Mouse-1

Product Introduction

1. Definition: Mouse-1 is a specialized decision model API designed to function as a lightweight, high-speed classification layer within an AI agent's operational stack. Technically, it is a fine-tuned, small-parameter model optimized for fast, deterministic choice-making from a predefined set of options.
2. Core Value Proposition: It exists to dramatically reduce the cost and latency of the hundreds of micro-decisions an AI agent makes during a complex task, such as routing, verification, and state management. By offloading these decisions from expensive, general-purpose frontier models (like GPT-6 Sol or Claude Opus), it enables faster, more reliable, and significantly cheaper agentic workflows.

Main Features

1. High-Speed Decision Layer: Mouse-1 operates with a sub-100ms average latency, often faster than a frontier model can generate its first token. It is engineered for batch processing, capable of answering up to 128 discrete questions in a single API request, maximizing throughput for agent harnesses.
2. Probability-Based Output: The model does not generate free-form text. Instead, it takes the agent's current state and a structured list of possible actions (options) as input and returns a probability distribution over those options. This deterministic output allows the agent's control logic (the harness) to execute decisions with high confidence.
3. Direct Harness Integration: Mouse-1 is designed to be plugged directly into an existing agent architecture alongside tools and a main LLM. It handles the "small decisions" (e.g., "is this task complete?", "which tool should be used next?", "is this output safe?"), freeing the primary, more expensive model to focus on complex reasoning, planning, and code generation.

Problems Solved

1. Pain Point: The prohibitive cost and slow speed of using large frontier models for every minor decision in an agentic loop. This creates scalability and latency bottlenecks for production AI applications.
2. Target Audience: AI engineers and developers building scalable, cost-effective autonomous agents; companies deploying AI agents for customer support, document processing, or content moderation; and researchers needing a reliable, fast classification component for complex AI systems.
3. Use Cases: Essential for real-time agent routing (e.g., document classification to tax_return or invoice), content moderation (classifying image/text safety), browser/UI state detection (identifying error, loading, or ready screens), and tool selection within an agent's action space.

Unique Advantages

1. Strengths & Limitations (Pros & Cons):

   • Pros: Exceptional cost-efficiency (163x cheaper than GPT-6 Sol per benchmark). Blazing fast inference speed (~100ms). High accuracy on structured decision tasks (89.1% on unfamiliar tools benchmark). Robust against prompt injection (0% decision flip rate in tests). Zero data retention policy enhances privacy.
   • Cons: Not a generative model; cannot create net-new content or perform open-ended reasoning. Limited to selecting from provided options, requiring well-defined decision frameworks. Performance is dependent on the quality and structure of the options list provided by the developer.

2. Key Alternatives & Differentiation:

   • vs. General-Purpose LLMs (GPT-6 Sol, Claude Opus): Mouse-1 is not a replacement but a complement. It is 65x faster and over 160x cheaper for pure decision tasks, but lacks generative capability. The workflow differentiation is clear: use Mouse-1 for fast, cheap decisions within the harness, and a frontier model for heavy lifting.
   • vs. Other Decision APIs (OpenAI Decisions, Cloudflare Clef): Mouse-1 claims superior accuracy (89.1% vs. 82.5% for OpenAI) and significantly higher robustness to adversarial instructions (0% vs. 53.9% flip rate). Its pricing is competitive at $0.10 per 1k decisions, matching OpenAI's base rate.
   • vs. Open-Source Models (Self-hosted): Mouse-1 offers a managed, optimized API endpoint, eliminating the infrastructure overhead, tuning complexity, and latency variability of self-hosting smaller open-source classification models.

Frequently Asked Questions (FAQ)

1. What is Mouse-1 used for in AI agents? Mouse-1 is used as a dedicated decision engine within an AI agent's architecture to handle high-volume, low-level classification tasks like routing, gating, verification, and tool selection, which dramatically improves the agent's speed and reduces operational costs.
2. How does Mouse-1 compare to using GPT-4 or Claude for decisions? Mouse-1 is specifically fine-tuned for fast, deterministic decision-making, making it significantly cheaper (up to 163x) and faster (often sub-100ms vs. seconds) than using large generative models for the same task, while also being more resistant to prompt injection attacks.
3. Is Mouse-1 a generative AI model? No, Mouse-1 is not a generative model. It is a specialized classification model that selects from a predefined list of options you provide. It outputs probabilities, not creative text or code.
4. How is data privacy handled with the Mouse API? According to Mouse, they operate with a zero data retention policy. Requests are scored for the decision and then immediately dropped, not stored or used for model training, addressing key data privacy concerns in enterprise deployments.
5. Can Mouse-1 be used for content moderation? Yes, content moderation is a primary use case. You can structure a request to classify content (image or text) into categories like safe, spam, nudity, or violence, receiving probability scores for each to automate moderation decisions.