Ali Asif

The AI log, frontier · open weights · agents and coding

Covers 6 Apr 2026 to 12 Apr 2026

Anthropic gates Mythos Preview, Meta's Muse Spark, GLM-5.1

Project Glasswing hands a withheld frontier model to defenders, Meta shows its first Superintelligence Labs model, and GLM-5.1 opens its weights.

By , published

What shipped

  1. Anthropic

    Project Glasswing and Claude Mythos Preview

    A frontier model held back from general release and given to defenders first, backed by up to $100M in usage credits and $4M for open-source security groups.

  2. Meta Superintelligence Labs

    Muse Spark

    The first model from Meta's new lab is natively multimodal and proprietary, with a private API preview and a Contemplating mode that runs agents in parallel.

  3. Z.ai

    GLM-5.1

    Z.ai's release notes date GLM-5.1 to April 7 and describe a model that can work on one task autonomously for up to eight hours.

  4. Anthropic

    Claude Managed Agents

    Public beta on April 8 of hosted agent infrastructure that handles sandboxing, sessions and scoped permissions.

  5. arXiv

    ClawBench: Can AI Agents Complete Everyday Online Tasks?

    A benchmark for finishing everyday web tasks that topped Hugging Face's weekly papers.

  6. arXiv

    OpenWorldLib: A Unified Codebase and Definition of Advanced World Models

    An attempt to pin down what counts as a world model and give researchers one shared codebase.

  7. arXiv

    Video-MME-v2

    A harder benchmark for video understanding, second on the weekly papers list.

The biggest news this week was a model nobody can download.

Glasswing and Mythos Preview

On April 7 Anthropic announced Project Glasswing, built around Claude Mythos Preview. The company calls it an unreleased general-purpose frontier model, and says it has already found thousands of high-severity zero-day vulnerabilities across every major operating system and web browser. The examples it names are a 27-year-old OpenBSD bug that lets an attacker crash a machine remotely, and a 16-year-old FFmpeg flaw in a line of code that automated tests had hit five million times. Those are patched.

Launch partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA and Palo Alto Networks, plus more than 40 other organizations that maintain critical software. Anthropic commits up to $100M in usage credits and $4M in donations to open-source security groups. It also says it does not plan to make the model generally available, and promises a public report on lessons within 90 days. After the preview, pricing is $25 and $125 per million input and output tokens.

I keep returning to the release decision. It is the first time I have seen a lab say, in effect, that capability alone is a reason to narrow access.

Muse Spark

Meta introduced Muse Spark on April 8, the first model from Meta Superintelligence Labs. It takes text, image and video and answers in text, runs through meta.ai and the Meta AI app, and opens a private API preview to selected partners. Meta reports 58% on Humanity's Last Exam in its Contemplating mode, which orchestrates several agents reasoning in parallel, and says it reaches Llama 4 Maverick's capability with over an order of magnitude less compute. Those are Meta's numbers.

Open weights and agents

Z.ai's release notes put GLM-5.1 at April 7, a model it says can stay on one task for up to eight hours. On the tooling side, Claude Managed Agents went into public beta on April 8, which means Anthropic now hosts the sandbox, the sessions and the permissions that many of us built by hand. The week's top paper was ClawBench, a test of whether agents can finish everyday web tasks, and the world-model and video benchmarks followed.

What I'd watch next: whether other labs copy the gated-release pattern, and how the 90-day Glasswing report reads.

Sources

Topics: Anthropic, arXiv, GLM, Meta, Z.ai, Muse Spark, Managed Agents, Meta Superintelligence Labs