What is Claude Mythos 5.1?
Claude Mythos 5.1 is Anthropic’s Mythos-class model for demanding research, cybersecurity, biology, software engineering, and long-running agentic work. An agentic workflow is one in which a model can plan multiple steps, use tools, inspect results, and continue toward a larger objective rather than simply answering one prompt.
Anthropic released the model on September 1, 2026. The provider describes it as sharing the core specifications of Claude Fable 5.1 while using a different safeguard configuration intended for vetted cybersecurity and life-sciences users. That distinction is important: Mythos 5.1 is not simply a public Claude tier with a different name. Its access model and governance requirements are part of its practical identity.
The canonical Claude API model identifier is claude-mythos-5-1. Equivalent provider-specific identifiers are available through Amazon Bedrock, Google Cloud, and Microsoft Foundry, subject to the relevant access arrangements.
Access and position in Anthropic’s lineup
Claude Mythos 5.1 is invitation-only. Organizations generally need approval through Anthropic, AWS, or Google Cloud account teams, or through applicable trusted-access programs. Project Glasswing is one of the named routes for access.
This restricted availability places Mythos 5.1 at the specialized, high-capability end of Anthropic’s lineup. It is intended for approved organizations whose work justifies the model’s capability, cost, and additional governance considerations. It is not positioned as a general replacement for broadly available Claude models used for ordinary chat, routine summarization, or high-volume production requests.
Anthropic’s stated rationale for the specialized configuration relates to the model’s intended work in areas such as cybersecurity and biology, where advanced reasoning and coding can be valuable but can also create elevated misuse risks. Organizations should therefore evaluate access controls, monitoring, retention requirements, and use-case approval alongside technical performance.
Technical specifications at a glance
| Specification | Claude Mythos 5.1 |
|---|---|
| Provider | Anthropic |
| Release date | September 1, 2026 |
| Model ID | claude-mythos-5-1 |
| Context window | 1,000,000 tokens |
| Maximum output | 128,000 tokens |
| Knowledge cutoff | June 2026 |
| Input | Text and images |
| Output | Text |
| Reasoning | Adaptive thinking, always enabled |
| Default effort | High |
| Comparative speed | Slower than faster Claude tiers |
| Availability | Invitation-only |
The 1-million-token context window allows an application to provide very large bodies of material in one interaction, such as extensive technical documentation, source code, research records, or a long-running project history. The context limit is not the same as the maximum response size: Mythos 5.1 can accept a very large working context while generating up to 128,000 output tokens.
The listed knowledge cutoff is June 2026. External web search or other tools may provide newer information during use, but that does not change the model’s underlying training-data cutoff.
Reasoning and coding capabilities
Mythos 5.1 uses adaptive thinking that is always enabled. In practical terms, the model can spend more processing effort on complex, multistep problems instead of treating every request as a short direct-answer task. Anthropic lists the default effort as high, and effort controls can influence the depth and cost of processing.
The model is intended for work where planning, analysis, and verification matter. Examples include investigating a complex security issue, examining a large technical codebase, coordinating a sequence of research tasks, or producing a detailed scientific or engineering report. The model’s slower response profile is a trade-off for this type of extended reasoning.
Coding is a primary use case. Mythos 5.1 can help analyze unfamiliar repositories, propose or implement changes, explain dependencies, review code, and coordinate tool-assisted development workflows. The supplied research does not provide benchmark results, so claims about coding quality should be understood as capability positioning rather than a quantified performance guarantee.
Input, output, and tool support
Claude Mythos 5.1 accepts text and image input and produces text output. Image input can be useful for interpreting diagrams, screenshots, charts, or other visual material alongside written instructions. It does not mean that the model generates images. The model does not natively produce image, audio, or video output.
The model supports tool use and is designed for workflows that combine reasoning with external actions or information. The research identifies support for code execution, programmatic tool calling, memory, compaction, context editing, streaming-style API workflows, prompt caching, and batch processing for the Mythos and Fable 5.1 generation.
