AIMultiple/aimultiple-decision-models-browser

AIM Decision Model Browser Benchmark: 10-task sample

Browser-agent results for 56 decision models (System One models) and two general LLMs, from AIMultiple's decision model benchmark. The AIM System One Index combines these browser results with the classification and AI-Bias benchmarks. The leaderboard covers all 50 tasks. Ten of the tasks are published here with every model's result; the other 40 are withheld for later runs.

Files

• tasks.jsonl: the 10 sample tasks, each with its start URL, instruction, accepted final URLs and pass conditions.
• results.jsonl: one row per sample task and run, with pass or fail, how the attempt ended, seconds, browser actions and run date.
• leaderboard.jsonl: one row per run on all 50 tasks (60 runs): rank, tasks completed, success rate, median seconds per attempt, cost and run date.
• media/: two demo clips.

Leaderboard

Rank orders the 46 ranked decision models by tasks completed, and equal counts share a rank. GPT-6 Astra and Gemini 3.8 Flash are references and have no rank. Ten runs are not ranked; the note column says why. Cost is API fees, or for models we ran ourselves, GPU time at one A100 rate. Each task ran once per model, so a gap of a few tasks does not show which model is better.

How tasks are scored

An attempt passes only if the model declares the task done within the limits and a separate check of the final page confirms every condition: an accepted final URL, every selector condition and a fully loaded page. A page still loading at the first check counts as passed if it passes once loaded. Limits per attempt: 900 seconds, 60 browser actions and 120 decision calls, in a fresh browser session.

Conditions use CSS selectors. text is the element's whitespace-normalized text, texts the list over all matches, value a form value, count the number of matches and visible rendered and not hidden by CSS.

Sample

ID	Site	Type
B04	Books to Scrape	Category switch to product page
W01	Wikipedia	Article lookup
Q06	Quotes to Scrape	Dependent filters
Q08	Quotes to Scrape	Infinite scroll
V07	Web Scraper	Native dropdown
H05	Scrape This Site	AJAX table
C06	ScrapingCourse	Product options
M02	ScrapeMe	Sort
D04	web-scraping.dev	Load more
T02	TestPages	Form submit

The ten tasks cover every site in the benchmark and a mix of outcomes, so sample pass rates differ from the full run: Jev, Kev-9B and Solar Decide pass 5 of these 10 but completed 18, 20 and 10 of all 50.

Runtime

Every model ran on the open-source runtime browser-use/jev-ultrafast. At each step it turns the page into text plus a numbered list of visible controls, and the model picks an operation and a target. A separate text model, Mercury 2.5, types form values for every model; Mercury 3 did so on 4 of Celeris 1.1 Decision's tasks. Practice sites change over time, so a task can fail later because a site changed.

Demo clips

Five models on W01 (top row Jev, Kev-9B, Laya; bottom row Gemini 3.8 Flash, GPT-6 Astra and a results card):

Four models on Q06, filtering quotes by author and tag:

The clips come from separate demo runs, not from the scored attempts in results.jsonl.