Mythos Preview hits 93.9% SWE-bench, remains restricted

// 45d agoMODEL RELEASE

Mythos Preview hits 93.9% SWE-bench, remains restricted

Anthropic's restricted "super-frontier" model outperforms Opus 4.7 across all benchmarks, setting a new record of 93.9% on SWE-bench Verified. The model is currently limited to defensive cybersecurity partners in Project Glasswing due to its high capability for autonomous zero-day discovery and exploitation.

// ANALYSIS

Anthropic is building a "Manhattan Project" for cybersecurity, prioritizing infrastructure defense over general accessibility.

–The 13-point jump on SWE-bench Verified signals a massive leap in reasoning and autonomous software engineering.
–Autonomous discovery of 27-year-old vulnerabilities makes this model a high-risk asset that could weaponize hacking if leaked.
–Project Glasswing’s $100M in credits and $4M in donations aim to secure global software before offensive models catch up.
–A 100% score on Cybench marks the end of existing security benchmarks, requiring a total overhaul of AI evals.

// TAGS

llmai-codingagentreasoningbenchmarksafetyresearchclaude-mythos-preview

DISCOVERED

45d ago

2026-04-16

PUBLISHED

45d ago

2026-04-16

RELEVANCE

9/ 10

AUTHOR

Bijan Bowen

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

NEWS47m ago

Global LLM capabilities rapidly converge

A discussion on X highlights the noticeable convergence in capabilities among top-tier Large Language Models, noting that both Western and Chinese models—such as MiniMax's multimodal offerings—are exhibiting highly comparable performance, distinguished primarily by their unique quirks and specialized strengths. As foundation models across developers reach a parity plateau, industry observers are questioning if we have entered a phase of diminishing returns for current architectures and when the next definitive leap forward in artificial intelligence will emerge.

MODEL1h ago

MiniMax-M3 hits OpenRouter with 50% discount

MiniMax-M3 has launched on OpenRouter, offering a frontier-class open-weight model designed for long-context multimodal tasks, coding, and agentic workflows. To drive developer adoption, OpenRouter is offering a 50% discount on API usage for the model's first week.

UPDATE1h ago

OpenCode 1.15.13 promotes ACP support

OpenCode has released version 1.15.13, bringing significant under-the-hood enhancements and standardized Agent Client Protocol (ACP) support to the open-source terminal-native AI coding agent. The update introduces deeper v2 state plumbing, improves TUI/desktop session flows, and resolves provider transport issues across Vertex AI, OpenAI, MCP, and Windows PTY.