Mouse-1 launches a decision model for agent harnesses

Mouse-1 is a launch on Product Hunt that describes itself as a decision model for agent harnesses. According to the listing, an agent makes hundreds of small decisions during a task, and Mouse-1 is built to handle them. It runs alongside the existing stack inside the harness, takes the agent's state and options, and returns a probability, leaving the main frontier model to handle reasoning or code.

The listing claims Mouse-1 is 163 times cheaper and 65 times faster than GPT-6 Sol. A comment from the maker gives different figures, saying that in their evals adding Mouse to an agent harness made the agent about 190 times cheaper and 13 times faster than Opus 5.5, with the same output. These comparisons come from the maker and have not been independently verified.

The maker says the model can be tuned and calibrated for a specific harness and that LoRA fine-tuning is passed through from the H100 provider with no markup. The API is available from a single endpoint and is zero data retention for all users by default. The maker also says Mouse-1 is more accurate and cheaper than Jev.

Developers could move routine per-step calls such as classification and routing to a small model inside the harness, and reserve frontier model calls for reasoning or code. The benchmark numbers are self-reported and their baselines conflict, so it is worth measuring on your own harness before switching. If your agent handles sensitive data, check the zero data retention default and the fine-tuning terms before adopting it.