Too Dangerous, or Too Expensive? The Real Reason Anthropic Is Hiding Mythos

Anthropic did not open Claude Mythos Preview as self-serve, instead routing it through a limited-access research program called Glasswing. Access is granted to about 40 organizations, and it is only available via the Anthropic API, AWS Bedrock, Google Vertex AI, and Microsoft Azure Foundry.

• Pricing is $25 per 1M input tokens and $125 per 1M output tokens.
• Anthropic is providing participants with up to $100 million in credits.
• Microsoft's documentation mentions a 1M context window and 128k output.

The security-side rationale comes from Anthropic's frontier red-teaming results. Rather than merely repeating known vulnerabilities, Mythos demonstrated the ability to actually find previously undisclosed zero-days, and the risk report described the model as more agentic and more powerful than before. Anthropic believes that if such capabilities were released broadly, they could benefit attackers before defenders.

At the same time, signals of compute pressure were also strong. Anthropic announced a large-scale TPU expansion deal with Google and Broadcom, additional CoreWeave capacity, and a review of its own AI chip around the same time, alongside restrictions on agentic harnesses and an increase in outages. A leaked internal draft also included language to the effect that Mythos was not yet ready for general release and that operating costs were high.

So the piece doesn't settle on a single conclusion. The security risk appears real, but cost and infrastructure constraints also read as an important part of what pushed this toward a limited release.