There is an important tool-selection limitation: Anthropic’s guidance says forced tool selection using tool_choice values such as any or tool is not supported for Mythos 5.1. Applications should use automatic tool choice and validate tool inputs and outputs carefully rather than assuming that a particular tool can always be forced.
Thinking blocks may be preserved for compatible conversations, but older models may not be able to interpret them. Systems that replay conversations across model generations should account for this compatibility issue.
Pricing and cost controls
Anthropic’s listed standard pricing is $10 per million input tokens and $50 per million output tokens. Output tokens are priced substantially higher than input tokens, so applications that generate very long responses or extended agent traces should estimate output usage carefully.
Prompt caching is available. The listed cache-write prices are $12.50 per million tokens for five-minute writes and $20 per million tokens for one-hour writes. Cache reads are listed at $0.25 per million tokens. Caching can be relevant when an application repeatedly sends the same large instructions, documents, or project context.
The Batch API provides a 50% discount on standard input and output pricing. Actual charges can vary according to the platform, regional inference settings, caching behavior, batch processing, and provider-specific commercial terms. The figures above are token prices, not a guarantee of a fixed project cost.
Main strengths and limitations
Mythos 5.1’s clearest strengths are its unusually large context window, long maximum output, adaptive reasoning, and suitability for tool-assisted work. These features make it a candidate for tasks that are too large, interconnected, or long-running for a smaller and faster model.
- Large working context: The 1-million-token window supports extensive documentation, code, and research material.
- Extended responses: The 128,000-token output limit accommodates detailed analyses, reports, and multi-step results.
- Deep reasoning: Always-on adaptive thinking is designed for complex planning and investigation.
- Technical focus: Cybersecurity, biology, scientific research, coding, and agentic workflows are central intended uses.
- Multimodal understanding: The model can combine image input with text-based analysis.
- Tool-oriented workflows: Code execution, programmatic tool calling, memory, compaction, and context editing support longer processes.
The same characteristics also create limitations. The model is slower than faster Claude tiers and costs more than lower-tier options. Its restricted access can be a larger obstacle than its API features for organizations that have not been approved. The model produces text only, so it is not suitable when native image, audio, or video generation is required.
Its specialized safeguard configuration also means that governance is not optional. Organizations should confirm that their monitoring, authorization, data-handling, and review processes meet the requirements attached to access. The research does not specify every retention or monitoring term, so those details should be confirmed with Anthropic or the applicable cloud provider.
When to choose Claude Mythos 5.1
Choose Mythos 5.1 when the work is high-value, technically difficult, and benefits from long context or extended reasoning. Good fits include vetted cybersecurity defense, advanced biology and life-sciences research, complex software engineering, technical investigations, large document analysis, and long-running agents that need to maintain project context while using tools.
It is particularly appropriate when quality, depth, and autonomy matter more than minimum latency or the lowest token cost. A security team investigating a complicated incident, for example, may value the ability to combine extensive logs, code, documentation, and tool results in one workflow.
Another generally available Claude model may be more appropriate for everyday chat, routine summarization, simple coding assistance, latency-sensitive applications, or high-volume workloads where the additional cost and slower processing are difficult to justify. A model with image, audio, or video generation would also be a better choice for applications that need those output types. Mythos 5.1 should be selected for its specialized reasoning and research role, not merely because it has a large context limit.
Practical integration guidance
Applications using the Claude Messages API should specify claude-mythos-5-1 and should treat the model as an always-thinking system. Integrations should not attempt to disable adaptive thinking or send legacy manual thinking budgets intended for older Claude models.
Because the model can generate long outputs and operate in extended tool workflows, applications should set appropriate operational budgets, validate tool results, and handle streaming responses where useful. Prompt caching and batch processing can reduce costs in workloads with repeated context or non-urgent jobs, but their economics depend on the application’s actual token pattern.
Overall, Claude Mythos 5.1 is a specialized Anthropic model for approved organizations that need advanced reasoning across large technical or scientific workloads. Its million-token context, long output limit, multimodal input, and agent-oriented features are valuable differentiators, while invitation-only access, slower processing, high output pricing, text-only output, and governance requirements define when it is practical to deploy.