Mouse-1 – a decision model for your agent harness that you can tune

Mouse is introducing Mouse-1, a general-purpose decision model built for AI agent harnesses, as described in mouse.dev's report. It integrates into an existing harness through a single API and returns a decision in about 100 milliseconds of model time. Mouse designed it so agents complete more work independently, with fewer errors and fewer unnecessary interruptions. In Mouse's tests, Mouse-1 was 163× faster than GPT-6 Sol and returned a decision before Sol produced its first token.

This piece covers the accuracy tables, the scoring call, image input, tuning, and the cost of 1,000 decisions. It is for engineers who already run an agent harness and want a separate model to score options while a larger model plans and writes code.

Mouse-1 – a decision model for your agent: what actually changed

Mouse evaluated Mouse-1 on unfamiliar tools and repositories to measure how well it generalizes beyond its training data. Average decision accuracy, where higher is better, was 89.1% for Mouse-1, 85.8% for d3, 84.2% for Clef, 82.5% for OpenAI Decisions, 81.4% for d3-flash, and 79.1% for Jev. The Agent Tool & Security Suite averages BFCL, API-Bank, When2Call, and PhishNChips, and is distinct from the 37-task HF Decision Index. On that suite, d3 scored 86.0%, Clef 85.6%, Mouse-1 84.9%, and Jev 79.8%.

The task charts do not keep the suite order. BFCL tool selection scored Clef at 98.5%, Mouse-1 at 97.6%, d3 at 97.3%, and Jev at 96.0%. API-Bank tool selection scored Clef at 91.9%, d3 at 86.4%, Mouse-1 at 83.7%, and Jev at 82.9%. When2Call, when to call a tool, scored d3 at 86.4%, Mouse-1 at 85.8%, Jev at 78.5%, and Clef at 72.5%. PhishNChips phishing decisions scored Clef at 79.4%, d3 at 73.9%, Mouse-1 at 72.3%, and Jev at 61.9%.

Mouse-1 – a decision model for your agent: how it works

The harness supplies task context and candidate options, and Mouse-1 returns a probability for each option. Developers decide how those scores affect execution and which decisions require human review. The caller sends a state plus the options to score. Tuning sends examples from that harness with the correct answers through the API. Mouse tests each tuning against the base model before making it available. Calibration then sets a confidence cutoff for review. Mouse does not store decisions unless the customer turns Outcomes on.

Mouse-1 reads images on the same request, so every question sees them at the same probabilities and the same price per token. The production model answered one choice each. One image scored hot_dog above 99.99%. Content moderation ("Is this image safe to post?") scored safe 99.96%, spam 0.03%, nudity under 0.01%, and violence under 0.01%. Document routing ("Where does this document go?") scored tax_return 99.99%, invoice 0.01%, contract under 0.01%, and id_document under 0.01%. Browser use on app.example.com/dashboard ("What state is the screen in?") scored error 99.95%, loading 0.03%, ready 0.02%, and login_required under 0.01%.

Mouse-1 – a decision model for your agent: why it matters now

Agents make many decisions as they operate. Mouse says those decisions compound over time and are becoming more important as agents work for hours and even days at a time. Mouse built Mouse-1 as a common interface for those choices and says adding it to Mouse's own coding agent harness significantly improves latency. The posted speed mark remains 163× versus GPT-6 Sol at about 100 ms of model time. Separately, Mouse prices the model at up to 65× cheaper than frontier models per 1,000 decisions.

The chart uses 1,000 input tokens per decision and adds 50 output tokens for generative models. Per 1,000 decisions: 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). New keys include $5 in credits, no card.

Mouse-1 – a decision model for your agent: who is affected

Harness authors are the direct users. Mouse's diagram places a decide step, labeled Mouse-1 at about 100 ms, beside harness work that routes, gates, and verifies, a main model that plans and writes code, and tools that run a shell, edit files, and call MCP. The loop shown is propose, then execute, then results, with 0:35 beside the 163× GPT-6 Sol comparison. Developers choosing a scorer are comparing Mouse-1 with d3, Clef, OpenAI Decisions, d3-flash, and Jev on the accuracy charts above.

Prompt-injection results name a second group. Mouse hardened Mouse-1 after what it calls a common early complaint from the initial Jev model results. In a test that planted instructions, none of Mouse-1's decisions changed, a 0% flip rate. Lower is better on that chart: d3-flash flipped 27.6% of decisions, Clef 37.4%, d3 40.4%, OpenAI 53.9%, and Jev 65.4%. People sending images use the same request and the same price per token for moderation, document routing, and browser state. Stored decisions exist only when Outcomes is turned on.

Mouse-1 – a decision model for your agent: what to watch

Image input is in beta. Mouse has checked the production model on cases like the printed hot_dog, safe-to-post, tax_return, and dashboard examples, not on a published image benchmark yet. Mouse-1 leads the unfamiliar-repository average, at 89.1%. It does not lead the suite average (84.9%, behind d3 at 86.0% and Clef at 85.6%), API-Bank tool selection (83.7%, behind Clef at 91.9%), or PhishNChips (72.3%, behind Clef at 79.4% and d3 at 73.9%). When2Call has d3 at 86.4% and Mouse-1 at 85.8%.

Mouse tests each tuning against the base model before making it available, so a harness-specific model waits on that check. Calibration uses the examples to set the confidence cutoff that marks decisions for human review. Decisions are stored only when Outcomes is on. The 65× cost line compares Mouse-1 at $0.10 per 1,000 decisions with GPT-6 Sol at $6.50, under 1,000 input tokens per decision and 50 output tokens on generative models. Sol's output price on that sheet is $30 per 1M tokens, and Mouse-1 output is $0.