Claude Mythos on Vertex AI | Access Requirements and Glasswing Eligibility
If you're wondering whether you can use Claude Mythos through Vertex AI, this article covers the delivery format, access requirements, pricing, and differences from other cloud providers in detail, based on primary sources.
目次 (9)
- Can You Use Claude Mythos on Vertex AI? — The Bottom Line
- What Is Claude Mythos Preview? — A Security-Focused Model Exclusive to Glasswing
- What Vertex AI Offers — Enterprise Governance with Data Staying Within GCP
- Access Requirements — Project Glasswing and the Invitation-Only Model
- Pricing — Reported Unit Rates and Credit Support
- Differences from AWS Bedrock and Microsoft Foundry
- The "Preview" Label Disappearance — Is GA Around the Corner?
- Key Considerations for Japanese Organizations Using Mythos via Vertex AI
- Summary
Can You Use Claude Mythos on Vertex AI? — The Bottom Line
On April 8, 2026, Google Cloud announced via its official blog that Claude Mythos Preview is available as a Private Preview on Vertex AI. Access is limited to "a select group of Google Cloud customers" and is positioned as part of the Project Glasswing framework. Unlike standard Claude models, you cannot simply enable it through Model Garden — the invitation-only nature of access is the most important prerequisite.
In short, the answer to "can I use it on Vertex AI?" is: "Technically, Vertex AI is one of the official delivery channels, but access is restricted to organizations participating in Glasswing." It's also worth noting that this is not positioned as a general-purpose business model — Anthropic describes it as a model with a renewed focus on reducing cybersecurity risks.
Source: Claude Mythos Preview on Vertex AI — Google Cloud Blog
What Is Claude Mythos Preview? — A Security-Focused Model Exclusive to Glasswing
Claude Mythos Preview is Anthropic's latest frontier-class model, announced in April 2026. Anthropic states it surpasses previous Claude models in cybersecurity, large codebase comprehension, and complex reasoning, and is capable of identifying and demonstrating high-severity software vulnerabilities with minimal human intervention. It has already discovered numerous previously undetected vulnerabilities in major operating systems and web browsers.
Rather than being offered as a standalone model, it is distributed as the dedicated model for Project Glasswing, an industry-wide initiative launched by Anthropic. Glasswing aims to give defenders an advantage before AI's code-analysis capabilities can be weaponized, and targets organizations protecting critical infrastructure, major operating systems, and large-scale open-source software. Access through Vertex AI cannot be separated from membership in Glasswing.
Source: Project Glasswing — Anthropic
What Vertex AI Offers — Enterprise Governance with Data Staying Within GCP
The primary reason to use Mythos via Vertex AI is the ability to apply Google Cloud's enterprise governance controls directly. Google Cloud's official blog describes the Mythos Preview as being combined with "Vertex AI's enterprise-grade foundation for building, scaling, and governing AI applications and agents."
For organizations using Mythos for security work, the following controls built into Vertex AI provide practical reasons for choosing this path:
- Communication boundary control via VPC Service Controls, with data designed to remain within Google Cloud's perimeter
- Data residency (storage region) and access audit logs for compliance
- Anthropic's policy of not using customer data accessed via Vertex AI to train foundation models
- Performance stability through dedicated capacity and prompt caching
- Extensibility to build agents integrated with BigQuery and Cloud API Registry
Given that vulnerability data is among the most sensitive information imaginable, the ability to control where results flow — in alignment with your organization's governance requirements — carries particular weight for a model like Mythos.
Source: Claude on Google Cloud Vertex AI — Claude
Access Requirements — Project Glasswing and the Invitation-Only Model
Even though Vertex AI is one of the delivery channels, self-serve sign-up is not possible. Anthropic has explicitly stated that "there are no plans for general availability (GA) at this time," and access is by invitation only. Eligible parties are Glasswing founding partners and certified organizations responsible for critical infrastructure.
Glasswing founding partners include AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks, and more, joined by over 40 organizations that build and maintain critical software. The eligibility benchmark is described as "maintaining software that affects hundreds of millions of users," and use cases are strictly limited to defensive security.
If you are considering access via Vertex AI, here is a practical path forward:
- Confirm whether your organization qualifies for Glasswing (critical infrastructure, major OSS, defensive security responsibility)
- Align with legal and compliance teams that use will be strictly limited to defensive security
- Apply for individual review through your Google Cloud account team or through Anthropic's contact channels
- After passing review, wait for access to be activated on Vertex AI for your designated project
Source: Project Glasswing — Anthropic
Pricing — Reported Unit Rates and Credit Support
The Google Cloud official blog does not disclose pricing for Mythos Preview. However, multiple media outlets and explainer articles report pricing of $25 per million input tokens and $125 per million output tokens. This is reported as a rate shared across delivery channels — including the Claude API, Amazon Bedrock, and Microsoft Foundry — not specific to Vertex AI.
