Mouse-1

Mouse-1 is a decision model and fast decision layer designed for AI agent harnesses. It is built for developers and teams who run agents that make many decisions during a task. Instead of sending every small decision to a large frontier model, the harness can plug Mouse-1 into its current stack and use it to resolve those decisions. The product takes the agent's state and the options available, returns a probability, and lets the harness route, gate, or verify while the main frontier model handles reasoning or code. The website presents Mouse-1 as something you plug directly into your agent harness and says users can see measurable results right away. An agent makes hundreds of small decisions during every task. These decisions are not the main reasoning or coding work, but they still need to be handled. Using a frontier model for all of them can consume time and money. Mouse-1 is built to handle these decisions so the agent does not have to spend frontier-model capacity on them. The website frames this as a decision layer that saves time and money, and Product Hunt metadata describes it as a decision model for an agent harness. The core problem is straightforward: agents need many small decisions resolved, and a fast, low-cost decision model can take that work while the main model focuses on the tasks that require deeper reasoning. Mouse-1 works by taking the agent's state and options and returning a probability. That probability lets the harness handle small decisions while the main frontier model takes care of reasoning or code. The harness is the part of the agent stack that can route, gate, and verify; Mouse-1 supplies the decision signal. The website's side-by-side diagram shows Mouse-1 at roughly 100 ms for Decide and labels the harness steps as Route, gate, verify. The main model is labeled Plan & write code, and tools are labeled Shell, edit, MCP. The sequence shows Propose, Execute, and Results. In this arrangement, Mouse-1 does not replace the frontier model; it adds a fast decision layer alongside the current stack. The website emphasizes speed. It says Mouse-1 answers in about 100 ms, and one request can answer up to 128 questions. It also says this is faster than a chat model's first token. In a side-by-side example, the same prompt is sent to Mouse-1 and a frontier model. The website says Mouse-1 finishes the task before the frontier model returns its first token. This is a strong illustration of the difference between a decision model and a chat model: the purpose is not to generate prose or code, but to return a choice quickly from the provided options. The ~100 ms figure and the up-to-128-questions-per-request detail are the key performance facts stated on the page. The page shows three image-based examples. In a content moderation example, an image of hot dogs is shown with the label hot_dog >99.99% and a moderation question: Is this image safe to post? The returned options and probabilities are safe 99.96%, spam 0.03%, nudity <0.01%, and violence <0.01%. In a document routing example, a 2020 Form 1040-ES is shown with the question Where does this document go? and the options tax_return 99.99%, invoice 0.01%, contract <0.01%, and id_document <0.01%. In a browser use example, Chrome's offline page is shown with the question What state is the screen in? and the options error 99.95%, loading 0.03%, ready 0.02%, and login_required <0.01%. These examples show Mouse-1 returning probability distributions over explicit options. The benchmarks section reports average decision accuracy on unfamiliar tools and repositories. Mouse-1 is listed at 89.1%, followed by d3 at 85.8%, Clef at 84.2%, OpenAI at 82.5%, d3-flash at 81.4%, and Jev at 79.1%. The page also reports a decision flip rate, described as decisions changed by a planted instruction, where lower is better. Mouse-1 is listed at 0%, followed by d3-flash at 27.6%, Clef at 37.4%, d3 at 40.4%, OpenAI at 53.9%, and Jev at 65.4%. The page says the 0% result supports the claim that Mouse-1 is hard to inject. These benchmarks are presented as evidence that a dedicated decision model can perform well on small decisions. The page lists zero data retention as a feature. It says Mouse-1 scores the request and drops it, and nothing is stored or used for training. It also says Mouse-1 is hard to inject because it only picks from the options given; planted instructions changed 0% of decisions in the tests. This matters for agent harnesses that process user content, documents, or browser screens, because the decision model is not asked to generate free-form output. It chooses among options. The combination of zero retention and option-only selection is presented as a security and privacy advantage. Pricing is listed as cost per 1,000 decisions. Mouse-1 is $0.10, with input at $0.10 per 1M tokens and output free. The page says this is up to 65× cheaper than frontier models. For comparison, the table lists GPT-6 Sol at $6.50 per 1,000 decisions, Claude Opus 5 at $6.25, Claude Sonnet 5 at $3.75, Gemini 3.7 Flash at $1.88, DeepSeek V4 at $1.52, Claude Haiku 4.5 at $1.25, Cloudflare Clef at $0.24, OpenAI Decisions at $0.10, and TypeSafe Jev at $0.042. The table notes 1,000 input tokens per decision and says generative models add 50 output tokens. Mouse-1 is positioned as a low-cost decision layer rather than a general-purpose generative model. The benefits stated on the page are time and money savings, fast decisions, low latency, low cost, zero data retention, and resistance to planted instructions. The use cases shown are content moderation, document routing, and browser use state detection, plus the general harness workflow of routing, gating, and verifying small decisions. The target users are developers and teams building agent harnesses, especially those already using a frontier model and tools such as shell, edit, and MCP. Mouse-1 is added alongside the current stack rather than replacing it. The page invites users to start building in the console and read the documentation, and it offers a copyable agent prompt. Mouse-1's primary value proposition is a fast, cheap decision layer for agent harnesses. It handles the hundreds of small decisions an agent makes during a task, returns probabilities over the options it is given, and leaves reasoning and code to the main frontier model. The website supports this with ~100 ms latency, up to 128 questions per request, 89.1% average decision accuracy on unfamiliar tools and repositories, a 0% decision flip rate in planted-instruction tests, zero data retention, and a $0.10 per 1,000 decisions price. For teams building agents, Mouse-1 is presented as a way to keep the frontier model focused while the harness gets fast, measurable decisions.