Claude Mythos 5

What is Claude Mythos 5?

Claude Mythos 5 is Anthropic’s specialized Claude model for advanced research in areas where technical complexity and potential risk are both high. Its stated focus includes cybersecurity, vulnerability research, biology, healthcare, biodefense screening, and broader life-sciences work. Rather than targeting everyday conversation or inexpensive bulk generation, the model is designed for investigations that may require very large documents, multiple research steps, external tools, and substantial reasoning.

Anthropic released Mythos 5 on June 9, 2026. The model is listed as active but invite-only. Access is provided through Project Glasswing and other trusted-access or approved partner programs, so having an Anthropic account or access to a standard Claude product does not necessarily provide access to Mythos 5.

In Anthropic’s current lineup, Mythos 5 is positioned separately from the more broadly available Claude Fable line. The distinction is practical: Mythos 5 is intended for vetted organizations conducting high-risk or specialized research, while a generally available model is usually a better fit for ordinary writing, coding, summarization, and routine business workloads.

Specifications and supported modalities

The model’s primary identifier is claude-mythos-5. On Amazon Bedrock, a compatible hosted identifier is anthropic.claude-mythos-5. The verified specifications supplied for the model are:

A token is a unit of text used for processing, and the context window is the amount of information the model can consider across an interaction. A 1-million-token window can accommodate unusually large collections of documents, long technical records, or extended tool-assisted investigations, although the practical limit also depends on how an application assembles prompts and manages outputs.

Mythos 5 accepts images as well as text, which supports visual analysis alongside written research material. It does not natively generate images, audio, video, speech, music, or embeddings. Its direct output is text, including ordinary prose, code, and structured text responses. Image understanding should therefore not be confused with image generation.

Reasoning, coding, and tool support

Adaptive thinking is always enabled for Mythos 5, with a high default effort setting. In practical terms, the model is configured to spend more effort on difficult problems rather than treating every request as a short conversational exchange. This is useful for multi-step analysis, code review, vulnerability investigation, experimental planning, and technical synthesis. It can also increase latency and cost compared with a faster, less expensive model.

The supplied research describes Mythos 5 as supporting tool use, streaming, structured outputs, prompt caching, and batch processing. Tool use allows an application to connect the model to approved functions or external systems, while structured outputs help return information in a predictable machine-readable shape. These features are relevant to research pipelines that need to extract findings, classify evidence, call analysis tools, or pass results to downstream software.

Coding is one of the model’s intended strengths, especially when programming is part of a larger technical investigation. Examples include examining source code for vulnerabilities, creating analysis scripts, interpreting technical logs, and working through long codebases or documentation sets. The research rates its reasoning and coding capability highly, but those ratings are editorial evaluations rather than Anthropic-published benchmark scores. They should be treated as guidance about positioning, not as a standardized performance result.

Amazon Bedrock documents response streaming, implicit and explicit prompt caching, and access through its Bedrock Mantle endpoint. Platform-specific behavior can differ, so developers should verify the exact deployment documentation before depending on a particular parameter, sampling behavior, or tool configuration.

Pricing and usage economics

Anthropic’s supplied pricing is based on tokens rather than a fixed consumer subscription. Standard input pricing is $10 per million input tokens, and standard output pricing is $50 per million output tokens.

Prompt caching changes the economics for applications that repeatedly reuse the same instructions or reference material. Five-minute cache writes cost $12.50 per million tokens, while one-hour cache writes cost $20 per million tokens. Cache reads cost $1 per million tokens. The Batch API offers a 50% discount on input and output pricing, making it more suitable for work that does not require immediate responses.

These rates make Mythos 5 substantially more expensive than a lower-cost model for routine requests. They can nevertheless be reasonable when a task would otherwise require dividing a very large research corpus across many calls, using several separate analysis stages, or manually coordinating a complex investigation. Caching is particularly relevant when the same large background corpus is reused across multiple prompts.

