TL;DR: Anthropic built an AI model so good at finding software vulnerabilities they won’t release it publicly. Meanwhile, Meta and ZAI dropped new models, Google added interactive simulations to Gemini, and Seed Dance 2.0 finally hit the US.
This was one of those weeks where you blink and miss five announcements. The creator of this roundup, Matt Wolfe from Future Tools, pulled together everything worth knowing from the past seven days, and there’s a lot to unpack.
The biggest story is Claude Mythos and Project Glasswing. Anthropic built a model that scores 83.1% on cybersecurity vulnerability reproduction (up from 66.6% for Opus 4.6) and absolutely dominates software engineering benchmarks. The catch? They’re not releasing it. According to Anthropic’s own words, this model “found thousands of high severity vulnerabilities, including some in every major operating system and web browser.” We’re talking a 27-year-old vulnerability in OpenBSD and a 16-year-old one in FFmpeg. Instead of a public launch, they created Project Glasswing, giving cybersecurity specialists at major tech companies access to patch vulnerabilities before models this powerful proliferate.
As the expert points out, there’s always a marketing element to “this model is too scary to release” announcements. GPT-2 got the same treatment back in 2019. But this time the reasoning feels more grounded. The model wasn’t even trained specifically for cybersecurity. It just got so good at code that hacking became a side effect.
📦 Two new models you can actually use
- Meta Muse Spark – First release from Meta’s Super Intelligence Labs (post-Yann LeCun). Jumped from last place to fourth on the Artificial Analysis Intelligence Index. Not the best at anything specific, but competitive across the board and extremely token-efficient, meaning it should be cheap to run. Available now at Meta AI.
- GLM-5.1 from ZAI – This is the sleeper hit. An open-source model under MIT license scoring 58.4 on SWEBench Pro, beating GPT 5.4 (57.7) and Opus 4.6 (57.3). You can download the weights from HuggingFace right now and fine-tune it for your own use cases. Open-weight models matching frontier lab performance is a massive deal.
📌 Other notable drops this week
- Gemini interactive simulations – Google’s pro model now generates interactive visualizations with adjustable sliders, similar to what OpenAI and Anthropic shipped recently
- Gemini Notebooks – Basically Google’s version of Claude Projects or ChatGPT Projects, with dedicated spaces for organizing chats and files. Syncs with NotebookLM
- Seed Dance 2.0 – Finally available in the US through Runway and CapCut. Fast generation, solid quality, arriving right after Sora’s exit
- HeyGen Avatar V – Captures your identity from just 15 seconds of video. Lip-sync still needs work, but the speed of cloning is impressive
- OpenAI’s new $100/month tier – Sits between Plus ($20) and the old Pro ($200), offering 5x more Codex usage
- Claude Managed Agents – New feature tying agents to tools like Notion, Slack, and Asana through the Claude console
- Anthropic cut third-party tool access – Claude subscriptions no longer cover usage in tools like OpenClaw, frustrating users who built workflows around it
- Perplexity + Plaid – Connect your financial accounts for a personalized finance dashboard with read-only access
- HappyHorse 1.0 – A mysterious video model (reportedly from Alibaba) topping leaderboards above Seed Dance 2.0
- Google AI Edge – Fully offline dictation app for iOS using the Gemma model
- Spotify prompted podcasts – AI-generated podcast playlists based on your interests
The pace isn’t slowing down. If you want the full breakdown with demos and visual comparisons, the complete video is worth your time.