Metric
index (points / 100)
Status
Still discriminating
Category
Agentic & Tool Use
Maintainer
Epoch AI
What it tests
Planning and sustained execution in an unfamiliar campaign game. Agents manage strategy, tactics, and written notes while pursuing objectives across the first five in-game days.
How it works
Default runs contain eight practice playthroughs followed by two scored playthroughs. The topline uses the better of the final two, measured as the share of 21 available objectives. Epoch also compares context-compaction settings for its canonical leaderboard.
How to read a score
Higher means more objectives achieved after practice. This is the topline performance score, not a measured learning gain; showing improvement from practice requires the source's separate controlled comparison.
Limitations
Best-of-two selection, note-taking, compaction, and the game harness influence the result. The benchmark covers a limited campaign segment and does not establish general on-the-job learning ability.
Leaderboard
Checked Nov 17, 2026, 07:00 UTC · scores last changed Nov 11, 2026Best Earthborne Rangers (EBR-bench) score per model, ranked. Each score links to its source.
Run your own prompt on the top 3
Claude Fable 5.2, Gemini 4 Carbon, GPT-6.1 Terra, side by side on the same prompt, with each one's time and cost. A free account includes starter credit.
Compare the leaders
More Agentic & Tool Use benchmarks
tau-bench (Airline)
A tool-using agent must serve a simulated customer and follow policy.
BrowseComp
Hard web-browsing tasks that require finding hard-to-locate facts.
OSWorld
Computer-use agents operating a real desktop to finish tasks.
WebDev Arena
Human preference between models building real web apps.
Cybench
Capture-the-flag cybersecurity tasks solved autonomously, end to end.
TheAgentCompany
Can an AI agent do real knowledge work at a simulated software company?
GDPval
Real, economically-valuable tasks from 44 occupations, graded against human experts.
BALROG
Long-horizon agentic reasoning tested inside games, from NetHack to Baba Is You.
APEX-Agents
Completing professional workflows across documents, spreadsheets, and workplace applications.
DeepResearchBench (FutureSearch)
Finding and synthesizing web evidence in a reproducible research environment.
METR: Average Task Success
Average task performance in the evaluation data used for METR's time-horizon analysis.
OSWorld (original)
Completing desktop workflows through computer-use actions in real operating systems.
PostTrainBench
Autonomously improving small language models within a fixed training-compute budget.
Remote Labor Index
Delivering complete freelance projects to a professional acceptance standard.
Earthborne Rangers (EBR-bench) is maintained by Epoch AI. Scores on this page are collected from openly-licensed sources and each links to its origin. See the full matrix to compare every model across every benchmark.