Best AI for Spanish: GPT Astra led 17 models with 4 of 5 right, and only it knew the RAE idiom

Which AI is best for Spanish?

Only one of 17 models knew the Spanish idiom in our Multilingual LLM Bench, run on 2 October 2026. That was GPT Astra, which got 4 of 5, and the idiom is half of its lead. Seven models tied behind it on 2 of 5, Claude Opus, Gemini 3.1 Pro and GPT Sol among them. Each language gets just 5 questions, so one answer moves a score by 20 points. Split the answers by the type of question and you can see where Spanish still trips them up:

The idiom question is public. It asks for the colloquial phrase the Diccionario de la lengua española gives for making one person pay alone for what a group lost. The answer is cargar las cabras, and the dictionary also lists echar las cabras. Ten of the 17 models wrote pagar el pato. That's the famous one, but the RAE defines it as taking a punishment someone else deserved. Close, not right.

GPT Astra wrote "Echar las cabras." It was the only pass. You can read all 17 answers on our public question page.

This page ranks the chatbots that write, translate and answer in Spanish. If you want to learn Spanish, that's a different pick: tutor apps such as Langua and Speak are built around speaking practice, and none of them was in our test.

How did 17 models do on Whizi's own Spanish questions?

GPT Astra won with 4 of 5, and seven models shared second place with 2. Every model got the same 5 questions in each of 30 languages, at low reasoning. We graded each answer against a sourced key, and an answer in the wrong language scored zero.

Five questions is a small sample. One question moves a model's Spanish score by 20 points, so read this as a rough order, not a fine ranking.

The second local question decided the top. Only GPT Astra and GPT Luna got it, which is why the cheapest GPT we ran tied Claude Opus. Astra was the only model to add the idiom on top.

Don't read the Spain against Latin America split as a verdict on either. The Spain question sat on a harder difficulty level than the Latin American one, so 2 against 9 says more about our dial than about the dialect.

Claude Fable did something odd on the idiom. It reasoned aloud in English through more than a dozen Spanish phrases, then settled on cargar el mochuelo, which is wrong. Grok 4.7 got none of the 5 Spanish questions, though it scored 44% across all 30 languages.

Don't use our bench to rank Spanish against other languages either. Spanish averaged 25.9% across the 17 models. That says more about how hard each set of five questions was than about the language.

What do the public Spanish leaderboards say?

The public leaderboards put Spanish close to English for every big model, but they can't tell the top ones apart. Artificial Analysis scores Spanish with Global-MMLU-Lite, a multiple choice test. As we read it in October 2026, five models tie at 94: Gemini 3.1 Pro, Gemini 3 Pro, Gemini 3 Flash, Claude Opus 4.6 and Claude Opus 4.5. When a cheap Flash model can share first place, you know the test is near its ceiling. The board also lags the releases: Claude Opus 5.5, the GPT-6 models and Gemini 4 Argon weren't on it when we read it on 18 October.

The vendors publish their own numbers too. Anthropic's multilingual support page says Claude Sonnet 4.5 keeps 98.2% of its English score in Spanish, and Claude Haiku 4.5 keeps 96.4%. Spanish is the top language in that table. OpenAI's simple-evals results give its o3 model 0.911 in Spanish, on an MMLU test that professional translators rewrote. GPT-4.1 nano, its smallest model there, scored 0.748.

La Leaderboard is the one built by Spanish speakers. María Grandury of SomosNLP led the paper, with Hugging Face and the Barcelona Supercomputing Center among the authors, and it appeared at ACL 2025. Its first version combines 66 datasets in Basque, Catalan, Galician and Spanish varieties from Spain and Latin America, and scores 50 models. The board itself runs as a Hugging Face space.

Three of these four are translated MMLU tests. A translated question can't ask for a Spanish idiom, and that's the gap our five questions aim at.

Does AI write Spain Spanish or Latin American Spanish?

AI writes the Spanish your prompt asks for, so name the country. Spain and Latin America differ in grammar, not just accent. The Diccionario panhispánico de dudas is the style guide that the RAE wrote with the other academies of ASALE, 23 in all. It says that in all of the Americas, and in western Andalusia and the Canary Islands, "ustedes es la única forma empleada para referirse a varios interlocutores". That means vosotros is a form you'll only hear in most of Spain. In the Americas, ustedes is the plural "you" for friends and strangers alike.

