GPT-4 hits 90% accuracy in Portuguese, reshaping Brazil's classrooms

Arco Educação, a name that carries weight across more than 11,000 schools in Brazil, just dropped some hard numbers on why it hitched its wagon to OpenAI. The company’s data science team put the leading generative AI models through their paces in Brazilian Portuguese. The result wasn’t even close. GPT-4 clocked a 90% accuracy rate for creating and grading pedagogical content, leaving the next-best model in the dust at 73%. When Arco’s own pedagogical experts evaluated the output, they gave GPT-4 a 70% approval rating for generated questions against a 56% second-place finish. For a country where teachers burn a third of their time on administrative tasks—a full 40% above the global average—those percentage points translate directly into hours reclaimed.

But this isn’t a story about a chatbot dropped carelessly into a classroom. Arco built what it calls the Teacher Assistant through a co-creation process that, according to the company, involved hundreds of educators. The focus is relentlessly practical, targeting the soul-crushing grind of lesson planning, especially for the 1.7 million students enrolled in special needs programs in Brazil. Teachers typically spend hours with inclusion support centers to craft a single Individualized Education Plan. The Assistant, which codifies inclusion archetypes and protocols, lets a teacher describe a student’s needs and receive personalized, adapted activities. Fernando, a high school teacher, noted it saves him hours and “provides suggestions that serve as a gateway to new ideas.”

Marcia, a high school Director of Curriculum and Instruction, framed the impact in terms of confidence and autonomy, calling it “wonderful” for parents to know a school has a concrete, personalized plan. The privacy model here is also worth noting: only the teacher sees the uploaded student data. Arco is betting that this walled-garden approach builds the trust needed for sensitive applications. The economics are being managed by mixing models, fine-tuning lighter options like GPT-4o mini and GPT-3.5 to keep the quality high without letting costs spiral out of control.

The rollout is starting small but scaling fast. A pilot is underway in 50 schools, with plans to hit 600 schools and roughly 70,000 students by year-end. The real ambition is to blanket Arco’s entire base of over 3 million students through 2025. CEO Ari de Sá Cavalcante frames AI as a tool to let teachers focus on a student’s “unique learning journey.” If the 90% accuracy figure holds up outside the lab and across diverse regional dialects, this could set a hard benchmark for what localized AI in education actually requires. I’ll be watching to see if the public sector buys in.