#

Key takeaways

• OpenAI published new evaluations for coding models in low-resource programming languages.

• The report focuses on COBOL, Fortran, and Visual Basic, among others.

• MultiLangBench tests whether a model can build a SQLite-compatible database engine in 20 languages.

• OpenAI also created a COBOL migration evaluation that checks behavior preservation.

• The results are OpenAI's own benchmark results, not independent proof for every legacy system.

What OpenAI published

OpenAI published new evaluations of coding models in low-resource programming languages on October 19, 2026. The report says languages such as COBOL, Fortran, and Visual Basic have less training code and documentation available than Python or JavaScript. It also notes that these languages still appear in essential systems.

The company says it developed MultiLangBench to test a specific coding task across 20 languages. The task is to build a SQLite-compatible database engine. OpenAI says the benchmark starts with an initial database and SQL query. It then compares the modified database with reference SQLite output.

OpenAI also developed a separate COBOL migration evaluation. That evaluation checks whether a rewritten program preserves behavior. This is important because the practical issue is not only rewriting code. It is also keeping existing business rules and behavior intact during maintenance or migration.

Reported benchmark results

OpenAI reported results for several models in MultiLangBench. GPT-6 Astra averaged an 87.4 percent test pass rate at maximum reasoning effort. OpenAI says GPT-5.6 Sol averaged 26.8 percent in the same benchmark. GPT-6 Luna averaged 40.3 percent. OpenAI also estimated an average API cost per sample of $0.77 for GPT-6 Luna, compared with $3.48 for GPT-5.6 Sol.

In the separate COBOL migration evaluation, OpenAI reports that GPT-6.1 Sol reached 86.6 percent accuracy at low reasoning effort. OpenAI says GPT-5.6 Sol reached 5.4 percent in that setting. GPT-6.1 Sol reached 97.9 percent at higher effort.

These figures come from OpenAI's own benchmark materials. They are useful for comparing the models inside the report. They are not independent evidence of performance on every legacy application. That distinction matters because real systems can differ from benchmark tasks.

Why the benchmark design matters

The benchmark design focuses on behavior, not just syntax. That is a useful choice for legacy software work. A program can look correct after rewriting and still behave differently. OpenAI's COBOL migration evaluation tries to measure that risk directly.

MultiLangBench also uses a concrete output check. The model must modify a database in a way that matches reference SQLite output. That makes the task more specific than a general coding prompt. It also gives a clearer pass or fail signal.

The report suggests that low-resource languages need different evaluation methods from mainstream languages. OpenAI frames this around the lack of training code and documentation. The benchmark materials were published alongside the report, which gives readers a way to inspect the setup.

Morocco relevance

The source reports no Morocco-specific fact. For readers in Morocco, the global lesson is conditional: if you work with legacy code, benchmark design should test behavior preservation, not only code generation.

Bottom line

OpenAI's report adds a focused look at coding models in languages with less available training material. It also separates two related problems: building code in low-resource languages and preserving behavior during migration. The reported gains are strong inside OpenAI's own tests, but they should be read as benchmark results, not universal guarantees.

For teams evaluating similar tools, the main takeaway is simple. Measure the task that matters. If the goal is migration, check whether the rewritten program still behaves the same. If the goal is code generation, use a benchmark that reflects the target language and the target output.