Introduction to Sabiá
Maritaca AI's Portuguese-specialized models, known as Sabiá, are available through an API.
S5, our most intelligent model
S5 is our most intelligent model and the recommended place to start. Use it to analyze documents, work with data, write and edit text, and build assistants that perform tasks with tools.
In the API, S5 has three names:
S5: the model for ongoing and production use.S5-br-sp: the same model, with inference 100% in Brazil (30% more expensive).S5-free: free for a limited time, to try it at no cost (text only, no Batch API, Flex or built-in tools).
For simple, high-volume tasks, use S5-nano: the compact, open-weights version of S5, at R$ 0.10 (input) and R$ 0.60 (output) per 1M tokens.
See a complete example in Getting started with S5 and every model in Models.
Billing for model usage is based on the volume of tokens, considering both input and output data. The specialized training of the Sabiá models guarantees superior performance at a more accessible cost compared to other solutions on the market.
What you can do with Sabiá
The Sabiá family of models, developed by Maritaca AI, can simulate human interactions in Portuguese with high fidelity. In addition to generating texts that cover a wide range of topics and communication styles, from everyday dialogues to complex analyses, it is a useful and accessible tool for anyone seeking meaningful interaction in Portuguese.
Answer questions
It can answer a wide variety of questions on many different topics. The information provided is based on the training data, which includes a vast range of books, websites, and other materials.
Provide Information
It can follow the context of a conversation and adjust responses accordingly, providing information about historical events, science, technology, culture, and more.
Translation
It has the ability to translate texts to and from many different languages.
Teaching and tutoring
It can help explain complex concepts and assist in learning various subjects, providing detailed explanations about a variety of topics.
Creative text generation
It can create texts, stories, poems, dialogues, and other creative content in Portuguese.
Customer support
It is capable of helping with frequently asked questions and offering basic customer assistance.
Data analysis
It can help interpret and analyze provided data.
It is important to remember that, although the models are a powerful tool for language tasks, they do not have consciousness or real understanding and their responses are based on language patterns learned during training.
Performance
Thanks to our specialized training in Portuguese and Brazilian-specific contexts, our models achieve strong performance on national benchmarks at significantly lower cost than comparable alternatives.
Sabiá 4 Thinking
Sabiá 4 Thinking is the reasoning model of the Sabiá family: it reaches frontier quality in Portuguese and Brazilian contexts at the lowest cost among the evaluated models. Compared to Sabiá 4, it brings significant gains in tool use, legal tasks, and answer quality.
The table below shows Sabiá 4 Thinking's performance, benchmark by benchmark, against the leading frontier models across three fronts: function calling and agents, legal, and general tasks. The competitors (Gemini 3.1 Pro, GPT-5.4, and Opus 4.8) were evaluated at reasoning effort medium. Running the full suite costs less than half of GPT-5.4 and about a third of Opus 4.8 on Sabiá 4 Thinking. In bold, the best in each row.
| Benchmark | Sabiá 4 Thinking | Gemini 3.1 Pro(effort medium) | GPT-5.4(effort medium) | Opus 4.8(effort medium) |
|---|---|---|---|---|
| Total cost · R$ to run the suite | R$ 206 | R$ 281 | R$ 449 | R$ 590 |
| Function calling / Agents | ||||
| Pix-Bench · internal | 100% | 100% | 100% | 97% |
| Ticket-Bench · public | 98% | 100% | 98% | 96.7% |
| MARCA · public | 83.9% | 84.8% | 93.2% | 91.5% |
| Legal | ||||
| OAB (judge) · internal | 90.1% | 91.1% | 91.6% | 90.1% |
| Legal drafting · internal | 77.7% | 75.9% | 72.8% | 74.8% |
| Case extraction · internal | 92.3% | 91.4% | 95.7% | 94.3% |
| General | ||||
| BLUEX · public | 93% | 96.8% | 95.7% | 95.4% |
| ENAMED · public | 94.4% | 98.9% | 97.8% | 97.8% |
| POSCOMP · public | 90.8% | 94.6% | 94.6% | 96.2% |
| PoETa v2 · public | 83.7% | 85% | 83.3% | 86.3% |
| Sotaques Digitais · public | 94.6% | 97.6% | 97.8% | 97.8% |
Sabiá 4
Read the technical paper (arXiv) for more details about Sabiá 4.
The table below compares Sabiá 4, Sabiá 3.1, and competing models across a range of Brazilian evaluations—including legal exams, educational assessments, legislative knowledge, agentic capabilities, and conversational skills—considering both quality and API execution cost.
Sabiá 4 has higher accuracy on Brazilian Laws and consistent improvements over Sabiá 3.1 across every metric, with a notable lead in agentic capabilities. It also comes at a cost far below the frontier models listed.
This combination reinforces Sabiá 4 as a balanced choice for legal, educational, and institutional applications in Brazil that need local context and strong cost-performance.
| Benchmark | Description | Metric | Sabiá-4 | Sabiá-3.1 | GPT-4.1 | GPT-5.2 w/o reasoning | GPT-5.2 reasoning | Gemini-3-Pro (low) | Gemini-3-Pro (high) | Kimi-k2-thinking | Qwen3-235b-instruct-2507 | Deepseek-v3.2 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Cost | Cost to run the benchmarks below | Brazilian reais spent on tokens via API | R$ 80.49 | R$ 62.15 | R$ 182.49 | R$ 307.12 | R$ 752.41 | R$ 403.31 | R$ 804.07 | R$ 516.52 | R$ 44.36 | R$ 49.22 |
| Brazilian laws | Knowledge of Brazilian legislation | Accuracy (5 options) | 97.4% | 77.8% | 80.8% | 84.0% | 86.3% | 74.9% | 88.6% | 59.1% | 65.9% | 67.3% |
| OAB Bench | Legal writing (lawyer) – 21 exams | Average score (0–10) | 7.49 | 7.21 | 7.30 | 8.07 | 8.73 | 9.05 | 8.90 | 6.62 | 6.33 | 6.40 |
| Magis Bench | Legal writing (judge) – 24 exams | Average score (0–10) | 5.08 | 4.97 | 5.55 | 6.66 | 6.99 | 7.79 | 7.48 | 4.49 | 4.52 | 4.88 |
| Agentic capabilities | Use of tools and environments in Portuguese | Pass³ and success@1 | 72.2% | 43.1% | 73.3% | 81.1% | 85.7% | 90.4% | 90.1% | 77.3% | 67.8% | 40.5% |
| Brazilian exams | 13 exams (ENEM, USP, OAB, etc.) | Accuracy (4 and 5 options) | 86.6% | 82.4% | 86.1% | 88.0% | 92.9% | 93.3% | 95.0% | 83.0% | 82.0% | 84.0% |
| Portuguese Multi-IF | Ability to follow instructions | Strict, average of 3 turns | 82.0% | 80.7% | 82.7% | 83.7% | 87.2% | 86.0% | 88.0% | 86.0% | 84.4% | 81.5% |
| BRACEval | Conversational skills in Portuguese | Wins against GPT-4o | 53.8% | 44.6% | 50.2% | 59.0% | 60.2% | 70.8% | 68.1% | 56.9% | 65.6% | 60.8% |