The Headline: Anthropic’s Most Dangerous Model Isn’t Public
The biggest AI story this month isn’t a benchmark score — it’s a model Anthropic won’t let you use. Claude Mythos, released in limited preview via Project Glasswing, is a cybersecurity-focused model capable of identifying zero-day vulnerabilities across every major OS and browser. Anthropic has locked it inside a vetted coalition of roughly 40 partners: Microsoft, Apple, Amazon, and the NSA are among them.
The decision to keep Mythos gated has triggered what CNBC called a cybersecurity “hysteria” — and a predictable open-source response. Developer Kye Gomez published OpenMythos on GitHub, a theoretical reconstruction of Mythos’s architecture built from public research. His central guess: Mythos is a Recurrent-Depth Transformer, a looped architecture that runs a smaller stack through itself multiple times per forward pass rather than stacking layers. The repo crossed 10,000 stars within weeks.
OpenAI responded with Daybreak, its own cybersecurity initiative, granting EU vetted teams limited preview access to GPT-5.5-Cyber.
What’s Actually Shipping
OpenAI made GPT-5.5 Instant the default ChatGPT model after its April 23 release. The key improvements: hallucinated claims down more than 50% in high-stakes scenarios, and expanded memory — the model can now draw context from past conversations, uploaded files, and connected services like Gmail, with new “memory sources” controls that show users exactly what influenced each response.
OpenAI also dropped three real-time audio models for conversational agents:
- GPT-Realtime-2 — conversational task execution
- GPT-Realtime-Translate — multilingual translation across 70+ languages
- GPT-Realtime-Whisper — live transcription and captioning
Anthropic separately released Claude Opus 4.7, scoring 57.3 on the AA Intelligence Index with a 1M token context window. It’s positioned below Mythos in capability but available to standard API customers.
Google released Gemma 4, a new open model series built for reasoning and agentic workflows, under Apache 2.0. Google’s research team also presented TurboQuant at ICLR 2026 — an algorithm that compresses the KV cache (the main memory bottleneck in long-context inference) using PolarQuant vector rotation combined with Quantized Johnson-Lindenstrauss compression. The practical effect: models with massive context windows run far more efficiently, applying pressure to the assumption that bigger always means better.
Meta shipped Muse Spark (AA Intelligence Index: 52.1, 262K context) and launched Incognito Chat — a privacy mode for Meta AI in WhatsApp where, Meta claims, even Meta can’t read the conversation.
The Infrastructure Story That Matters More Than Any Model
Anthropicʼs Model Context Protocol (MCP) crossed 97 million installs in March 2026. That number signals a transition from experimental standard to foundational infrastructure. By May, virtually every major AI framework and enterprise tool ships with native MCP compatibility.
This matters because it shifts the competitive bottleneck. Model capability — which lab has the smartest weights — is no longer the primary constraint on what agents can do. The constraint is now integration: 46% of enterprise teams cite connecting agents to existing systems as their top challenge, according to a recent Arcade.dev survey.
Practical examples of where this is landing in production right now:
- Broadridge (May 13): agentic workflows chaining data and context to automate exception resolution in post-trade finance
- Notion: Developer Platform with an External Agent API for custom agent deployment inside workspaces
- Honeycomb: Agent Timeline views and reusable Canvas Skills for multi-agent debugging workflows
What the Pattern Suggests
Five things are true simultaneously right now:
- Frontier capability is bifurcating — Mythos-class models exist but are gated; the public frontier (GPT-5.5, Opus 4.7) is genuinely powerful but a tier below.
- Efficiency is the new scaling — TurboQuant, Flash Lite, and Gemma 4 all prioritize doing more with less over raw parameter growth.
- Audio and real-time are catching up — the gap between text and voice AI is closing fast.
- MCP won the protocol war — the tooling layer is standardizing around it whether or not you chose it deliberately.
- Agentic deployment has left experimentation — the question is no longer “can agents work?” but “how do we operate them reliably at scale?”
The labs are no longer racing purely to make models smarter. They’re racing to make models deployable.