ResearchOpenAI

OpenAI improves coding in COBOL and Fortran

Original title: Advancing coding in low‑resource programming languages

Plain language

What it is: OpenAI announced improvements in its AI models for writing and updating code in older or less common programming languages like COBOL, Fortran, and Visual Basic. The company also introduced new evaluation tests to measure these capabilities.
How it works: The latest models, such as GPT-6 Astra and GPT-6.1 Sol, achieve higher accuracy and lower costs when translating or writing code in these languages. The models were tested on tasks like building database engines and migrating old code to modern languages.
Where it is available: These capabilities are available through OpenAI's API. The test suite used to measure these capabilities has been released on GitHub.
Example: A team can use the models to migrate old COBOL programs to modern languages like Python or Java while keeping the exact same business rules and calculations.

What it is: OpenAI has updated its models to perform better on low-resource programming languages and released two new benchmarks, MultiLangBench and a COBOL migration evaluation. The updates apply to models including GPT-6 Astra, GPT-6 Luna, and GPT-6.1 Sol.
How it works: MultiLangBench evaluates models by having them build a SQLite-compatible database engine from scratch across 20 languages, using database files to verify output parity. The COBOL migration evaluation tests translating COBOL to modern languages while requiring byte-for-byte parity on observable outputs.
Where it is available: The updated models are accessible via the OpenAI API. The MultiLangBench evaluation setup and test suite are publicly available on GitHub.
Example: Migrating a COBOL codebase to C++, Java, Kotlin, Python, or Rust while preserving specific language behaviors like fixed-point arithmetic rounding rules.

The headline and explanation on this page are based on OpenAI's post. The full details are in the original.

Advancing coding in low‑resource programming languages