Cost planning should include both input and output. A long context does not mean that using the entire 1-million-token capacity is free, and the 128,000-token output ceiling is a maximum rather than a recommendation. Applications should request only the output length needed for the job, use caching when repeated context justifies it, and consider batch processing for offline workloads.

What Claude Mythos 5 is best for

Mythos 5 is best suited to organizations that have an approved access path and a problem where long context and intensive reasoning provide meaningful value. Strong use cases include:

• Cybersecurity research: analyzing vulnerability reports, source code, exploit-related evidence, defensive controls, and large collections of technical logs.
• Biology and life-sciences research: synthesizing extensive technical material, comparing findings, and supporting research workflows that involve long documents and specialized evidence.
• Healthcare research: reviewing and organizing complex research materials where appropriate governance, privacy controls, and human oversight are in place.
• Long-running technical investigations: maintaining a large working context while the model reasons through multiple stages of a problem.
• Tool-using research agents: combining model reasoning with approved functions, retrieval systems, analysis utilities, or other application tools.
• Structured extraction: returning findings in a predictable format for software systems or later review.

The model’s capabilities do not remove the need for expert review. In cybersecurity and life sciences especially, a plausible explanation can still be incomplete or wrong. Results should be checked against authoritative evidence, tested in controlled environments, and reviewed under the organization’s safety and compliance procedures.

Limitations and access constraints

The most important limitation is availability. Mythos 5 is not a generally accessible public model. Anthropic, AWS, or Google Cloud account teams may need to approve access, and eligibility is tied to trusted-access programs and organizational use cases. Amazon Bedrock identifies the model as a Preview Beta Service, so cloud availability and behavior may change.

The model also has a stated knowledge cutoff of January 2026. Its large context window can help it analyze current documents supplied by the user or retrieved through connected tools, but the cutoff does not make the model independently current. Time-sensitive claims still require verification.

Mythos 5 produces text only. It can inspect image inputs, but it is not a native image, video, audio, speech, or music generator. Teams seeking those output types need a different specialized system or a separate processing stage.

Premium pricing is another constraint. High reasoning effort, large prompts, and long outputs can make costs rise quickly. The model is therefore a poor choice for high-volume classification, simple rewriting, basic chat, or other workloads where a faster and cheaper model can meet the quality requirement. Its restricted-access status can also complicate prototyping if an organization has not yet completed the approval process.

Anthropic states that retirement will not occur sooner than June 9, 2027, but no exact shutdown date is published. This should be considered when designing long-lived production systems, particularly because the model is also described as preview or beta on at least one hosted platform.

When to choose Claude Mythos 5

Choose Claude Mythos 5 when the work is technically demanding, the input material is unusually large, and the organization can justify premium pricing and trusted-access requirements. It is a strong candidate for a vetted cybersecurity or life-sciences research team that needs to combine extensive context, adaptive reasoning, image understanding, tools, and structured responses in one workflow.

Choose another option when speed, low cost, broad availability, or simple consumer usability matters more than maximum context and research depth. A generally available Claude model may be more appropriate for ordinary writing, customer support, routine coding, and everyday analysis. A specialized non-text system may be more appropriate when the required result is an image, video, audio file, speech recording, or embedding rather than text.

The central trade-off is capability versus operational efficiency. Mythos 5 provides a very large context window and a configuration intended for difficult investigations, but it costs more, may respond more slowly than lightweight alternatives, and cannot be used freely by every developer. For the narrow audience it targets, those costs may be justified; for routine workloads, they usually are not.

Bottom line

Claude Mythos 5 is an invite-only Anthropic model built for high-risk and technically demanding research rather than general-purpose chat. Its defining specifications are a 1-million-token context window, 128,000-token maximum output, adaptive reasoning, text and image input, text output, tool support, structured outputs, caching, and batch processing. The model’s premium token pricing, January 2026 knowledge cutoff, restricted access, and lack of native media generation limit its usefulness outside specialized research programs. For approved organizations working on complex cybersecurity, biology, healthcare, or life-sciences problems, it offers a concentrated set of capabilities aimed at long-running technical analysis.