Additionally, Anthropic has stated it will provide up to $100 million USD in model usage credits for Project Glasswing participants and additional partner organizations throughout the research preview period. This means eligible organizations may have their initial costs offset by credits. That said, the reported figures above are not officially confirmed primary-source data, so budget planning should include a caveat that "pricing may be revised when GA launches."
※ Based on reporting as of June 2026. For official confirmed figures, refer to the Google Cloud official page.
Source: 3 ways to obtain Claude Mythos API — Apiyi
Differences from AWS Bedrock and Microsoft Foundry
Claude Mythos Preview is also available through the Claude API, Amazon Bedrock, and Microsoft Foundry in addition to Vertex AI. Even with the same model, constraints and usability vary depending on the underlying cloud platform.
- Vertex AI (Google Cloud): Private Preview. Allows direct application of Google Cloud governance features such as VPC Service Controls and data residency; BigQuery integration and the broader Google Cloud ecosystem are key strengths.
- Amazon Bedrock (AWS): Available only in the US East (us-east-1) region, pay-as-you-go only. Reported to currently lack support for Guardrails, Knowledge Bases, Agents, the Token Counting API, and cross-region inference.
- Microsoft Foundry / Claude API: Named as delivery channels, but fewer details have been published compared to Vertex AI and Bedrock.
If your organization is already centered on Google Cloud and you want to keep data within GCP boundaries, Vertex AI is the natural choice. Conversely, if your existing cloud contracts and region requirements are AWS-centric, Bedrock is the natural alternative.
Source: 3 ways to obtain Claude Mythos API — Apiyi
The "Preview" Label Disappearance — Is GA Around the Corner?
One topic that attracted attention around Vertex AI in May 2026 was the observation that the "Preview" label had disappeared from the base-model-claude-mythos entry in Google Cloud's internal model listing. Watchers on X reported that an entry that had not been visible the day before showed a changed label in Vertex AI's Model Garden on May 17, 2026.
Some interpreted this change as a sign that general availability was imminent, but commentary was measured: "This is not the precursor to a public launch — it is an infrastructure preparation stage." In terms of Anthropic's three-phase rollout roadmap, it was characterized as being "at the entry point between Phase 2 and Phase 3," with the note that an official announcement would precede any actual release. Crucially, no official announcement from Anthropic accompanied this change.
Therefore, even if the Mythos label in the Vertex AI console has changed, it would be premature to conclude that "anyone can now use it." Whether general availability has been granted should be determined by waiting for official announcements from Anthropic and Google Cloud.
Source: Claude Mythos in Google Cloud: What the Missing "Preview" Label Means — BuildFastWithAI
Key Considerations for Japanese Organizations Using Mythos via Vertex AI
Organizations based in Japan considering Mythos through Vertex AI should confirm several things before getting to the technical details:
- Whether your organization could qualify for Glasswing (do you bear responsibility for critical infrastructure, major OSS, or defensive security?)
- Whether use can be strictly limited to defensive security (general development assistance and analysis of third-party products are out of scope)
- Whether the data storage region and cross-border transfer satisfy your organization's information governance requirements
- Whether your organization has the internal resources to fix discovered vulnerabilities (discovering vulnerabilities and leaving them unpatched actually increases risk)
- Whether you can separate the payment channel for Mythos from your existing Claude (Sonnet, Opus) contracts
If any of these answers is "No," applying at this stage would be premature. For everyday development support, using the standard Claude models (Sonnet, Opus) on Vertex AI is more realistic — the prudent approach is to consider Mythos only once a clear defensive security need has been firmly established.
Summary
Putting together the answer to "vertex claude mythos":
- Delivery format: Vertex AI is one of the official channels for Mythos Preview, but it is Private Preview and invitation-only.
- Prerequisite: Using it requires participation in Project Glasswing; self-serve sign-up is not available.
- Vertex AI's strengths: Google Cloud governance features such as VPC Service Controls and data residency can be applied directly, making it easier to keep data within GCP boundaries.
- Pricing: No official unit pricing disclosed. Reported figures are $25 input / $125 output (per million tokens), with up to $100 million in credit support available. (※ Based on reporting as of June 2026. Refer to the Google Cloud official page for confirmed figures.)
- GA outlook: The May label disappearance remains an observation only; no official announcement has been made.
For critical infrastructure organizations that are centered on Google Cloud and want to strengthen defensive security, access via Vertex AI is a compelling option. For all other organizations, the rational approach is to first solidify their use of the standard Claude series on Vertex AI and wait for Anthropic's general availability announcement for Mythos.
Sources:
