What Is Project Glasswing? Anthropic's Restricted AI Program
What Is Project Glasswing? Anthropic’s Restricted AI Program
Project Glasswing is Anthropic’s invite-only enterprise program for accessing Claude Mythos Preview, their most powerful AI model. It launched on April 7, 2026 and is Anthropic’s first model deemed too capable for unrestricted public release.
Last verified: April 12, 2026
Why “Glasswing”?
Named after the transparent-winged butterfly, Project Glasswing reflects Anthropic’s stated commitment to transparency about model capabilities — even when those capabilities are dangerous. The program was announced alongside a detailed safety card from Anthropic’s red team.
What Is Claude Mythos Preview?
Claude Mythos Preview is a “step change” in AI capability, according to Anthropic. Key facts:
- Released: April 7, 2026 (Private Preview)
- Pricing: $25 / $125 per million input/output tokens
- Access: Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry
- Availability: Restricted to approved enterprise customers
What Makes It Different
Mythos Preview is notably more capable than Claude Opus 4.6 in several areas:
- Cybersecurity capabilities — Can find and exploit zero-day vulnerabilities in legacy software that human researchers missed
- Code analysis — Developed working exploits 181 times in Anthropic’s internal benchmark suite
- Reasoning depth — Significant improvements in multi-step reasoning and long-horizon planning
- General intelligence — Anthropic describes it as a “reckoning” in what AI systems can do
Why It’s Restricted
Anthropic’s red team found that Mythos Preview can:
- Identify software vulnerabilities in the OSS-Fuzz corpus that humans hadn’t flagged
- Achieve “register control” (near-complete exploit) on 29 additional targets
- Score consistently at Tier 4-5 severity (the highest levels) on crash-to-exploit ladders
The New York Times, The Guardian, and NBC News all covered the restrictions, with Anthropic explicitly stating the model is “too dangerous to publicly release” without safeguards.
How to Get Access
Project Glasswing is currently available to:
- Select Google Cloud customers — Via Vertex AI Private Preview
- Enterprise API customers — Through direct application to Anthropic
- Amazon Bedrock users — For approved enterprise accounts
- Microsoft Foundry users — With enterprise credentials
Anthropic has committed model usage credits to participants, suggesting they’re subsidizing adoption for research and enterprise security applications.
Pricing Comparison
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Claude Mythos Preview | $25.00 | $125.00 |
| Claude Opus 4.6 | $15.00 | $75.00 |
| Claude Sonnet 4.5 | $3.00 | $15.00 |
| GPT-5.4 | $10.00 | $30.00 |
Mythos Preview is the most expensive commercial AI model available, at roughly 1.7x the cost of Opus 4.6.
What This Means for AI
Project Glasswing represents a shift in how frontier AI companies release models. Instead of “ship everything to everyone,” Anthropic is gating access based on safety evaluations — a model that other labs may follow as capabilities increase.
For enterprise teams working on security, code auditing, or high-stakes reasoning tasks, Glasswing access could be worth the premium. For everyone else, Claude Opus 4.6 remains the publicly available frontier model from Anthropic.
The Takeaway
Project Glasswing is Anthropic betting that the most powerful AI models need controlled rollouts. Whether you agree with the approach or not, it signals that we’re entering an era where model capability itself becomes a gated resource — not just compute or data.
Last verified: April 12, 2026