# Earthborne Rangers (EBR-bench) (EBR-bench)

> Epoch AI's EBR-bench tests whether models get better at an unfamiliar game with practice: they play the campaign card game Earthborne Rangers repeatedly, take notes, and are scored on later playthroughs.

Canonical page: https://benchlm.ai/benchmarks/ebrbench

- Category: [Agentic](/agentic)
- Last updated: November 13, 2026

## About EBR-bench

- Year: 2026
- Tasks: Ten playthroughs of an Earthborne Rangers campaign segment per sample (eight learning, two scored)
- Format: Long-horizon game play with notes across playthroughs
- Difficulty: Learning an unfamiliar game over multi-hour playthroughs
- Paper: [Earthborne Rangers (EBR-bench), Epoch AI](https://epoch.ai/benchmarks/ebr-bench)

Earthborne Rangers is an obscure, mostly card-based campaign game that mixes deck-building with turn-by-turn tactics; one playthrough of the evaluated segment takes an experienced human two to four hours. Each sample has ten playthroughs: eight learning runs followed by two scored runs, and the topline score is the better scored run, as a share of 21 possible objectives. Models run with context compaction at 250K tokens (Fable-series models at 90% of their context window). Recent models play under a card-ban setting.

EBR-bench is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (29 models)

| Rank | Model | Creator | Score | Who ran it |
|------|-------|---------|-------|-----------|
| 1 | [Claude Fable 5.2](/models/claude-fable-5-2) | Anthropic | 90.5% | Run by EBR-bench |
| 2 | [Gemini 4 Carbon](/models/gemini-4-carbon) | Google | 87.6% | Run by EBR-bench |
| 3 | [GPT-6.1 Terra](/models/gpt-6-1-terra) | OpenAI | 78.1% | Run by EBR-bench |
| 4 | [GPT-6 Astra](/models/gpt-6-astra) | OpenAI | 76.2% | Run by EBR-bench |
| 5 | [Claude Opus 5.5](/models/claude-opus-5-5) | Anthropic | 71.4% | Run by EBR-bench |
| 6 | [Gemini 4 Argon](/models/gemini-4-argon) | Google | 64.8% | Run by EBR-bench |
| 7 | [Claude Fable 5.1](/models/claude-fable-5-1) | Anthropic | 57.1% | Run by EBR-bench |
| 8 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | OpenAI | 54.3% | Run by EBR-bench |
| 9 | [GPT-6 Sol](/models/gpt-6-sol) | OpenAI | 53.3% | Run by EBR-bench |
| 10 | [Claude Opus 5](/models/claude-opus-5) | Anthropic | 45.7% | Run by EBR-bench |
| 11 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | OpenAI | 44.8% | Run by EBR-bench |
| 12 | [Claude Haiku 5.5](/models/claude-haiku-5-5) | Anthropic | 43.8% | Run by EBR-bench |
| 13 | [Claude Fable 5](/models/claude-fable) | Anthropic | 39.5% | Run by EBR-bench |
| 14 | [GPT-5.5](/models/gpt-5-5) | OpenAI | 34.3% | Run by EBR-bench |
| 15 | [Muse Wonder](/models/muse-wonder) | Meta | 31.4% | Run by EBR-bench |
| 16 | [Grok 4.6](/models/grok-4-6) | xAI | 30.5% | Run by EBR-bench |
| 17 | [Claude Opus 4.8](/models/claude-opus-4-8) | Anthropic | 28.6% | Run by EBR-bench |
| 18 | [GPT-5.4](/models/gpt-5-4) | OpenAI | 25.4% | Run by EBR-bench |
| 19 | [GPT-5.2](/models/gpt-5-2) | OpenAI | 23.0% | Run by EBR-bench |
| 20 | [Claude Opus 4.7](/models/claude-opus-4-7) | Anthropic | 19.0% | Run by EBR-bench |
| 21 | [Gemini 3.1 Pro](/models/gemini-3-1-pro) | Google | 14.3% | Run by EBR-bench |
| 22 | [Claude Opus 4.5](/models/claude-opus-4-5) | Anthropic | 14.3% | Run by EBR-bench |
| 23 | [Claude Opus 4.6](/models/claude-opus-4-6) | Anthropic | 12.7% | Run by EBR-bench |
| 24 | [GPT-5 (high)](/models/gpt-5-high) | OpenAI | 12.7% | Run by EBR-bench |
| 25 | [Qwen3.7 Max](/models/qwen3-7-max) | Alibaba | 9.5% | Run by EBR-bench |
| 26 | [GLM-5.2](/models/glm-5-2) | Z.AI | 9.5% | Run by EBR-bench |
| 27 | [Gemini 3.5 Flash](/models/gemini-3-5-flash) | Google | 4.8% | Run by EBR-bench |
| 28 | [Kimi K2.6](/models/kimi-2-6) | Moonshot AI | 2.4% | Run by EBR-bench |
| 29 | [Claude Sonnet 4.5](/models/claude-sonnet-4-5) | Anthropic | 2.4% | Run by EBR-bench |

## FAQ

### What does EBR-bench measure?

Whether a model can learn on the job: improve at Earthborne Rangers, a little-known card game, over repeated playthroughs using its own notes.

### Which model leads EBR-bench?

Claude Fable 5.2 at max effort leads with 90.5%, ahead of GPT-6.1 Terra (78.1%), GPT-6 Astra (76.2%) and Claude Opus 5.5 (71.4%).

## Compare Top Models on EBR-bench

- [Claude Fable 5.2 vs Gemini 4 Carbon](/compare/claude-fable-5-2-vs-gemini-4-carbon)
- [Gemini 4 Carbon vs GPT-6.1 Terra](/compare/gemini-4-carbon-vs-gpt-6-1-terra)
- [GPT-6.1 Terra vs GPT-6 Astra](/compare/gpt-6-1-terra-vs-gpt-6-astra)
- [GPT-6 Astra vs Claude Opus 5.5](/compare/claude-opus-5-5-vs-gpt-6-astra)