In an unusual display of transparency, Anthropic has publicly admitted that its newly released Claude Opus 4.7 is "less broadly capable" than an unreleased model—the company's own Claude Mythos Preview. This concession, announced on April 16, 2026, marks a significant shift in how AI labs communicate about model capabilities and safety limitations.
The acknowledgment came as Anthropic shipped Opus 4.7, its most powerful publicly available AI model to date. While the upgrade delivers meaningful improvements over Opus 4.6—including better coding performance, sharper vision, and enhanced self-verification—Anthropic confirmed it deliberately built the model with reduced cyber capabilities as part of Project Glasswing's safety framework.
Why Mythos Preview remains locked away
Claude Mythos Preview represents the frontier of Anthropic's model development. The system has demonstrated capabilities that exceed Opus 4.7 across every evaluation metric, particularly in cybersecurity applications. According to reports, Mythos Preview scored 83.1% on the CyberGym benchmark compared to Opus 4.7's 73.1%—a 10-percentage-point gap with profound implications for security use cases.
Anthropic revealed that Mythos Preview has autonomously identified thousands of zero-day vulnerabilities across every major operating system and web browser. This dual-use capability—powerful for defensive security but equally dangerous in malicious hands—prompted the company to restrict access through Project Glasswing, launched April 7, 2026.
Project Glasswing: A tiered approach to AI access
Project Glasswing represents a new paradigm in tiered AI access. Anthropic partnered with major technology and financial institutions—including Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, and the Linux Foundation—to provide controlled access to Mythos Preview for defensive security work.
"AI capabilities have crossed a threshold that fundamentally changes the urgency required to protect critical infrastructure from cyber threats,"
Anthropic's Project Glasswing announcement
Anthropic committed $100 million in usage credits and $4 million in donations to open-source security organizations to support this defensive initiative. Participants access Mythos Preview at $25/$125 per million input/output tokens through the Claude API, Amazon Bedrock, Google Cloud's Vertex AI, and Microsoft Foundry.
The implications for AI development
This strategy of releasing a "civilian" version while gating frontier capabilities behind verified partnerships signals a maturing approach to AI safety. Rather than deploying their most capable models universally, Anthropic is experimenting with "efforts to differentially reduce these capabilities" during training—deliberately limiting cyber offense potential while maintaining utility for legitimate applications.
For growing organizations navigating this complex landscape, the tiered access model presents both opportunities and challenges. While specialized partners can navigate both generally available models and restricted-access programs, the broader question remains: how will the industry balance capability democratization against legitimate safety concerns as frontier AI systems grow increasingly powerful?
Leave a Comment
Sign in to join the discussion and share your thoughts.
Login to Comment