Anthropic's Claude Mythos Preview is no longer just a leaked frontier model. It is now part of a guarded cybersecurity program that gives selected organizations access through platforms including Google Cloud's Vertex AI.
Anthropic has put its most sensitive new Claude model into a controlled rollout, and the signal for startups is clear: the next phase of frontier AI will be shaped as much by access, cloud distribution and security controls as by raw benchmark gains.
The company announced Claude Mythos Preview on April 7 as part of Project Glasswing, a cybersecurity initiative built around defensive use of a model Anthropic says is its most capable yet for coding and agentic tasks. Participants can access Mythos through the Claude API, Amazon Bedrock, Google Cloud's Vertex AI and Microsoft Foundry, with Anthropic setting pricing at $25 per million input tokens and $125 per million output tokens for Project Glasswing participants.
That matters because Mythos is not being treated like a normal model launch. Anthropic says the model has already identified thousands of zero-day vulnerabilities across critical infrastructure, which is useful for defenders but uncomfortable if the same capability reaches attackers first. According to TechCrunch, 12 partner organizations are involved in Project Glasswing, while about 40 organizations in total are expected to get access to the preview.
The practical takeaway for startups is immediate. If Mythos becomes the shape of future premium Claude tiers, teams building agentic coding tools, security products or high-context enterprise workflows should expect a different cost profile. The model is powerful, but it is also expensive, gated and operationally sensitive. That combination changes product planning.
Why Vertex AI matters
Google Cloud's role is important because Vertex AI is already where many enterprise customers manage model access, permissions, logging and compliance. Putting Mythos inside that environment gives Anthropic a way to reach large customers without asking them to move sensitive workflows into a new operational setup.
For founders, this is a reminder that model choice and cloud choice are becoming harder to separate. A startup that runs most of its stack on AWS or Azure may still need a GCP path if the best access to a specific Claude capability sits inside Vertex AI. That does not mean every AI startup should move to Google Cloud. It means infrastructure planning now has to account for where frontier models actually show up, not just where compute is cheapest.
There is also a procurement angle. Large companies often prefer buying AI through existing cloud contracts because security reviews, billing and compliance workflows are already in place. Anthropic benefits from that channel. Google benefits because model availability gives customers another reason to stay inside its platform. Startups that sell into enterprises should pay attention, because the cloud marketplace can become part of the product strategy, not just the deployment layer.
The pricing question
Mythos' published Project Glasswing pricing points to a future where frontier inference is not a commodity line item. At $25 per million input tokens and $125 per million output tokens, it is positioned for selective use, not casual experimentation at scale.
That distinction matters in product design. A company building a code review assistant, an autonomous security scanner or an internal research agent may not want every task routed through Mythos. A better architecture may use lower-cost models for routine work, then escalate to Mythos only when the task requires deeper reasoning, vulnerability analysis or complex software changes.
This is where smaller startups can still compete. They may not have the same reserved capacity as major enterprises, but they can design systems that spend carefully. Routing, caching, evaluation layers and model fallbacks will become more important as the gap between everyday models and restricted frontier models widens.
Access is now part of the product
The guarded rollout also shows why security policy is becoming a commercial feature. Anthropic is clearly trying to balance two truths at once: defenders need stronger AI tools, and stronger AI tools can also lower the barrier for offensive cyber activity.
That concern is not theoretical. TechRadar, citing Bloomberg, reported in April that Anthropic was investigating claims of unauthorized access to Claude Mythos Preview through a third-party vendor environment. Anthropic said it had no evidence that access extended beyond that vendor environment, but the episode underlined the pressure around restricted model programs.
For startups, the lesson is not simply to wait for public access. Teams should build for uneven availability. If a product depends entirely on one gated model, it risks becoming fragile the moment access terms, pricing or compliance rules change. A more durable approach is to separate the application logic from the model layer, so teams can switch between Claude tiers, other frontier models and lower-cost alternatives as the market shifts.
Anthropic's Mythos rollout is current because it sits at the intersection of three live questions: how powerful AI models should be distributed, which cloud platforms will control access to them, and how much startups will pay to use them responsibly. The next thing to watch is whether Mythos remains a defensive research preview or becomes the template for a new class of premium Claude models built for enterprise-scale coding and cybersecurity work.
Also read: Mustafa Suleyman's 18-month warning forces startups to rethink work • Mustafa Suleyman's 18-Month Claim Forces Startups to Rethink Hiring and Product Bets • Mustafa Suleyman's 18‑Month Claim Forces Startups to Choice: Rebuild Teams or Rebuild Strategy
