The Mouse-1 decision model never writes a sentence. You give it your AI agent’s situation and a short list of options, and in about 100 milliseconds it returns a probability for each one. Its maker says that is enough to take the hundreds of small “what next?” calls off the expensive frontier model running your agent.
Mouse-1 launched on 20 October 2026 and placed tenth on Product Hunt’s daily board on 21 October. The pitch is appealing, so here is how it works, what it costs and where its own numbers fall short.
Key takeaways
- Mouse-1 is a scoring model. It picks only from options you supply and returns probabilities, not text.
- Pricing is $0.10 per million input tokens, and output is free. Mouse puts that at about $0.10 per 1,000 decisions.
- Mouse says it scored 89.1% on unfamiliar tools and repositories and 0% on a prompt-injection flip test.
- On a mixed tool-and-security suite it came third, behind d3 and Clef.
- Every benchmark so far is the vendor’s own. Nobody has checked them independently yet.
What is the Mouse-1 decision model?
First, a definition. An agent harness is the software wrapped around an AI model that turns it into an agent. In the words of Modular’s handbook, it keeps conversation state, exposes tools, runs approved actions and decides whether to prompt the model again or stop.
That last job adds up to a lot of small judgement calls. Is this page still loading? Should I call a tool or answer directly? Is this file an invoice or a contract? Most harnesses send every one of those questions to the same large model that writes the code. That is slow, and you pay for every token.
According to Mouse’s API page, Mouse-1 sits next to that model as a “fast decision layer”. Your harness sends the state and the candidate options. Mouse-1 returns a score for each option, and your code decides what to do with them, including which low-confidence calls go to a human.
Think of a railway switch operator. They don’t drive the train or plan the route. They just flip the lever the right way, quickly, many times a day. Mouse-1 is the switch operator, and the frontier model is the driver.
How does it work in practice?
One request can carry up to 128 questions. The examples Mouse shows are refreshingly ordinary:
- Content moderation: is this image safe, spam, nudity or violence?
- Document routing: tax return, invoice, contract or ID?
- Browser agents: is the screen showing an error, still loading, ready or asking for a login?
Image input is in beta. As Tech Bytes noted in its 20 October write-up, Mouse has tested images on only a handful of cases and has not published an image benchmark.
You can also tune it. Send examples from your own harness with the correct answers, and Mouse trains a tuned version, which it tests against the base model before releasing it. Calibration then sets a confidence cutoff that sends uncertain decisions for human review. On Product Hunt, maker Pete McGrath said LoRA fine-tuning (a cheap way to adapt a model by training a small add-on) is passed through from the GPU provider “with zero mark up”.
On privacy, Mouse says requests are scored and then discarded, with nothing stored or used for training unless you turn on its optional Outcomes feature.
How much does Mouse-1 cost compared with frontier models?
Mouse prices each model per 1,000 decisions, assuming 1,000 input tokens per decision and, for chat-style models, 50 output tokens. These are the vendor’s own figures from its launch page:
| Model | Cost per 1,000 decisions |
|---|---|
| Jev | $0.042 |
| Mouse-1 | $0.10 |
| OpenAI Decisions | $0.10 |
| Clef | $0.24 |
| Claude Haiku 4.5 | $1.25 |
| Claude Sonnet 5 | $3.75 |
| Claude Opus 5 | $6.25 |
| GPT-6 Sol | $6.50 |
Two things stand out. Mouse-1 is not the cheapest dedicated decision model: Jev costs less than half as much, and OpenAI Decisions costs the same. And the headline multipliers don’t agree with each other. The Product Hunt post says “163x cheaper, and 65x faster than GPT-6 Sol”, while the API page and Tech Bytes put it the other way round: about 65x cheaper and 163x faster. The table supports the 65x cost figure.
New API keys come with $5 of credit and need no card, which at these rates covers tens of thousands of test decisions.
Do the benchmarks hold up?
Partly. On Mouse’s own test of unfamiliar tools and repositories, Mouse-1 averaged 89.1%, ahead of d3 (85.8%), Clef (84.2%) and OpenAI Decisions (82.5%). Its best result is the prompt-injection test, in which planted instructions try to change a decision. Mouse-1 changed 0% of its decisions, while OpenAI’s changed 53.9%.
The broader Agent Tool & Security Suite, which averages BFCL, API-Bank, When2Call and the phishing test PhishNChips, shows a different picture. There d3 scored 86.0%, Clef 85.6% and Mouse-1 84.9%. On phishing detection Mouse-1 managed 72.3%, behind Clef’s 79.4%.
My take: publishing a suite on which you come third is a point in Mouse’s favour. Still, every number here comes from the company’s own tests. Before you trust it with real traffic, run it against your own cases. Our 7-point AI agent evaluation checklist is a good place to start.
Who should actually try it?
It suits you if your agent makes many repetitive, closed-choice decisions, such as classifying, routing or checking state. If your bottleneck is reasoning quality, a scorer won’t help.
Speed may matter more than price. One Product Hunt commenter argued that for a router, the delay before the first word reaches the user is what people notice. McGrath separately said that across 300+ test runs, harnesses using Mouse-1 averaged 3x faster and 15x cheaper than those without it.
It also fits a wider trend of non-generative helper models, such as Cloudflare’s Clairic extractor. If you are working out where a component like this belongs, see our maturity model for production-grade AI agents.
What happens next
Mouse has not published a roadmap. The near-term steps it names are finishing the image-input beta and checking customers’ tuned models before release.
The real test is outside scrutiny. If independent developers reproduce the 0% injection flip rate and the 100 ms latency, Mouse-1 becomes an easy default for agent routing. If they don’t, it joins a crowded field that already includes Jev, Clef, d3 and OpenAI’s own decision model.
FAQ
Does Mouse-1 replace my main AI model?
No. It only scores options you provide. Your frontier model still does the reasoning, planning and code writing.
How much does Mouse-1 cost?
$0.10 per million input tokens, with free output. Mouse estimates about $0.10 per 1,000 decisions, and new keys get $5 of credit.
Are the Mouse-1 benchmarks independent?
No. All the accuracy, speed, cost and injection figures published so far come from Mouse’s own tests.
Next step: take 50 real decisions from your agent’s logs, run them through Mouse-1 on the free credit, and compare the results with what your current model chose.
Leave a comment