Plenty of people distrust AI, with reason, and would rather trust a dictionary. For reading Latin, though, a dictionary has a problem no amount of care solves: it can tell you every case an ending could be, but not which one it is in this sentence. That choice is the whole difficulty of reading Latin, and it is the one thing a language model is good at. So before building MnemoReadâs Latin reader on AI, we measured whether it deserves that trust: against scholarsâ annotations, against a classic dictionary program, and model against model.
THE SHORT VERSION
Most accurate
gpt-6-luna and gpt-6-astra, tied: 2 words in 100 with the case or tense wrong.
Best value
gpt-6-luna, at $0.32 per 1,000 lookups: about 70 times cheaper than astra for the same accuracy.
Against a dictionary
Whitakerâs Words alone gets 47 words in 100 wrong, because it cannot choose between readings. The AI, given its readings, gets 2 wrong.
Why Latin is a hard test for AI
Latin tells you who does what with endings, not word order. The trouble is that one ending can mean several things: puellae can be genitive, dative or nominative plural; -is can be dative or ablative plural, or a genitive singular, or a verb. A model that guesses the wrong case gives you a translation that sounds fine and means something else. That is why we did not test whether models can translate Latin in general. We tested whether they get the grammar of each word right, in its sentence.
A reader has three ways to settle an ending. A dictionary, printed or as a program, lists every reading the form can have and leaves the choice to you. A trained Latin tagger, the kind classicists use on large corpora, reads the context but gives no translation or explanation. Or a language model reads the sentence and answers in words. We measured the first and the third on the same 200 words.
Why not just a dictionary?
Whitakerâs Words is the dictionary program Latin students have used for thirty years: give it a form and it returns every dictionary word and every case, number, tense and mood that form could be. On our 200 words its readings disagreed for 99: half the time, the dictionary alone leaves the reader to decide. Taking its first reading, the most common word it lists, gets the case or tense of 47 words in 100 wrong. Even a reader who always picked the right reading from its list would still be wrong on 15 in 100, because it has no entry at all for 31 of the words: colloquial forms in Petronius, names, rare compounds.
A Latin dictionary program against AI, on the same 200 words
How many of 100 words each approach gives the wrong case or tense for. Whitakerâs Words lists every reading a form could have but cannot choose between them; the AI models choose using the sentence. Shorter is better.
An older small AI model on its own does about as well as that perfect dictionary reader. Give it the dictionaryâs readings to choose from and it does better than either; the newest models, given the same readings, get 2 words in 100 wrong. The dictionary and the AI are not rivals: the dictionary rules out the impossible readings, the model picks among the possible ones. That is how MnemoReadâs reader works.
What this test does not cover: trained Latin taggers such as LatinCy or Stanza. They also read context, and their own benchmarks report about 93% on full morphology, a different measure from ours. They are free to run, but they translate nothing and explain nothing, which is what a reader taps a word for.
How we tested
- The words. 200 words chosen at random from the test set of the Latin Perseus treebank, classical texts (Ovid, Petronius, Phaedrus and others) in which scholars have marked the dictionary form, case, tense and mood of every word by hand. Nouns, verbs and adjectives only.
- The question. Each model got the word, its sentence, and the same instructions MnemoRead uses in its reader: give the dictionary entry, the case or the tense and mood, and a translation. Each also got the possible parses of the word from Whitakerâs Words, a classic Latin dictionary program, to choose from.
- The score. A word counts as right when the modelâs case (for a noun or adjective) or tense and mood (for a verb) matches the treebank. We report it as mistakes per 100 words.
We tried 16 models. Thirteen answered all 200 words. Two Gemini models and Mistral Large were cut short by rate limits on our accounts; their partial results are in the table. We could not include Claude or DeepSeek in this round.
Two hundred words is enough to separate a model that makes 2 mistakes per 100 from one that makes 10, not one that makes 2 from one that makes 3: treat differences under about 3 points as a tie.
