SimpleBench
Integrity rank #12 of 65 · 80 models scored · top score 78.28 · Claude Fable 5
Basic human common-sense reasoning that resists memorized knowledge
Strongest on discrimination, weakest on freshness.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | Claude Fable 5 | 78.28 | unverifiedT1 | 2026-06-09 |
| 2 | Claude Opus 5 | 76.72 | unverifiedT1 | 2026-07-24 |
| 3 | Gemini 3.1 Pro | 75.52 | unverifiedT1 | 2026-02-19 |
| 4 | GPT-5.5 Pro | 72.28 | unverifiedT1 | 2026-04-23 |
| 5 | Gemini 3.5 Flash | 72.04 | unverifiedT1 | 2026-05-19 |
| 6 | Gemini 3 Pro | 71.68 | unverifiedT1 | 2025-11-18 |
| 7 | Grok 4.6 | 71.08 | unverifiedT1 | 2026-08-12 |
| 8 | Muse Spark 1.2 | 69.4 | unverifiedT1 | 2026-08-05 |
| 9 | GPT-5.4 Pro | 68.92 | unverifiedT1 | 2026-03-05 |
| 10 | GPT-5.6 Sol | 66.04 | unverifiedT1 | 2026-07-09 |
| 11 | Qwen3.7-Max | 64.48 | unverifiedT1 | 2026-05-19 |
| 12 | Grok 4.5 | 64 | unverifiedT1 | 2026-07-08 |
| 13 | GPT-5.5 | 62.8 | unverifiedT1 | 2026-04-23 |
| 14 | Claude Opus 4.6 | 61.12 | unverifiedT1 | 2026-02-05 |
| 15 | Claude Opus 4.8 | 57.76 | unverifiedT1 | 2026-05-28 |
| 16 | Qwen 3.6 Max (Preview) | 55.6 | unverifiedT1 | 2026-04-20 |
| 17 | Gemini 2.5 Pro (Jun 2025) | 54.88 | unverifiedT1 | 2025-06-05 |
| 18 | Claude Opus 4.5 | 54.4 | unverifiedT1 | 2025-11-24 |
| 19 | Claude Opus 4.7 | 54.04 | unverifiedT1 | 2026-04-16 |
| 20 | GPT-5 Pro | 53.92 | unverifiedT1 | 2025-10-07 |
| 21 | Gemini 3 Flash | 53.32 | unverifiedT1 | 2025-12-17 |
| 22 | DeepSeek V4 Flash 0731 | 53.32 | unverifiedT1 | 2026-07-31 |
| 23 | Kimi K3 | 52.84 | unverifiedT1 | 2026-07-16 |
| 24 | Claude Sonnet 5 | 52.72 | unverifiedT1 | 2026-06-30 |
| 25 | Grok 4 | 52.6 | unverifiedT1 | 2025-07-09 |
| 26 | Claude Opus 4.1 | 52 | unverifiedT1 | 2025-08-05 |
| 27 | GLM-5.2 | 50.56 | unverifiedT1 | 2026-06-16 |
| 28 | Claude Opus 4 | 50.56 | unverifiedT1 | 2025-05-22 |
| 29 | Kimi K2.7 Code | 49.48 | unverifiedT1 | 2026-06-12 |
| 30 | GPT-5.2 Pro | 48.88 | unverifiedT1 | 2025-12-11 |
| 31 | GPT-5 | 48.04 | unverifiedT1 | 2025-08-07 |
| 32 | GLM-5.1 | 46.12 | unverifiedT1 | 2026-04-07 |
| 33 | Claude Sonnet 4.5 | 45.16 | unverifiedT1 | 2025-09-29 |
| 34 | GLM-5 | 43.84 | unverifiedT1 | 2026-02-11 |
| 35 | GPT-5.1 | 43.84 | unverifiedT1 | 2025-11-13 |
| 36 | o3 | 43.72 | unverifiedT1 | 2025-04-16 |
| 37 | Gemini 2.5 Pro (Mar 2025) | 41.92 | unverifiedT1 | 2025-03-31 |
| 38 | Inkling | 40 | unverifiedT1 | 2026-07-15 |
| 39 | GPT-5.6 Terra | 38.68 | unverifiedT1 | 2026-07-09 |
| 40 | GLM-4.7 | 37.24 | unverifiedT1 | 2025-12-22 |
| 41 | Kimi K2.5 | 36.16 | unverifiedT1 | 2026-01-27 |
| 42 | GPT-5.6 Luna | 36.16 | unverifiedT1 | 2026-07-09 |
| 43 | Claude 3.7 Sonnet | 35.68 | unverifiedT1 | 2025-02-24 |
| 44 | DeepSeek-V4-Flash | 35.56 | unverifiedT1 | 2026-04-24 |
| 45 | MiniMax-M3 | 34.96 | unverifiedT1 | 2026-06-01 |
| 46 | GPT-5.2 | 34.96 | unverifiedT1 | 2025-12-11 |
| 47 | Claude Sonnet 4 | 34.6 | unverifiedT1 | 2025-05-22 |
| 48 | o1-preview | 30.04 | unverifiedT1 | 2024-09-12 |
| 49 | Claude 3.5 Sonnet (October 2024) | 29.68 | unverifiedT1 | 2024-10-22 |
| 50 | Gemini 2.5 Flash (Jun 2025) | 29.44 | unverifiedT1 | 2025-06-17 |
| 51 | DeepSeek-R1 (May 2025) | 28.96 | unverifiedT1 | 2025-05-28 |
| 52 | o1 | 28.12 | unverifiedT1 | 2024-12-17 |
| 53 | DeepSeek-V3.1 | 28 | unverifiedT1 | 2025-08-21 |
| 54 | o4-mini | 26.44 | unverifiedT1 | 2025-04-16 |
| 55 | Grok 3 | 23.32 | unverifiedT1 | 2025-04-09 |
| 56 | Qwen 3.6 Flash | 22.24 | unverifiedT1 | 2026-04-27 |
| 57 | GPT-4.5 | 21.4 | unverifiedT1 | 2025-02-27 |
| 58 | Gemini 2.0 Flash (Dec 2024) | 17.32 | unverifiedT1 | 2024-12-11 |
| 59 | Qwen3-235B-A22B | 17.2 | unverifiedT1 | 2025-04-28 |
| 60 | DeepSeek-R1 | 17.08 | unverifiedT1 | 2025-01-20 |
| 61 | Gemini 2.0 Flash Thinking (Jan 2025) | 16.84 | unverifiedT1 | 2025-01-21 |
| 62 | Llama 4 Maverick | 13.24 | unverifiedT1 | 2025-04-05 |
| 63 | Claude 3.5 Sonnet | 13 | unverifiedT1 | 2024-06-20 |
| 64 | DeepSeek-V3 (Mar 2025) | 12.64 | unverifiedT1 | 2025-03-24 |
| 65 | Gemini 1.5 Pro (Sept 2024) | 12.52 | unverifiedT1 | 2024-09-24 |
| 66 | GPT-4.1 | 12.4 | unverifiedT1 | 2025-04-14 |
| 67 | Kimi K2 (Jul 2025) | 11.56 | unverifiedT1 | 2025-07-12 |
| 68 | GPT-4 Turbo (Apr 2024) | 10.12 | unverifiedT1 | 2024-04-09 |
| 69 | Claude 3 Opus | 8.2 | unverifiedT1 | 2024-02-29 |
