Claude Mythos Preview

Overview

Freshness note: Model capabilities, limits, pricing, and availability can change quickly. This profile is a point-in-time snapshot last verified on July 6, 2026.

Claude Mythos Preview is Anthropic’s gated research-preview model announced on April 7, 2026. Anthropic describes it as a general-purpose frontier model and, more specifically, as its most capable model yet for coding and agentic tasks. The launch matters because Anthropic is not presenting Mythos as just another public Claude plan option. It introduced the model together with Project Glasswing, an early-access initiative focused on securing critical software with frontier AI.

That distinction is important. Project Glasswing is not a separate model or a replacement product surface. It is the deployment and evaluation initiative built around Mythos access, giving selected partners and additional critical-software organizations access for defensive work. Anthropic’s current model docs now list Claude Mythos 5 as a separate limited-availability route that shares Claude Fable 5 specs and pricing, while Mythos Preview remains invitation-only with no self-serve signup.

Capabilities

Anthropic’s public framing is unusually direct: Mythos Preview is a general-purpose model whose most striking current strength is cybersecurity. The official Glasswing materials say the model has already found thousands of high-severity vulnerabilities across critical infrastructure, while the RED-team writeup shows it identifying and exploiting difficult bugs in browsers, operating systems, and major open-source targets.

Anthropic also emphasizes that these cyber capabilities are downstream effects of broader improvements in code, reasoning, and autonomy. That is why the model is relevant beyond security specialists. The same characteristics that make it unusually capable at vulnerability discovery also make it important for autonomous coding, long-running agent loops, and hard software-engineering tasks over large codebases.

Anthropic’s June Project Glasswing expansion also says partners are using Mythos Preview for patch writing, pre-release vulnerability checks, penetration testing, threat detection and response, and memory-safe rewrites of legacy codebases. In other words, Anthropic is positioning Mythos as a frontier coding-and-agency step change whose current release path happens to be security-first.

Technical Details

The current official Claude model overview and Project Glasswing materials list these practical anchors:

• 1M token context window
• 128K synchronous max output
• text and image input with text output
• adaptive thinking support

Anthropic now also documents claude-mythos-5 as a newer Mythos model ID with the same published specs as Claude Fable 5. Treat Mythos Preview and Mythos 5 as related but separate routes, especially when checking pricing, account eligibility, and cloud-provider access.

Anthropic’s own public announcement materials focus more on demonstrated behavior than on a polished standalone spec sheet. The RED-team writeup says Mythos Preview performs strongly across the board, but is especially capable at computer-security tasks. It also says many of the observed exploit-development abilities were not explicitly trained as a separate objective, but emerged from broader gains in code understanding, reasoning, and autonomy.

Operationally, the important detail is not just “bigger Claude.” It is that Anthropic is treating Mythos as a model that can sustain coherent work over long horizons, act more autonomously on difficult coding problems, and cross a threshold where safety controls and access restrictions become central deployment concerns.

Pricing & Access

Anthropic’s current official position is deliberately narrow:

• gated research preview only
• access prioritized for defensive cybersecurity use cases
• available to Project Glasswing participants and additional selected critical-software organizations
• invitation-only, with no self-serve signup
• accessible through the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry

Anthropic’s project page lists pricing for participants at:

• Input: $25 per 1M tokens
• Output: $125 per 1M tokens

Anthropic now also documents Claude Mythos 5 at $10 input and$50 output per 1M tokens with limited Project Glasswing availability. Do not assume Mythos Preview and Mythos 5 share the same commercial terms; verify the approved route with Anthropic or the relevant cloud account team.

Anthropic also says it has committed up to $100M in usage credits for Project Glasswing and related participants during the research preview. The June 2026 expansion post points toward wider Mythos-level cyberdefense access over time, including a Cyber Verification Program, but teams should not treat this as a normal self-serve model purchase.

Best Use Cases

Mythos Preview is best understood as a controlled frontier model for defensive security, autonomous software work, and long-running agents over large codebases or critical systems. It fits environments where the organization can justify a gated-research posture, strong internal review, and careful use boundaries.

It is not the right default for everyday assistant work, broad public product rollout, or casual experimentation. Anthropic’s own release posture makes that clear: this is a restricted model for organizations that can absorb both its capability upside and its safety burden.

Comparisons

• Claude Opus 4.8 (Anthropic): Opus is the public premium Claude route for hard coding and agentic work; Mythos Preview is the more restricted security-first research tier above that public surface.
• Claude Mythos 5 (Anthropic): Newer limited-availability Mythos route with Fable 5 specs and pricing for approved Project Glasswing customers.
• GPT-5.5 (OpenAI): GPT-5.5 remains the more broadly available frontier generalist, while Mythos Preview is positioned for narrower, higher-risk, cyber-heavy use.
• Gemini 3.5 Flash (Google): Gemini is the more conventional broadly distributed route for multimodal and agentic workflows; Mythos Preview differentiates through restricted access and unusually strong defensive-security framing.