Claude Mythos Preview

What is Claude Mythos Preview?

Claude Mythos Preview was an invitation-only research model from Anthropic, identified in the API by claude-mythos-preview. It was developed within Project Glasswing and targeted organizations working on defensive cybersecurity, autonomous coding, vulnerability discovery, attack-surface auditing, and other long-running agent workflows.

This was not a broadly available consumer model or a standard generally available API offering. Access was limited to approved Project Glasswing organizations, and availability could vary across Anthropic-operated API routes and cloud integrations. Anthropic later deprecated the preview on June 9, 2026 and recommended migrating to Claude Mythos 5.

For readers evaluating the model today, its most important distinction is therefore its combination of very large context and specialized research-preview access. The model could process extremely large amounts of material in one request, but it was not a practical default for a new application that requires stable availability and predictable commercial terms.

Purpose and position in Anthropic’s lineup

Claude Mythos Preview occupied a specialized position in the Claude model family. It was designed for demanding technical work rather than everyday chat, routine summarization, or broad consumer access. Project Glasswing organizations could use it for security research, large-scale code analysis, autonomous programming tasks, and agents that needed to work through extended sequences of reasoning and tool calls.

The model’s successor, Claude Mythos 5, is the relevant migration target identified by Anthropic. The migration is described as broadly compatible at the Messages API and tool-use levels, although developers need to account for differences in thinking configuration and output behavior. Claude Mythos 5 uses always-on adaptive thinking, while the preview supported both configurable manual thinking and adaptive thinking.

Verified capabilities and limits

The one-million-token context window is the model’s clearest technical advantage. A context window is the amount of information the model can consider during a request and conversation. At this size, suitable workloads could include whole-codebase review, large document collections, extensive security logs, or long agent sessions without repeatedly splitting the material into smaller pieces.

The maximum output was 128,000 tokens. That limit describes the amount of text the model could potentially produce in one response; it does not mean every request would use or need that much output. Long outputs can increase latency and cost, so a well-designed application should still request only the amount of explanation, code, or analysis it needs.

Input and output modalities

Claude Mythos Preview accepted text and image input and generated text output. Image input made it possible to include visual material alongside written instructions, such as screenshots, diagrams, interface captures, or other supported visual evidence. The supplied specifications do not indicate native image, audio, video, speech, music, embedding, or other non-text output.

This distinction matters for product planning. The model could understand multimodal evidence, but it was primarily a text-and-code reasoning system rather than a media-generation model. An application needing generated images, audio, or video would require a different model or a separate specialized service.

Reasoning, coding, and agent workflows

Claude Mythos Preview supported extended reasoning features. Manual extended thinking allowed a developer to provide an explicit thinking-token budget. Adaptive thinking allowed the model to determine how much reasoning to use for a task. Thinking tokens were counted as output tokens for billing purposes, so deeper reasoning could affect both usage and operational cost.

These controls were relevant to tasks where a quick answer was less important than careful multi-step analysis. Examples include tracing a vulnerability through a large codebase, reviewing an attack surface, designing a remediation, or coordinating a sequence of tool calls. They could also help with autonomous coding workflows in which the model must inspect files, reason about dependencies, propose changes, and continue working through a larger task.

The model’s coding capabilities were especially relevant to defensive security. It was positioned for vulnerability discovery, code auditing, attack-surface analysis, and autonomous coding rather than merely producing short code snippets. However, the model’s output should still be reviewed by qualified engineers and security professionals. The supplied research does not establish that the model’s security conclusions or generated code were guaranteed to be correct.

Tools, web search, and deployment options

Claude Mythos Preview supported tool use, allowing an application to connect the model to external functions or services. This is important for agentic work because the model can reason about a task and then use application-provided tools to retrieve information or perform controlled operations. Tool support does not mean that the model independently has unrestricted access to systems; the application or platform determines which tools exist and what permissions they have.