| 70 | Llama 3.1-405B | 7.6 | unverifiedT1 | 2024-07-23 |
| 71 | o3-mini | 7.36 | unverifiedT1 | 2025-01-31 |
| 72 | Grok-2 (Dec 2024) | 7.24 | unverifiedT1 | 2024-12-12 |
| 73 | Mistral Large 2 (Jul 2024) | 7 | unverifiedT1 | 2024-07-24 |
| 74 | gpt-oss-120b | 6.52 | unverifiedT1 | 2025-08-05 |
| 75 | Llama 3.3 70B | 3.88 | unverifiedT1 | 2024-12-06 |
| 76 | DeepSeek-V3 | 2.68 | unverifiedT1 | 2024-12-26 |
| 77 | o1-mini | 1.72 | unverifiedT1 | 2024-09-12 |
| 78 | GPT-4o (Aug 2024) | 1.36 | unverifiedT1 | 2024-08-06 |
| 79 | Command R+ | 0.88 | unverifiedT1 | 2024-08-30 |
| 80 | GPT-4o mini | 0 | unverifiedT1 | 2024-07-18 |
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.
Eighty models are scored on this benchmark, and the gap between the best and worst scores spans 78.28 points. That wide spread supports real separation between models, meaning rank differences are likely to reflect genuine capability gaps rather than close-score noise.
3 cited facts
The benchmark's integrity score is 98, its integrity grade is A, and its integrity rank is 18 out of 62 benchmarks in the universe, framing its health as a discriminator of frontier models under a disclosed harness. The weakest component in its integrity breakdown is age, which costs the reader freshness: the benchmark may discriminate less sharply among the most recently released frontier models because its design is older.
5 cited facts
This benchmark is not saturated, with a top score of 78.28 and 21.72 points of headroom remaining to the ceiling. Two models cluster within a couple of points near the top. Because the benchmark is not saturated, there is still genuine room to separate the field at the top, so small differences among top models may reflect real capability gaps rather than noise.
5 cited facts
Per harness comparability, this benchmark's scores are produced under a consistent harness, so they are directly comparable when generated with that same harness; however, harness notes indicate that answer extraction and refusal handling can affect results, so cross-harness comparisons should be made cautiously. Per test set privacy, the test set is held out, and contamination history describes the held-out design as the primary contamination defense, with only a limited public sample and no confirmed train-set leakage. In practice, a held-out test set means high task performance under the disclosed harness warrants less contamination scrutiny than a public test set would, while the noted harness sensitivity still means scores should not be treated as deployed capability.
6 cited facts
This benchmark measures basic human common-sense reasoning that resists memorized knowledge, using questions intended to be easy for an average person yet hard for a model. The test was assembled by hand as a set of adversarial multiple-choice questions with several options each, only a small subset of which was released publicly while the remainder stayed private. Its design assumes those questions isolate reasoning from memorized knowledge and that the held-out set generalizes as a reasoning proxy. The sharpest caveat is that external auditability is limited by the tiny public slice, the human comparison rests on a very small unspecialized sample, and the item set is small enough that per-question noise and trick-question framing may reward a particular answer style.
7 cited facts
SimpleBench: A multiple-choice text benchmark of adversarial everyday-reasoning questions (spatio-temporal, social, and trick questions) on which unspecialized humans outperform frontier models. 213 questions total but only 10 are public; the rest are held out.
Claude Fable 5 leads SimpleBench at 78.28. The full leaderboard above lists every recorded measurement, not just the headline number.
80 models have recorded scores on SimpleBench, spanning a score spread of 78.28.
tensor.news grades SimpleBench A for integrity (score 97/100), ranking #12 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — SimpleBench still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is held-out, which lowers — but does not eliminate — the risk that training data leaked into the questions.
Scores are reported under a consistent harness, so comparisons on SimpleBench are reasonably apples-to-apples.
Every SimpleBench measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites simple-bench.com.