Earthborne Rangers (EBR-bench) (EBR-bench)
We show this table for reference; we do not rank on it.
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.
Benchmark score on EBR-bench — November 13, 2026
We compile the EBR-bench rows from benchmark-owner or independent runs. Claude Fable 5.2 leads the table at 90.5%, followed by Gemini 4 Carbon (87.6%) and GPT-6.1 Terra (78.1%). We do not use these results to rank models overall.
Claude Fable 5.2
Anthropic
Run by EBR-benchGemini 4 Carbon
GPT-6.1 Terra
OpenAI
Run by EBR-bench29 modelsAgenticCurrentDisplay onlyUpdated November 13, 2026
Benchmark score table (29 models)
| Rank | Model / configuration | Score | Who ran it | Parameters (B) | Open / closed |
|---|---|---|---|---|---|
| 1 | Claude Fable 5.2AnthropicRun by EBR-bench | 90.5% | Run by EBR-bench | Not reported | Closed |
| 2 | Gemini 4 CarbonGoogleRun by EBR-bench | 87.6% | Run by EBR-bench | Not reported | Closed |
| 3 | GPT-6.1 TerraOpenAIRun by EBR-bench | 78.1% | Run by EBR-bench | Not reported | Closed |
| 4 | GPT-6 AstraOpenAIRun by EBR-bench | 76.2% | Run by EBR-bench | Not reported | Closed |
| 5 | Claude Opus 5.5AnthropicRun by EBR-bench | 71.4% | Run by EBR-bench | Not reported | Closed |
| 6 | Gemini 4 ArgonGoogleRun by EBR-bench | 64.8% | Run by EBR-bench | Not reported | Closed |
| 7 | Claude Fable 5.1AnthropicRun by EBR-bench | 57.1% | Run by EBR-bench | Not reported | Closed |
| 8 | GPT-6.1 SolOpenAIRun by EBR-bench | 54.3% | Run by EBR-bench | Not reported | Closed |
| 9 | GPT-6 SolOpenAIRun by EBR-bench | 53.3% | Run by EBR-bench | Not reported | Closed |
| 10 | Claude Opus 5AnthropicRun by EBR-bench | 45.7% | Run by EBR-bench | Not reported | Closed |
| 11 | GPT-5.6 SolOpenAIRun by EBR-bench | 44.8% | Run by EBR-bench | Not reported | Closed |
| 12 | Claude Haiku 5.5AnthropicRun by EBR-bench | 43.8% | Run by EBR-bench | Not reported | Closed |
| 13 | Claude Fable 5AnthropicRun by EBR-bench | 39.5% | Run by EBR-bench | Not reported | Closed |
| 14 | GPT-5.5OpenAIRun by EBR-bench | 34.3% | Run by EBR-bench | Not reported | Closed |
| 15 | Muse WonderMetaRun by EBR-bench | 31.4% | Run by EBR-bench | Not reported | Closed |
| 16 | Grok 4.6xAIRun by EBR-bench | 30.5% | Run by EBR-bench | Not reported | Closed |
| 17 | Claude Opus 4.8AnthropicRun by EBR-bench | 28.6% | Run by EBR-bench | Not reported | Closed |
| 18 | GPT-5.4OpenAIRun by EBR-bench | 25.4% | Run by EBR-bench | Not reported | Closed |
| 19 | GPT-5.2OpenAIRun by EBR-bench | 23.0% | Run by EBR-bench | Not reported | Closed |
| 20 | Claude Opus 4.7AnthropicRun by EBR-bench | 19.0% | Run by EBR-bench | Not reported | Closed |
| 21 | Gemini 3.1 ProGoogleRun by EBR-bench | 14.3% | Run by EBR-bench | Not reported | Closed |
| 22 | Claude Opus 4.5AnthropicRun by EBR-bench | 14.3% | Run by EBR-bench | Not reported | Closed |
| 23 | Claude Opus 4.6AnthropicRun by EBR-bench | 12.7% | Run by EBR-bench | Not reported | Closed |
| 24 | GPT-5 (high)OpenAIRun by EBR-bench | 12.7% | Run by EBR-bench | Not reported | Closed |
| 25 | Qwen3.7 MaxAlibabaRun by EBR-bench | 9.5% | Run by EBR-bench | Not reported | Closed |
| 26 | GLM-5.2Z.AIRun by EBR-bench | 9.5% | Run by EBR-bench | Not reported | Open |
| 27 | Gemini 3.5 FlashGoogleRun by EBR-bench | 4.8% | Run by EBR-bench | Not reported | Closed |
| 28 | Kimi K2.6Moonshot AIRun by EBR-bench | 2.4% | Run by EBR-bench | Not reported | Open |
| 29 | Claude Sonnet 4.5AnthropicRun by EBR-bench | 2.4% | Run by EBR-bench | Not reported | Closed |
- Lab-reported
- Benchmark-run
- Independent
How to read this leaderboard
Higher is better. The score is the share of 21 objectives reached in the better of two scored playthroughs. Claude Fable 5.2 leads at 90.5%; Epoch reports that frontier models now beat the median human baseline on average.
Operator receipt: 29 sourced rows are currently displayable on this page; the leading published row is Claude Fable 5.2 at 90.5%.
Honest limit: The score reflects one game and one agent setup, and recent models play under a card ban that older runs did not use. Epoch now analyzes its learning-curve experiments internally, so the public number is the topline score, not the learning gain itself. BenchLM shows EBR-bench as display-only; it does not enter overall or category rankings.
Among the reported EBR-bench rows, Claude Fable 5.2 is first at 90.5%. The third row is 12.4 points behind. The broader top-10 range is 44.8 points, so the table still separates the published systems.
29 models have been evaluated on EBR-bench. The benchmark falls in the Agentic category. We keep these results separate from the weighted global scoring system, so they remain source-specific evidence. EBR-bench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.
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
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.
Freshness and provenance
Version
EBR-bench 2026
Refresh cadence
Quarterly
Staleness state
Current
Question availability
Public benchmark set
BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.
Questions
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%).
Know when it’s worth switching models
The model to choose, the cheaper alternative, and the release we would wait on.
Read a sample issueJoin 7,000+ subscribers.
One email each week. Unsubscribe anytime.