GPT-6 legacy programming improves as GPT-6.1 Sol scores 97.9% on COBOL test

What it claims

OpenAI reports that GPT-6 models are better at COBOL, Fortran and Visual Basic, measured on two benchmarks it built. The COBOL migration benchmark translates COBOL into C++, Java, Kotlin, Python or Rust. GPT-6.1 Sol scores 97.9% accuracy at higher reasoning effort. OpenAI's stated reason for the work is that these languages have less training code than Python but run critical infrastructure. The article cites IBM estimates that COBOL underpins over 40% of online banking systems and handles 95% of ATM transactions.

Evidence

MultiLangBench asks a model to build a SQLite-compatible database engine from scratch, and the score is the average test pass rate across 20 languages:

• GPT-6 Astra: 87.4% at maximum reasoning effort.
• GPT-6 Luna: 40.3%.
• GPT-5.6 Terra: 37.9%.

The existing summary leaves out Luna's score. Luna sits only slightly above the previous-generation Terra, so the large jump belongs to Astra, not to the GPT-6 family as a whole. The page gives no reasoning-effort setting for Luna or Terra.

The COBOL migration benchmark measures accuracy of translation into the five target languages:

• GPT-6.1 Sol: 86.6% at low reasoning effort, rising to 97.9% at higher effort.
• GPT-5.6 Sol: 5.4%, effort setting not stated.

Limits

This is a secondhand news report, produced with AI assistance and reviewed by Cryptonomist editors. It is not OpenAI's own write-up. Both benchmarks are OpenAI's own, and the page does not say whether they are public.

The page leaves out several things a reader would need:

• the list of the 20 MultiLangBench languages, or any per-language scores;
• any separate result for Fortran or Visual Basic;
• how migration "accuracy" is defined or graded;
• the size of either benchmark;
• results for each COBOL target language;
• cost or latency at each reasoning effort.

Before relying on the COBOL figure for a migration project, check that the target language is one of the five tested. Then find OpenAI's primary source for how translations are scored.