Anthropic Mythos Pricing $25/$125: Why the most expensive and most dangerous AI model is not made public

Core conclusion

In April 2026, Anthropic quietly released its most powerful AI model - Claude Mythos Preview, priced as high as input $25/ million tokens, output $125/ million tokens, which is nearly 10 times the standard version of Claude. But it’s not the price that has aroused heated debate in the industry. It’s that Anthropic strictly limits the use of this model to about 40 security research institutions, claiming that its autonomous zero-day vulnerability discovery capabilities have crossed the security red line. However, during the same period, Anthropic successively signed computing power purchase agreements, third-party agent restriction policies, and leaked internal documents, making the outside world question: Is it not disclosed because it is too dangerous or because it is too expensive?

Key Points

• Event Date: April 7, 2026 (Project Glasswing Released)
• Model Pricing: $25/ million input tokens + $125/ million output tokens (Claude Opus is $15/$75)
• Core capabilities: 1M token context window, independent discovery and exploitation of zero-day vulnerabilities
• ACCESS RESTRICTIONS: Only available to approximately 40 invited institutions in the Glasswing program, not publicly available
• Controversy behind the scenes: Security restrictions vs insufficient computing power costs, evidence from both parties coexists

Background: What exactly is Mythos Preview?

Mythos Preview is a cutting-edge security research model released by Anthropic through the Project Glasswing initiative. Unlike OpenAI's high-profile release of GPT-5 and DeepSeek V4 as open source, Anthropic chose the most low-key approach - no blog announcement, no waitlist, and no developer invitation link.

According to Anthropic's technical documentation, Mythos Preview is only available through four cloud channels: Anthropic Direct API, Amazon Bedrock, Google Vertex AI, Microsoft Azure Foundry. However, no matter which channel, it is only open to those who have passed the Glasswing plan review.

Microsoft's technical documentation discloses the detailed specifications of Mythos: it supports 1M token context window and has a maximum output of 128,000 tokens. It is one of the longest context models currently in public documents. What's more, Anthropic's Frontier Red Team report shows that Mythos demonstrated in testing the ability to autonomously discover and exploit zero-day vulnerabilities—not by reproducing known vulnerabilities from training data, but by attacking real software flaws that have never been publicly disclosed.

Double Game: Safety or Cost?

Evidence from the safety faction

Anthropic makes it clear that access is restricted for security reasons:

"Without necessary security protections, these powerful cyber capabilities could be used to attack critical software infrastructure."

Axios quotes Anthropic's frontline red team leader as saying that Mythos' autonomy and vulnerability discovery capabilities "crossed a threshold" and that it was a reasonable decision not to disclose it until adequate protections were established.

The Glasswing program is also organized to support the security narrative — 40 select organizations (enterprise security companies, cloud providers, critical infrastructure operators), $1 billion in compute credits, joint AWS/Microsoft/Google Cloud/Cisco endorsement. Such an investment scale seems unnecessary if the computing power is not enough.

Doubts from the computing power faction

However, three things happened during the same period that made the security narrative seem less pure:

During the same period, Anthropic also restricted the third-party Agent framework's calls to Claude. The official explanation was that it "brings excessive pressure to the system." The investigation by "The Guardian" directly pointed out: The narrative itself of "Too powerful to release" is helping Anthropic win the public relations war.

The true cost of running Mythos

To understand the argument of the calculation school, we need to do some calculations.

For a 70 billion parameter model with a 128K context, the KV cache requires about 40GB of video memory. Mythos supports 1M token context, which means the KV cache requires about 320GB — this does not include model weights, activation values, and other overhead. For thousands of concurrent agent sessions, each potentially involving multi-step autonomous workflows, the computing power requirements increase exponentially.

This architecture determines that Mythos is inherently difficult to deploy on a large scale — secure or not.

Estimate the cost of a single call

That is, an in-depth Mythos Agent session might cost $25-$41. This is acceptable for vulnerability research, but far out of reach for the average developer or content creator.

Implications for AI automation practitioners

Regardless of Anthropic’s true motivations, the Mythos release model provides three important signals for AI automation practitioners:

1. The cost paradox of long context + agent If your automated workflow requires a large number of contexts to function, the running cost will increase linearly with the context length. Token economics must be taken into consideration when designing Agent workflows. Blindly pursuing long context can make operational costs spiral out of control.

2. Safety moat = cost moat When model capabilities reach a level that "may cause security risks," access control itself becomes a new business barrier. Only trusted institutions can use the most cutting-edge models, which means that in addition to technical capabilities, trust and compliance also become competitiveness.

3. The third-party Agent framework is facing tightening Anthropic's policy of restricting third-party agent calls is not an exception. As model capabilities improve, API access control will become more stringent. If your automation pipeline relies heavily on a single model provider's API, be sure to allow for migrations and fallbacks.

Tool entry (trigger tool floating card)

In the AI automation workflow, tools such as Claude / Anthropic / OpenAI / DeepSeek / n8n / LangGraph need to be flexibly combined according to cost and capability characteristics. The story of Mythos reminds us: when choosing an AI model, we should not just look at capabilities, but also at usability and economics.

Internal link guidance

• Want to learn how to replace expensive APIs with open source models? Watch: Claude Code + DeepSeek V4 Building Tutorial: API Fees Dropped by 90% (15 Minutes)
• Learn how to use n8n + OpenAI to build automated workflows and control token costs: How to use n8n + OpenAI to build an automated content collection and publishing workflow: from zero to one in 30 minutes
• Real case: He used Claude + n8n to build an AI automation system: He Built an AI Automation Stack with Claude + n8n — $4K to $12K/mo in 6 Months

Reference sources

• Kingy AI: Too Dangerous to Release — Or Just Too Expensive?
• Axios: Anthropic Mythos Preview
• Reuters: Anthropic CoreWeave
• The Guardian: Anthropic