Voseo is the second split. The DPD calls Argentina, Uruguay and Paraguay "áreas de voseo generalizado", and it says vos is just as general in five Central American countries, from Guatemala down to Costa Rica. Chile mixes tú and vos. In Mexico, vos is only alive in Tabasco and Chiapas. So a prompt that only says "Spanish" leaves the model to guess between tú, vos, vosotros and ustedes.

Vocabulary is the third. The DLE marks ordenador, for a computer, as a Spain word ("Esp."), and its entry for computadora carries no such mark. Tell the model the country, and check the words your reader will notice first.

```

Write in Spanish for readers in [Spain / Mexico / Argentina / Colombia / Chile].

Use [vosotros / ustedes] for plural "you" and [tú / vos / usted] for one reader.

Use the everyday words people in that country use, not a neutral mix.

Keep every name, number, date and amount exactly as written.

At the end, list any word or phrase that would sound foreign in that country.

Task: [a product email / a school essay / a WhatsApp message] about [topic], about [200] words.

```

For translation, swap the task line for "Translate this English text into Spanish for [country]" and keep the rest. Our guide to using AI to translate covers the checks to run after.

Should you use a Spanish model like ALIA or LatamGPT?

Use ALIA or LatamGPT when you want to run a model yourself, or when you need Catalan, Basque or Galician next to Spanish. If it's just Spanish quality you want, the big chatbots have the scores, and these two don't have any yet. Neither is in Artificial Analysis's Spanish top five, and neither was in our run.

Spain's ministry calls ALIA "an open project, a continuous process in which experience, research, training and new knowledge are added". The ALIA-40b model card says the BSC doubled the weight of Spanish in training, which brought it to 16.12% of the corpus.

LatamGPT is infrastructure, not a chatbot. Its model card lists Spanish, Portuguese and English and says it "is not optimized for languages other than Spanish, Portuguese, and English". Open Source For You reported on 11 February 2026 that the first version cost about $550,000, and that the money came from CENIA and from CAF, the Development Bank of Latin America. The card doesn't list any benchmark results.

How much do ChatGPT, Claude and Gemini cost in Spain and Latin America?

Colombia pays the most for ChatGPT Plus and for Claude Pro: COP 99,900 a month for either one, which is $31.38. We read each app's App Store listing in five Spanish speaking stores on 18 October 2026. Dollar figures use the Currency API rates of 17 October 2026, rounded to the cent.

Apple shows a bare "$" in the Mexican, Colombian and Chilean stores. We read those as local pesos, since that's the only way the amounts make sense. Argentina's store bills all three apps in US dollars. Google AI Pro is the cheapest of the three main plans in Spain, Mexico and Colombia. In Chile, ChatGPT Plus is.

GPT Astra, the model that topped our test, isn't in ChatGPT Plus chats. OpenAI's help page says "GPT-6 Pro, powered by GPT-6 Astra, is available in ChatGPT for Pro $100, Pro $200, Business and Enterprise plans." Plus gets GPT-6 Sol in chat. It gets Astra only in ChatGPT Work and Codex. Claude Pro includes Claude Opus, per claude.com/pricing. For more countries, see Claude Pro price by country. If your card keeps failing, paying for AI when your card is declined covers the workarounds.

Where does Whizi fit for Spanish?

Whizi is our product, so here's where it helps and where it loses. GPT Astra, the only model with 4 of 5 in our Spanish test, is on Whizi Powerhouse at $49.99 a month, less than half the price of ChatGPT Pro 100 in all five stores above. Powerhouse's 10,000 monthly credits buy 100 Astra messages, though, so it suits checking hard Spanish, not chatting all day. Pro at $29.99 has Claude Opus, ChatGPT Sol and Gemini Pro, three models that each got 2 of 5, with side by side model chats to compare them on one prompt. Whizi loses on a single plan: Claude Pro alone is cheaper in four of the five stores (Colombia is the exception), and Whizi carries no ALIA or LatamGPT model. Side by side model chats shows how compare works.