Mouse-1 compared

Side-by-side pages for the decision models and judges people weigh against Mouse-1. Where we ran the competitor ourselves, the numbers come from the same suites; where we didn't, the page says so.

Mouse-1 vs AnyJev L0 AnyJev turns stock Gemma 4 31B into a decision model with no training at all. It scores well on public tool benchmarks; its probabilities are poorly calibrated, and Mouse-1 leads on When2Call.
Mouse-1 vs Claude Haiku 5.5 as an agent judge Haiku 5.5 costs the same per input token, so this is a race on time. The same agent finished faster with Mouse-1 answering its decisions, and spent less than half as long waiting on them.
Mouse-1 vs Claude Opus 5.5 as an agent judge Using a frontier model to judge an agent's every step is thorough and slow. In five head-to-head runs, the same agent finished 4.7× faster and about 7× cheaper per task with Mouse-1 answering its decisions.
Mouse-1 vs Cloudflare Clef Mouse-1 is more accurate on agent decisions, ignores planted instructions and costs less than half as much per token. Clef has open weights, longer inputs and slightly higher public tool-benchmark scores.
Mouse-1 vs Darwin-27B-ZTC Mouse-1 has the higher average on the four public tool benchmarks, driven by phishing detection. Darwin is far ahead on When2Call and ships open weights.
Mouse-1 vs Decider 31B Decider 31B scores higher on public tool benchmarks, but you have to host it on a Blackwell GPU. Mouse-1 is an API call, and it is better at knowing when to call a tool at all.
Mouse-1 vs Fastino GLiDE Both are hosted System One–style APIs. Mouse-1 costs a third less, keeps nothing by default and has the higher average on public tool benchmarks; GLiDE reads longer inputs.
Mouse-1 vs GPT-6 Sol for agent decisions Sol is a frontier model; Mouse-1 answers one kind of question. For the yes/no and pick-one decisions inside an agent loop, Mouse-1 answers many times faster at a twentieth of the input price, with no output to pay for.
Mouse-1 vs Jev (TypeSafe) Same API shape, same primitives. Mouse-1 is more accurate on every agent decision suite we ran and no planted instruction changed its answers; Jev costs less per token.
Mouse-1 vs LiteLLM safeguards classifier LiteLLM lets your proxy answer Claude Code's tool-call checks with any LLM you configure. Mouse answers the same checks with a model built for it, in a fraction of the time.
Mouse-1 vs OpenAI Decisions Same price per token. Mouse-1 was more accurate on every agent decision suite we ran and ignored planted instructions; OpenAI reads far longer inputs and offers EU data residency.
Mouse-1 vs OpenRouter Armor Both judge an agent's tool calls against plain-English rules and answer in Claude Code's format. Armor lives inside OpenRouter and is free in beta; Mouse works with any provider and publishes how its model scores.
Mouse-1 vs Perplexity Decider Perplexity Decider v1.2 is cheaper and scores higher on public tool benchmarks. Mouse-1 is the one with zero data retention by default, calibrated abstention and fine-tuning on your own decisions.
Mouse-1 vs Sieve 27B Sieve 27B v2 has one of the best public tool-benchmark averages on the board, led by near-perfect phishing detection. Mouse-1 scores higher on BFCL and When2Call and runs as an API.
Mouse-1 vs Torchcast Decision 27B The two are within a point on public tool benchmarks. The difference is the licence: Torchcast's weights are non-commercial, and Mouse-1 is built to be used in a product.