The results
The newest models are close to perfect on case and tense, and the ranking has little to do with price: gpt-6-luna, one of the cheapest models we tried, makes fewer mistakes than models that cost 10 to 70 times more. The warning is for older small models: gpt-4.1-nano gets 1 word in 5 wrong, gpt-4o-mini 1 in 6.
How often each AI model gets a Latin wordâs grammar wrong
200 words of classical Latin, the same for every model. For each word we checked the case (nouns, adjectives) or the tense and mood (verbs) the model gave against scholarsâ annotations. Each bar counts the wrong answers per 100 words: shorter is better.
Price is where the models really differ. gpt-6-luna and gpt-6-astra make the same number of mistakes, but astra costs about 70 times more per lookup. gpt-5.5 is both dearer and slower, at 9 seconds a lookup against 5 for luna.
Accuracy against price for 13 AI models
Each dot is one model. Up is more grammar mistakes on our 200 Latin words; right is a higher price. The best value sits in the bottom-left corner: few mistakes, low price.
All the numbers
| Model | Mistakes per 100 | Case right | Tense & mood right | Dictionary form right | Seconds per lookup | Cost per 1,000 lookups |
|---|---|---|---|---|---|---|
| gpt-6-lunaOpenAI | 2.0 | 97.8% | 98.6% | 90.5% | 5.3 | $0.32 |
| gpt-6-astraOpenAI | 2.0 | 97.8% | 98.6% | 90.0% | 6.4 | $22.60 |
| gpt-6-solOpenAI | 2.5 | 97.1% | 98.6% | 90.0% | 4.0 | $4.76 |
| gpt-5.4-miniOpenAI | 3.0 | 96.3% | 98.6% | 90.5% | 3.3 | $2.37 |
| gpt-5.5OpenAI | 3.0 | 96.3% | 98.6% | 90.5% | 9.0 | $20.57 |
| gpt-4.1OpenAI | 5.5 | 92.6% | 98.6% | 91.0% | 2.8 | $4.00 |
| gpt-5-miniOpenAI | 7.0 | 91.2% | 97.3% | 88.5% | 6.0 | $1.67 |
| Mistral MediumMistral | 7.5 | 89.7% | 98.6% | 89.5% | 2.8 | $3.25* |
| gpt-5.4-nanoOpenAI | 8.0 | 89.7% | 97.3% | 85.5% | 3.8 | $0.78 |
| gpt-5-nanoOpenAI | 9.0 | 88.2% | 97.3% | 78.0% | 11.4 | $0.65 |
| gpt-4.1-miniOpenAI | 9.5 | 86.8% | 97.3% | 88.0% | 3.0 | $0.81 |
| gpt-4o-miniOpenAI | 16.5 | 77.2% | 97.3% | 87.0% | 2.9 | $0.28 |
| gpt-4.1-nanoOpenAI | 20.0 | 74.3% | 93.2% | 74.0% | 2.6 | $0.19 |
| Gemini 3.5 Flash-Lite156 of 200 answered | 3.2 | 96.0% | â | 89.7% | â | $0.69 |
| Mistral Large90 of 200 answered | 7.8 | 88.9% | â | 91.1% | â | $0.42 |
| Gemini 3.8 Flash18 of 200 answered | â | â | â | â | â | $0.15 |
List prices of October 2026. *Mistral Mediumâs price is from third-party listings. Grayed rows answered only part of the 200 words. The dictionary-form column tops out near 90% for every model: most of the remaining misses are the treebankâs own conventions (it files multis under multa), not real errors.
Does giving the AI a Latin dictionary help?
In a first run on the same 200 words, we asked three models twice: once on their own, once with Whitakerâs Wordsâ list of possible parses in the prompt. The dictionary does not know which parse is right. It only rules out the impossible ones, and the model picks from what is left using the sentence.
Does a Latin dictionary help the AI pick the right case?
