Today we’re introducing mouse-1, a general purpose decision model specifically built for AI harnesses.
It integrates into your existing harness through a single API and returns decisions in about 100 milliseconds of model time. It’s designed to help agents complete more work independently, with fewer errors and unnecessary interruptions.
On average mouse-1 is 12× faster than GPT-6 Sol.
GPT-6 Sol
We evaluated mouse-1 on unfamiliar tools and repositories to measure how well it generalizes beyond its training data. It achieved 89.1% average decision accuracy, compared with 85.8% for d3 and 81.4% for d3-flash.
Unfamiliar tools and repositories (%) · Higher is better
| mouse-1 | 89.1% |
|---|---|
| d3 | 85.8% |
| d3-flash | 81.4% |
Average of BFCL, API-Bank, When2Call and PhishNChips (%) — distinct from the 37-task HF Decision Index · Higher is better
| d3 | 86.0% |
|---|---|
| clef | 85.6% |
| mouse-1 | 84.9% |
| JEV-27B | 79.8% |
Tool selection (%) · Higher is better
| clef | 98.5% |
|---|---|
| mouse-1 | 97.6% |
| d3 | 97.3% |
| JEV-27B | 96.0% |
Tool selection (%) · Higher is better
| clef | 91.9% |
|---|---|
| d3 | 86.4% |
| mouse-1 | 83.7% |
| JEV-27B | 82.9% |
When to call a tool (%) · Higher is better
| d3 | 86.4% |
|---|---|
| mouse-1 | 85.8% |
| JEV-27B | 78.5% |
| clef | 72.5% |
Phishing decisions (%) · Higher is better
| clef | 79.4% |
|---|---|
| d3 | 73.9% |
| mouse-1 | 72.3% |
| JEV-27B | 61.9% |
Prompt Injection
We hardened mouse-1's resistance to prompt injection, which was a common early complaint from the initial Jev model results. We tested this in an evaluation using planted instructions, which resulted in 0 change of mouse-1 decisions, meaning it performs significantly better at squashing these types of attacks.
Decisions changed by a planted instruction (%) · Lower is better
| mouse-1 | 0% |
|---|---|
| d3-flash | 27.6% |
| d3 | 40.4% |
Built for improving agent harness cost and speed.
Agents make many decisions as they operate from tool calls to verification. These decisions compound over time and are becoming more important as agents now work for hours and even days at a time.
We built mouse-1 to support these choices through a common interface. The harness supplies the task context and candidate options and our model returns a probability for each option. Developers can decide how those scores affect the execution and even delegate to which decisions require human review.
We’ve found that adding mouse-1 to our own coding agent harness significantly improves its latency.
The model is available as an API. Get a key and send a state plus the options you want scored. $5 in credits, no card.