIT license — completely open, no restrictions on commercial use, modification, or redistribution.
- 744B total parameters (MoE architecture, 40B active per forward pass)
- 200K context window
- SWE-Bench Pro #1 — reportedly beat both Claude Opus 4.6 and GPT-5.4 on expert-level real-world software engineering tasks
- API cost: ~$1/$3.2 per million tokens; self-host: free
This is now the strongest coding model you can run on your own hardware, by at least one credible benchmark. Zhipu AI has been on a steady climb: GLM-4.5 → 4.6 → 4.7 → 5 → 5.1, each iteration more capable and more openly licensed.
🔵 Gemma 4: Google's Open-Source Push
Google open-sourced the Gemma 4 series on April 1, including:
- Gemma 4 27B — text + image + audio, Apache 2.0
- Gemma 4 26B-A4B and E2B/E4B variants
- Free to self-host, Apache 2.0 license
Gemma 4 focuses on reasoning and agentic workflows, a departure from Google's earlier fine-tuning-heavy approach.
📋 Full Model Release Log (April 1–8, 2026)
| Date | Model | Developer | Type | License | Price |
|---|---|---|---|---|---|
| Apr 1 | Gemma 4 27B | Text+Image+Audio | Open (Apache 2.0) | Free | |
| Apr 1 | GLM-5V-Turbo | Zhipu AI | Vision+Code | Proprietary (API) | API pricing |
| Apr 1 | Bonsai 8B | PrismML | Text | Open | Free |
| Apr 2 | Qwen 3.6-Plus | Alibaba | Text+Agentic | Open | ~$0.28/M |
| Apr 2 | MAI-Transcribe-1 | Microsoft | Speech-to-Text | Proprietary | Azure pricing |
| Apr 2 | MAI-Voice-1 | Microsoft | Voice Generation | Proprietary | Azure pricing |
| Apr 2 | MAI-Image-2 | Microsoft | Image+Video Gen | Proprietary | Azure pricing |
| Apr 7 | GLM-5.1 | Zhipu AI | Text+Reasoning | Open (MIT) | ~$1/$3.2/M |
| Apr 7 | Claude Mythos | Anthropic | Text+Reasoning+Cyber | Gated | ~$25/$125/M |
🧩 This Week's Broader Trends
Multimodal Embedding Goes Mainstream Hugging Face published major updates to sentence transformers enabling joint text+image embedding and reranking — a foundational capability for better search, recommendation, and generation systems.
On-the-Job Learning for AI Agents IBM Research and Hugging Face unveiled ALTK-Evolve, a technique enabling AI agents to learn and adapt in real-time without explicit supervision — potentially changing how agents improve post-deployment.
Enterprise AI at Scale OpenAI published a case study showing CyberAgent (Japan) scaling ChatGPT Enterprise and Codex across advertising, media, and gaming — improving decision speed and quality across the board.
RAG Gets Practical Towards Data Science released an enterprise-grade guide to grounding LLMs with Retrieval-Augmented Generation (RAG) for knowledge base applications — signaling RAG has matured from experiment to production standard.
📌 The Bigger Picture
Early April 2026 isn't just about benchmark scores. It's about a philosophical split in the AI industry:
- Anthropic's path: Build the most capable model, gate it behind a defensive program, charge premium prices for a closed club
- Zhipu's path: Build a frontier-competitive model, give it away freely, let the community build on it
The price gap this week: free to $125 per million output tokens — for models that are reportedly close in capability. The split is no longer about capability. It's about who gets to use it, and on whose terms.
Sources: whatllm.org, llm-stats.com, tokencalculator.com, Build Fast With AI, dev.to, Hugging Face blog, OpenAI blog, Towards Data Science