OpenAI Releases MultiLangBench and Tests GPT-6 on Rare Languages

On October 19, 2026, OpenAI published a report on model performance in low-resource programming languages and released the MultiLangBench benchmark, where a model writes a SQLite-compatible database engine from scratch offline. GPT-6 Astra, at its maximum reasoning level, averaged 87.4% of tests across 20 languages, compared to 26.8% for GPT-5.6 Sol.

Compiled by Sergey KostenchukPublished 2026-10-21Updated 2026-10-21

2026-10-21 Research OpenAI

🤖 OpenAI Releases MultiLangBench and Tests GPT-6 on Rare Languages

On October 19, OpenAI released the MultiLangBench benchmark on GitHub (Apache-2.0): a model without internet access writes a database engine compatible with SQLite 3.40.1 from scratch, which is then verified by 24,000 tests from the SQLite TCL suite. GPT-6 Astra, at its maximum reasoning level, passed an average of 87.4% of tests across 20 languages, compared to 26.8% for GPT-5.6 Sol.

🌍 Critical infrastructure relies on COBOL and Fortran: according to IBM, COBOL handles 95% of ATM transactions, while Fortran is used in weather forecasting and nuclear reactor safety. The ability to write and port such code reduces migration costs in the face of a shortage of specialists.

👤 The benchmark is open and can be run and reproduced. OpenAI's closed migration test is nearly exhausted—new, stricter evaluations are needed.

Source 1: https://openai.com/index/low-resource-programming-languages/ Source 2: https://github.com/openai/MultiLangBench

Sources