Metric
accuracy (%)
Status
Still discriminating
Category
Reasoning & Knowledge
Maintainer
Epoch AI
Chance baseline
9.22%
What it tests
Planning from a textual game state without relying on a publicly known evaluation identity. Epoch withholds the game's name to reduce targeted preparation for the benchmark.
How it works
The suite contains 100 programmatically generated positions, each filtered to have one best next move. A minimal agent submits its move through a tool. The normalized move must match the answer key; illegal or missing moves earn no credit.
How to read a score
Higher is better. Epoch's metadata gives a 9.22% random baseline. This is move-selection accuracy, not full-game win rate.
Limitations
The undisclosed game, prompts, and transcripts make independent auditing harder. A small sample and a single game distribution limit conclusions about general reasoning or broad game-playing skill.
Leaderboard
Checked Nov 17, 2026, 07:00 UTC · scores last changed Nov 11, 2026Best Mystery Game Puzzles score per model, ranked. Each score links to its source.
Run your own prompt on the top 3
Claude Fable 5.2, GPT-6 Astra, GPT-6.1 Sol, side by side on the same prompt, with each one's time and cost. A free account includes starter credit.
Compare the leaders
More Reasoning & Knowledge benchmarks
GPQA Diamond
Graduate-level science questions that are hard even with a search engine.
Humanity's Last Exam
A broad, extremely hard closed-ended exam at the frontier of human knowledge.
MMLU
The classic 57-subject knowledge exam, now largely solved by frontier models.
ARC-AGI-2
Novel visual-grid puzzles built to resist memorization and brute-force scaling.
ARC-AGI-1
The original abstraction-and-reasoning grid puzzles, now solved cost-efficiently.
BIG-Bench Hard
23 reasoning tasks that once beat models, now largely saturated.
SimpleQA Verified
Short factual questions that probe hallucination and calibration.
HellaSwag
Commonsense sentence completion that is trivial for people and long resisted models.
WinoGrande
Large-scale Winograd pronoun resolution that turns on one commonsense word.
PIQA
Physical commonsense: which everyday method actually works?
OpenBookQA
Grade-school science questions needing a known fact plus commonsense to combine it.
TriviaQA
Open-domain factual recall scored by strict exact match.
ANLI
Adversarially-collected natural-language inference across three rounds.
SuperGLUE
The harder successor to GLUE, a language-understanding suite now surpassed by LLMs.
LAMBADA
Predict the final word of a passage that requires the whole context.
SimpleBench
Trick questions where ordinary people comfortably beat frontier models.
MMLU-Pro
Academic and professional reasoning across a broad set of subjects with harder multiple-choice distractors than MMLU.
GPQA Main
Graduate-level science reasoning on the main GPQA question set.
NarrativeQA
Reading comprehension and answering questions about narratives.
Natural Questions (open-book)
Finding an answer to a naturally occurring search question within supplied Wikipedia evidence.
Natural Questions (closed-book)
Recalling factual answers to naturally occurring search questions without supplied Wikipedia evidence.
MedQA
Medical knowledge and reasoning on professional medical board-examination questions.
CLEAR
Identifying whether a patient note supports, rejects, or leaves uncertain the presence of a medical condition.
MEDEC
Detecting whether a clinical narrative contains a medical error.
HEAD-QA
Biomedical knowledge and reasoning on multiple-choice examination questions.
Medbullets
Applying medical knowledge to patient scenarios in USMLE-style questions.
MedMCQA
Medical entrance-examination knowledge across clinical and foundational subjects.
PubMedQA
Interpreting biomedical abstracts to answer research questions with yes, no, or maybe.
n2c2 Clinical Trial Matching
Determining whether a patient's clinical notes satisfy clinical-trial eligibility criteria.
MedHallu
Detecting unsupported statements in biomedical question-answer pairs.
AI2 ARC Challenge
Multiple-choice school science questions selected to resist simple retrieval.
Chess Puzzles
Finding the uniquely best next move from a chess position described in text.
DTBench
Reasoning correctly about decision problems involving predictions and correlated agents.
LMCA
Judging conceptual arguments in agreement with expert human ratings.
Mystery Game Puzzles 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.