Claude Mythos Preview

Preview · Family: Claude

Anthropic's unreleased general-purpose frontier model from the Mythos line, with exceptional vulnerability discovery and agentic-coding skills; access limited (Project Glasswing).

⚠ Deprecated⏳ Limited accessLLMReasoning modelMultimodalTool-using model📁 Claude

Context window

Nieujawnione publicznie

tokens

Parameters

Nieujawnione publicznie

parameters

🏢AnthropicProducer · 2026

Access:APIDeployment:☁ Cloud

Overview

Claude Mythos Preview (API ID: claude-mythos-preview) is Anthropic's unreleased, general-purpose frontier model from the Mythos line. Anthropic disclosed it on April 7, 2026 alongside the Project Glasswing initiative as a model with exceptional abilities to find and exploit software vulnerabilities.

The model has found thousands of high-severity zero-day vulnerabilities, including ones in every major operating system and web browser — some of which had survived decades of human review. On CyberGym (vulnerability reproduction) it scored 83.1% versus 66.6% for Claude Opus 4.6.

Its advanced cyber capabilities stem from strong agentic coding and reasoning: SWE-bench Verified 93.9%, Terminal-Bench 2.0 82.0%, GPQA Diamond 94.6%. The model accepts text and image input and supports agentic tool use and computer use.

Access is restricted: under Project Glasswing the model was made available to founding partners and over 40 organizations that maintain critical infrastructure, and later to verified security professionals. Anthropic does not plan to make the model generally available; after the research period the price for participants is $25 per 1M input tokens and $125 per 1M output tokens. The model is marked deprecated (as of June 9, 2026).

Classification

LLMReasoning modelMultimodalTool-using model

Family: Claude

Access & deployment

API

Cloud

Weights: Closed

Key parameters

📏 Context: Nieujawnione publicznie

🧩 Parameters: Nieujawnione publicznie

✓ Tools

📥 Input: text, image

Platforms

Anthropic Claude API Amazon Bedrock Vertex AI

Technical specification

Context window

Nieujawnione publicznie

tokens

Parameters

Nieujawnione publicznie

parameters

License

Proprietary (Anthropic) — dostęp wyłącznie przez programy weryfikacyjne (Project Glasswing / Cyber Verification Program)

Hardware requirements

Not applicable — API-only access; the model is hosted on Anthropic infrastructure and cloud partners (Amazon Bedrock, Google Vertex AI).

Features:✓ Tool use

Modalities

⬇ Input

textimage

⬆ Output

textcode

Capabilities and applications

Native model capabilities

Reasoning

The model's ability to reason logically and solve complex problems.

Category: reasoning

Multi-step reasoning

Carrying out multi-step chains of reasoning across long, complex tasks.

Category: reasoning

Coding

Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.

Category: coding

Agentic coding

Multi-hour, multi-step programming tasks performed autonomously by the model: cloning a repository, running tests, iterating on fixes, integrating with CLI tools. Characteristic of Codex variants (GPT-5.1-Codex-Mini, Codex-Max).

Category: coding

Cybersecurity

The model ability to perform computer-security tasks: vulnerability analysis, proof-of-concept exploit generation, patching, and cybersecurity question answering.

Category: other

Computer use

The model's ability to operate a computer interface by interpreting screenshots and generating actions such as clicks, typing, and navigating applications.

Category: planning

Tool use

The model's ability to call external functions, APIs and tools during a conversation: calculator, search engine, code editor, database. The model decides when and how to use a tool and interprets its result.

Category: planning

Planning

Forming and executing action plans for complex tasks.

Category: planning

Agentic capability

The model's ability to autonomously plan and execute multi-step tasks by sequentially using tools, maintaining context, and adapting to intermediate results.

Category: planning

Benchmark results

11 benchmarks

CyberGym

accuracy · vulnerability reproduction (vs Claude Opus 4.6 66.6%)

83.1%