Out of 100 nouns, adjectives and participles, how many each model put in the wrong case: on its own, and when its prompt also listed every reading that Whitakerâs Words, a Latin dictionary program, gives for the word. Fewer is better.
For gpt-4.1-mini, the dictionary cut case mistakes by a third. gpt-4.1 gained a little, and gpt-5-mini made slightly more mistakes with it than without, within the margin of a 200-word test. The lesson: a dictionary makes a weaker model trustworthy, and the newest models have most of it already.
Where the models still go wrong
Parsing single words is half the job. We also asked models to break whole sentences into clauses, and read the answers ourselves. Three sentences tripped up the models MnemoRead used before:
Scylax, canino scilicet usus ingenio, ⌠triclinium implevit
Petronius
Before: gpt-4.1 and gpt-4.1-mini both called canino usus ingenio an ablative absolute.
After: gpt-6-luna: a participial phrase. Usus is the participle of utor, âto useâ, which takes the ablative ingenio: âusing his canine witâ.
Inventum medicina meum est
Ovid, Metamorphoses I
Before: gpt-4.1 read medicina as a genitive: âthe discovery of medicine is mineâ.
After: gpt-6-luna: medicina is nominative, âmedicine is my inventionâ, as Apollo says it.
Male mulcatus graculus
Phaedrus
Before: gpt-4.1-mini translated âbadly soothedâ, mixing up mulcare, to beat, with mulcere, to soothe.
After: gpt-6-luna: âbadly beaten, the jackdawâŚâ.
What we chose, and what it means for you
MnemoReadâs Latin reader now uses gpt-6-luna, with Whitakerâs Wordsâ parses in every word lookup. When you tap a word, you get its dictionary entry with macrons (rÄx, rÄgis, m.), its case or tense in this sentence, and a translation. Tap a sentence and you get its clauses: ablative absolutes, indirect statements, purpose clauses.
If you use AI for Latin anywhere else, three habits help:
- Prefer a recent model. The gap between the newest and the older small ones is larger than any price gap.
- Ask for the dictionary entry and the parse, not just a translation, and check the entry against a dictionary.
- Be most careful with ambiguous endings (-ae, -is, -a, -um) and with participles, where we saw most mistakes.
Try it on a real book
MnemoReadâs free Latin library runs from Fabulae Faciles and Ora Maritima to Caesar, Ovid and Vergil, each book with a measured difficulty level.
Browse Latin booksQuestions about AI and Latin
Can ChatGPT translate Latin?
ChatGPT runs on OpenAI models like the ones in this test, and the recent ones read classical Latin well: gpt-6-luna got the case or tense of 98 words in 100 right. Older small models are another matter: gpt-4.1-nano got one word in five wrong. Whatever you use, check the dictionary entry it gives before you trust a parse.
Which AI is best for Latin grammar?
In our test, gpt-6-luna and gpt-6-astra tied at 98% of words with the case or tense right, followed closely by gpt-6-sol, gpt-5.4-mini and gpt-5.5. gpt-6-luna is the one we chose: it costs about 70 times less than gpt-6-astra for the same accuracy.
Is AI better than a Latin dictionary program?
For reading in context, yes. Whitakerâs Words lists every reading a form could have, but on our 200 words its readings disagreed for 99 of them, and taking its first reading gets the case or tense of 47 words in 100 wrong. gpt-6-luna, choosing from those same readings with the sentence in front of it, gets 2 in 100 wrong. The best setup is both together.
Does giving an AI a Latin dictionary make it more accurate?
For weaker models, yes. With the possible parses from Whitakerâs Words in its prompt, gpt-4.1-mini went from 19.9 to 13.2 wrong cases per 100 nouns, adjectives and participles. The strongest model we tried it with, gpt-5-mini, did not improve.
Tested on 9 and 10 October 2026. The words and their annotations come from the UD Latin Perseus treebank (CC BY-NC-SA); we publish only our aggregate results.