This set has 1 other set of results, from earlier versions of the questions or the answer key. We do not draw them on the same chart, because a score computed against different questions is not the same measurement.
Quick summary
Results
How many questions each model answered correctly, from the exact same set. The bars show the count of correct answers, not a general ranking of model ability.
The model is asked to explain what a single word means in free text. An answer counts as correct if it contains every word of one of our approved keys.
A model whose name does not end in a date is called through an id that does not pin the version. The provider may swap the model behind that id at any time without notice, so the figures apply to whichever version was active when the run happened.
How these figures are computed
arti-kata-v1KeteranganHow word-meaning questions are asked. The model is asked to explain what a single word means, with no options offered, so the answer is free text.scoring kunci-alias-v1KeteranganHow word-meaning answers are scored. An answer counts as correct if it contains every word of one of our approved keys, compared on word stems. Affixes, word order, and extra surrounding text do not disqualify it.question version a7ba9fa24e64KeteranganA 12-character fingerprint of every question the model saw: item number, question text, answer options, and option order. Changing a single word changes the fingerprint. Scores under different fingerprints answer different questions, so this page only joins figures that share one.answer key version 873b6f168502KeteranganA 12-character fingerprint of the answer key: the reference, the other answers we accept, and speaker verdicts for items no machine may score. The questions can stay identical while the key changes, and when that happens scores move without a single word of the question changing. That is why it is kept apart from the question version.Per-model detail
The same figures as the chart above, plus the cost per correct answer, the wait time, and the date of the last run. Any column that cannot be guessed from its name carries its own explanation.
| Model | CorrectKeteranganHow many answers matched the key, out of the items scored. | Held backKeteranganThe score on items whose key we have never published. The column to its left covers items whose key is public, so a score there cannot be told apart from a model that read it. This column is what tells them apart. On this set, a gap only counts as a signal above roughly 20.8 points. Anything smaller cannot be separated from chance. Publishing this score also wears it down over time, because every figure released is information about the items we hold back. That is why the held-back slice has a limited lifespan, and will be rotated. | $/correctKeteranganThe cost of one full run divided by ALL its correct answers, including those in the Held back column. The Correct column shows the public slice only, so the denominator here is the sum of both columns. NOT the cost per call: a cheap model that gets a lot wrong can cost more per correct answer than a model that costs more per call. | LatencyKeteranganMedian wait for a single call. Median rather than mean, so one stalled call does not move the number. | Last runKeteranganThe last date we ran this model on the same questions and answer key. |
|---|---|---|---|---|---|
| Gemini 3.5 FlashGoogle | 128/13594.8% | 58/6490.6% | $0.002985$0.5552 | 2,711 msmedian | |
| Gemma 4 31BGoogle | 116/13585.9% | 51/6479.7% | $0.000012$0.0021 | 1,063 msmedian | |
| GPT-5.6 LunaOpenAI | 112/13583.0% | 53/6482.8% | $0.000051$0.0084 | 1,438 msmedian | |
| DeepSeek V4 Flash 0731DeepSeek | 107/13579.3% | 49/6476.6% | $0.000054$0.0085 | 3,123 msmedian | |
| Claude Haiku 4.5Anthropic | 86/13563.7% | 41/6464.1% | $0.000292$0.0371 | 1,635 msmedian | |
| Ling-3.0-flashAnt Group | 69/13551.1% | 27/6442.2% | $0.000036$0.0035 | 1,390 msmedian | |
| Llama3 8B CPT Sahabat-AI v1 Instructlocal | 61/13545.2% | 27/6442.2% | - | 528 msmedian | |
| Gemma-SEA-LION v4.5 E2B-ITlocal | 60/13544.4% | 27/6442.2% | - | 734 msmedian | |
| Solar Pro 4Upstage | 52/13538.5% | 24/6437.5% | $0.000013$0.0010 | 1,142 msmedian | |
| Qwen3 30B A3B Instruct 2507Alibaba | 51/13537.8% | 23/6435.9% | $0.000020$0.0015 | 816 msmedian | |
| Llama 3.1 8B Instructlocal | 26/13519.3% | 11/6417.2% | - | 532 msmedian | |
| Nemotron 3.5 LightningNVIDIA | 25/13518.5% | 9/6414.1% | $0.001081$0.0368 | 3,088 msmedian | |
| Mistral Small 3Mistral | 25/13518.5% | 6/649.4% | $0.000070$0.0022 | 843 msmedian | |
| Hunyuan A13B InstructTencent | 21/13515.6% | 12/6418.8% | $0.000118$0.0039 | 1,747 msmedian |
Models marked local run on our own hardware, so there is no bill to record and the wait time measures our machine, not a service anyone else can buy. The score is still comparable: the items, the key, and the temperature are identical to every other row here.
Every figure in this table can be recomputed from the raw data. Open the data →
Cost and tokensCheapest per correct answer: Gemma 4 31B ($0.000012). Most expensive: Gemini 3.5 Flash ($0.002985).Show chartsHide
All cost figures are in US dollars.
Cost per correct answer
The model with the highest score is not always the cheapest one. This divides the cost of one full test by its number of correct answers.
Cost of one full run
The raw figure, before dividing by the number of correct answers. Every model answered the exact same questions, so this is directly comparable: it is what you pay to run this benchmark once on each model.
Tokens used
Input and output are kept apart because they are priced differently, often tenfold. A long output bar means the model talks a lot, and that is where almost all of the cost difference between models comes from. A model with no bar has no token record, which is not the same as zero.
Earlier versions
The questions or the key changed after these numbers were measured, so they cannot be compared directly with the results above. We keep them rather than delete them.