Eighteen hours after Anthropic published the Project Glasswing update on Thursday, the AWS Bedrock model catalog quietly grew a new entry. anthropic.claude-mythos-preview-20260529-v1:0 appeared in us-east-1 and us-west-2 on Friday morning, listed as a preview-tier model with provisioned throughput capped at 50 units, an enterprise-only allowlist, and a model card linking directly to the Glasswing white paper. There was no Anthropic blog post. There was no AWS launch announcement. There was a console update, a documentation diff, and a SageMaker JumpStart entry that went live at 6:14 AM Pacific.
This is a launch. It is the launch. The Glasswing report was the marketing. The Bedrock listing is the distribution.
The mechanic is worth pulling apart because it is the new shape of frontier-model rollouts and almost nobody outside enterprise procurement is tracking it. A year ago a model release meant a flagship blog post, a benchmark sweep, a livestream, and a Twitter thread. Now the flagship blog post lands two days before the model itself, the benchmarks are buried inside a security-research PDF, and the actual GA event is a row in a JSON manifest that AWS enterprise account teams email to their top fifty customers. The buyer who matters is not on Twitter. The buyer who matters has a procurement portal open and a Bedrock invoice line item to defend at the next quarterly business review.
The Mythos Bedrock listing carries three details that tell you what Anthropic is actually selling. First, the model is gated by an account-level allowlist, which means it is not in the model-access self-service flow. You ask your AWS account team. They ask Anthropic. Anthropic decides. This is the enterprise sales motion wearing an AWS console as its skin. Second, the documented context window is 500K tokens, which is the same as Opus 4.7 but with a different throughput profile: lower tokens-per-second, higher cost-per-token, and a hard rate limit during preview. Security scanning is a batch workload, not an interactive one, and the pricing reflects that. Third, the model card explicitly names the workloads it is qualified for: code review, vulnerability discovery, log triage, and a fourth category called “structured incident reasoning” that nobody in the security space has a clean definition for yet but everybody seems to want.
The competitive read is that the AWS Marketplace channel is now the fastest way to get a frontier model in front of a Fortune 500 security budget, and Anthropic figured this out about eighteen months before the rest of the field. OpenAI is still trying to make ChatGPT Enterprise the buying surface. Google is still bundling Gemini through Workspace. xAI is still selling Grok through X Premium. Anthropic is selling Mythos through the same procurement portal that already has a five-year contract for EC2, S3, and CloudWatch, and the procurement officer signs a CO. The friction profile is not the same.
The Clank-flavored read is that we are watching the end of the era when AI launches looked like consumer-tech launches. The frontier-model launch of 2026 is a row in a manifest, a model card written in compliance language, and a security-research PDF that does the marketing for a product the customer cannot actually buy off the shelf. The cool people on Twitter will keep arguing about benchmark scores. The CISO at JPMorgan Chase will keep filling out the access request. Both of them are participating in the same launch, and only one of them gets to use the model.