Web search support was platform-dependent. Anthropic documentation indicated that web search was supported for this model on the Claude API, Google Cloud, and Microsoft Foundry. It was not available for the model on Amazon Bedrock or Claude Platform on AWS. Teams should therefore verify the exact integration before designing a workflow around live web retrieval.

Streaming API interactions and prompt caching were also supported. Streaming can allow partial output to reach a user or application before the complete response is finished. Prompt caching can be useful when the same large instructions or reference material are reused across requests, although the supplied research does not provide a standalone price or savings figure for this model.

Pricing and cost considerations

A confirmed standalone public price for Claude Mythos Preview was not available in the supplied Anthropic documentation. Its pricing may have depended on the research-preview arrangement, contract terms, or the cloud platform used. Organizations should not automatically apply Claude Mythos 5 pricing to the preview.

The absence of a verified price is itself a practical limitation. A one-million-token context window and long reasoning runs can be valuable for difficult technical work, but they may also increase resource use and latency compared with smaller, faster models. Because the preview’s cost structure was not confirmed, a team cannot reliably estimate total operating cost from the available public information alone.

In general terms, the model was aimed at capability-intensive workloads rather than speed-sensitive, low-cost interactions. A smaller generally available model may be more appropriate for routine classification, short summaries, simple code transformations, or high-volume user requests. That is a trade-off based on workload requirements, not a provider-published benchmark comparison.

Main strengths and limitations

Where Claude Mythos Preview was strongest

• Large-scale analysis: The one-million-token context window supported very large codebases, document sets, and extended working contexts.
• Security-focused workflows: Its Project Glasswing positioning emphasized defensive cybersecurity, vulnerability discovery, and attack-surface auditing.
• Complex reasoning: Manual and adaptive thinking options supported tasks requiring multiple stages of analysis.
• Agentic coding: Tool use, streaming, and coding capability supported longer-running software and security workflows.
• Visual understanding: Image input allowed visual evidence to be considered together with text.

Where it was limited

• Restricted access: It was available only to approved Project Glasswing organizations.
• Deprecated status: Anthropic deprecated the model on June 9, 2026, so it should not be treated as a new long-term deployment target.
• Unconfirmed pricing: No reliable standalone public price was supplied.
• Text-only output: It did not natively generate images, audio, video, speech, or music.
• Platform variation: Web search and other features could differ between the Claude API and cloud integrations.
• Operational cost: Large contexts and extended thinking may increase latency and usage compared with smaller models.

When to choose this model

For new work, Claude Mythos Preview is mainly relevant when evaluating an existing historical Project Glasswing deployment or maintaining compatibility with an older integration. Its characteristics would have made it a strong fit for approved organizations conducting defensive cybersecurity research, reviewing very large codebases, analyzing extensive technical documentation, or building agents that needed long-running reasoning and tool use.

It would be a poor choice for a new project that needs public self-service access, a stable supported model, transparent pricing, or a general-purpose production endpoint. In those situations, the documented successor Claude Mythos 5 is the more appropriate starting point, subject to verifying its current availability and behavior.

A smaller or faster model may also be preferable when the task involves short responses, high request volume, simple coding assistance, or strict latency and budget targets. Conversely, a specialized non-text generation model would be more appropriate for applications that must create images, audio, or video. Claude Mythos Preview’s value was concentrated in large-context text reasoning, code, visual understanding, and controlled tool-driven workflows.

Status and migration guidance

Anthropic deprecated Claude Mythos Preview on June 9, 2026 and identified Claude Mythos 5 as its recommended successor. The Messages API and tool-use concepts are largely compatible, but migration is not necessarily a drop-in replacement for every application. Developers should review thinking parameters, output behavior, platform-specific tool availability, and any assumptions about the preview model’s access or pricing.

Because the preview was gated and is now deprecated, its specifications are most useful for understanding an earlier Anthropic research offering and for planning a migration. Teams should use currently supported Anthropic documentation to confirm the replacement model’s limits and commercial terms before moving a production workflow.