
Mouse-1 – a decision model for your agent harness that you can tune
A general-purpose decision model built for agent harnesses.
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
A general-purpose decision model built for agent harnesses.
Today we’re introducing mouse-1, a general-purpose decision model specifically built for AI agent 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.
In our tests, Mouse-1 was 163× faster than GPT-6 Sol, returning a decision before Sol produced its first token.
Mouse-1 · ~100 ms Decide Harness Route, gate, verify Main model Plan & write code Tools Shell, edit, MCP Propose Execute Results 0:35
163×
faster than 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, 84.2% for Clef, 82.5% for OpenAI Decisions, 81.4% for d3-flash, and 79.1% for Jev.
Unfamiliar tools and repositories (%) · Higher is better
Mouse-1
d3
Clef
OpenAI
d3-flash
Jev
Average decision accuracy, Unfamiliar tools and repositories (%) Mouse-189.1%d385.8%Clef84.2%OpenAI82.5%d3-flash81.4%Jev79.1%
Agent Tool & Security Suite
Average of BFCL, API-Bank, When2Call and PhishNChips (%) — distinct from the 37-task HF Decision Index · Higher is better
d3
Clef
Mouse-1
Jev
Agent Tool & Security Suite, Average of BFCL, API-Bank, When2Call and PhishNChips (%) — distinct from the 37-task HF Decision Index d386.0%Clef85.6%Mouse-184.9%Jev79.8%
BFCL
Tool selection (%) · Higher is better
Clef
Mouse-1
d3
Jev
BFCL, Tool selection (%) Clef98.5%Mouse-197.6%d397.3%Jev96.0%
API-Bank
Tool selection (%) · Higher is better
Clef
d3
Mouse-1
Jev
API-Bank, Tool selection (%) Clef91.9%d386.4%Mouse-183.7%Jev82.9%
When2Call
When to call a tool (%) · Higher is better
d3
Mouse-1
Jev
Clef
When2Call, When to call a tool (%) d386.4%Mouse-185.8%Jev78.5%Clef72.5%
PhishNChips
Phishing decisions (%) · Higher is better
Clef
d3
Mouse-1
Jev
PhishNChips, Phishing decisions (%) Clef79.4%d373.9%Mouse-172.3%Jev61.9%
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, and none of mouse-1's decisions changed.
Decisions changed by a planted instruction (%) · Lower is better
Mouse-1
d3-flash
Clef
d3
OpenAI
Jev
Decision flip rate, Decisions changed by a planted instruction (%) Mouse-10%d3-flash27.6%Clef37.4%d340.4%OpenAI53.9%Jev65.4%
Images
mouse-1 reads images too. Attach them to the same request and every question sees them, with the same probabilities and the same price per token. These are real answers from the production model on the images shown, one choice question each.
Image input is in beta: we have checked it on cases like these, not on a published image benchmark yet. Read the guide.
You can now tune mouse-1 on decisions from your own agent harness. Send examples with the correct answers through our API. We test each tuning against the base model before making it available.
Calibration uses your examples to set a confidence cutoff, so your harness knows which decisions to send for review. We don’t store your decisions unless you turn on Outcomes.
Agents make many decisions as they operate. 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 choose which decisions require human review.
TypeSafe Jev
$0.042
Mouse-1
$0.10
OpenAI Decisions
$0.10
Cloudflare Clef
$0.24
Claude Haiku 4.5
$1.25
DeepSeek V4
$1.52
Gemini 3.7 Flash
$1.88
Claude Sonnet 5
$3.75
Claude Opus 5
$6.25
GPT-6 Sol
$6.50
Cost per 1,000 decisions across models TypeSafe Jev $0.042 ($0.042 / 1M in · $0 out) Mouse-1 $0.10 ($0.10 / 1M in · $0 out) OpenAI Decisions $0.10 ($0.10 / 1M in · $0 out) Cloudflare Clef $0.24 ($0.24 / 1M in · $0 out) Claude Haiku 4.5 $1.25 ($1 / 1M in · $5 / 1M out) DeepSeek V4 $1.52 ($1.32 / 1M in · $3.96 / 1M out) Gemini 3.7 Flash $1.88 ($1.50 / 1M in · $7.50 / 1M out) Claude Sonnet 5 $3.75 ($3 / 1M in · $15 / 1M out) Claude Opus 5 $6.25 ($5 / 1M in · $25 / 1M out) GPT-6 Sol $6.50 ($5 / 1M in · $30 / 1M out)
1,000 input tokens per decision. Generative models add 50 output tokens.
We’ve found that adding mouse-1 to our own coding agent harness significantly improves its latency.
The model is available as an API at $0.10 per 1M input tokens with free output. Get a key and send a state plus the options you want scored. $5 in credits, no card.
$5 in credits on a verified email. No card.
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。