Mystery Game Puzzles
Integrity rank #11 of 61 · 35 models scored · top score 54.84 · Claude Opus 5
unknown
Strongest on discrimination, weakest on saturation headroom. Caveat: the public test set does not rule out training-data contamination.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | Claude Opus 5 | 54.84 | reproducedT1 | 2026-07-24 |
| 2 | GPT-5.6 Sol | 53.73 | reproducedT1 | 2026-07-09 |
| 3 | GPT-5.5 | 51.53 | reproducedT1 | 2026-04-23 |
| 4 | Claude Fable 5 | 47.12 | reproducedT1 | 2026-06-09 |
| 5 | Qwen 3.8 Max | 31.7 | reproducedT1 | 2026-07-19 |
| 6 | Gemini 3.7 Flash | 30.6 | reproducedT1 | 2026-08-13 |
| 7 | GPT-5.4 | 30.6 | reproducedT1 | 2026-03-05 |
| 8 | Claude Opus 4.8 | 29.5 | reproducedT1 | 2026-05-28 |
| 9 | Claude Sonnet 5 | 28.4 | reproducedT1 | 2026-06-30 |
| 10 | GPT-5.6 Terra | 28.4 | reproducedT1 | 2026-07-09 |
| 11 | Gemini 3.1 Pro | 27.3 | reproducedT1 | 2026-02-19 |
| 12 | Grok 4.6 | 27.3 | reproducedT1 | 2026-08-12 |
| 13 | DeepSeek V4 Flash 0731 | 27.3 | reproducedT1 | 2026-07-31 |
| 14 | Gemini 3.5 Flash | 25.09 | reproducedT1 | 2026-05-19 |
| 15 | Qwen3.7-Max | 25.09 | reproducedT1 | 2026-05-19 |
| 16 | Gemini 3.6 Flash | 22.89 | reproducedT1 | 2026-07-21 |
| 17 | Claude Opus 4.7 | 20.69 | reproducedT1 | 2026-04-16 |
| 18 | Qwen 3.6 Plus | 19.59 | reproducedT1 | 2026-04-01 |
| 19 | Gemini 3 Flash | 18.48 | reproducedT1 | 2025-12-17 |
| 20 | Kimi K3 | 18.48 | reproducedT1 | 2026-07-16 |
| 21 | Claude Opus 4.6 | 17.38 | reproducedT1 | 2026-02-05 |
| 22 | GPT-5 | 15.18 | reproducedT1 | 2025-08-07 |
| 23 | GPT-5.2 | 15.18 | reproducedT1 | 2025-12-11 |
| 24 | Claude Opus 4.5 | 14.08 | reproducedT1 | 2025-11-24 |
| 25 | Claude Opus 4.1 | 12.98 | reproducedT1 | 2025-08-05 |
| 26 | GPT-5.6 Luna | 12.98 | reproducedT1 | 2026-07-09 |
| 27 | Gemini 3.5 Flash-Lite | 10.77 | reproducedT1 | 2026-07-21 |
| 28 | Kimi K2.6 | 9.67 | reproducedT1 | 2026-04-20 |
| 29 | Claude Sonnet 4.5 | 8.57 | reproducedT1 | 2025-09-29 |
| 30 | Qwen 3.5 Plus (hosted 397B-A17B) | 7.47 | reproducedT1 | 2026-02-16 |
| 31 | Qwen 3.6 Max (Preview) | 7.47 | reproducedT1 | 2026-04-20 |
| 32 | Claude Sonnet 4.6 | 7.47 | reproducedT1 | 2026-02-17 |
| 33 | Qwen 3.5 Flash (hosted 35B-A3B) | 6.37 | reproducedT1 | 2026-02-25 |
| 34 | Qwen 3.6 Flash | 5.27 | reproducedT1 | 2026-04-26 |
| 35 | Inkling-Small | 0 | reproducedT1 | 2026-07-30 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
This benchmark evaluates 35 models, and the spread between its top and bottom scores is 54.84 points. A spread this wide suggests that observed rank differences reflect genuine capability gaps rather than close or indistinguishable performance.
3 cited facts
The benchmark's integrity index is 99, corresponding to grade A, placing it 11th in a universe of 61 benchmarks. Its weakest integrity component is saturation, meaning the ceiling is crowded and top scores cluster, which complicates its function as a discriminator of frontier models under the disclosed harness.
5 cited facts
The benchmark is not saturated, with a top score of 54.84 and 45.16 points of headroom to the ceiling. Only two models cluster near the top within a couple of points, so the leaderboard still has real room to separate the field rather than being dominated by noise.
4 cited facts
Scores are task performance under a consistent disclosed harness, so results are directly comparable within this harness but should not be treated as deployed capability. The test set privacy label is unknown, so whether the test set is public or held out cannot be confirmed; a public set would make high scores warrant greater scrutiny for possible contamination. Contamination history is unknown, leaving the benchmark's susceptibility to contamination and harness sensitivity qualitatively uncharacterized.
3 cited facts
This benchmark's methodology is not documented in the available facts: what it measures, how it was constructed, and its key assumptions are all listed as unknown. Because the intended measurement is undocumented, the benchmark's design cannot be described in plain terms from these fields. No human baseline is documented, so there is no reported comparison point for interpreting results. The sharpest caveat a reader should hold onto is that this benchmark's limitations are undocumented, so any conclusion drawn from its results lacks a stated boundary on that evidence.
6 cited facts
Mystery Game Puzzles: Mystery Game Puzzles as reported in Epoch AI's Capabilities Index CSV.
Claude Opus 5 leads Mystery Game Puzzles at 54.84 — independently reproduced. The full leaderboard above lists every recorded measurement, not just the headline number.
35 models have recorded scores on Mystery Game Puzzles, spanning a score spread of 54.84.
tensor.news grades Mystery Game Puzzles A for integrity (score 99/100), ranking #11 of 61 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — Mystery Game Puzzles still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is unknown, so contamination risk is on the table: a high score may partly reflect training-data overlap rather than capability.
Scores are reported under a consistent harness, so comparisons on Mystery Game Puzzles are reasonably apples-to-apples.
Every Mystery Game Puzzles measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites epoch.ai/data/eci_benchmarks.csv.