OpenAI says its newest model can now move old COBOL programs into modern languages almost every time. In a post on Monday, Oct. 19, 2026, the company said GPT-6.1 Sol passed 86.6% of its COBOL migration evaluation at low reasoning effort, up from 5.4% for GPT-5.6 Sol, and reached 97.9% at higher effort, "nearly saturating the evaluation."

What OpenAI measured

COBOL, Fortran and Visual Basic are "low-resource" languages: there is far less public code for models to learn from than for Python or JavaScript, yet OpenAI notes COBOL supports more than 40% of online banking systems and 95% of ATM transactions, figures it cites from outside sources.

  • COBOL migration: models rewrite COBOL programs in C++, Java, Kotlin, Python or Rust. Hidden tests require byte-for-byte matching output, and a migration only counts if every test passes. OpenAI says GPT-6.1 Sol's low-effort run cost 61% of GPT-5.6 Sol's API cost.
  • MultiLangBench: models build a SQLite-compatible database engine from scratch in 20 languages. GPT-6 Astra averaged 87.4%, up from 26.8%, including COBOL rising from 10.9% to 72.1% and Fortran from 21.4% to 83.5%. The cheaper GPT-6 Luna reached 40.3% at about $0.77 a sample versus $3.48 for GPT-5.6 Sol.

What is still unproven

OpenAI released MultiLangBench on GitHub, but not the COBOL migration test behind the headline number, so outsiders cannot check how many programs it holds or where they came from, AImagazine reported after reviewing the post and repository. It also noted the public repo's README tells users to run a default subset of 272 tests, and that OpenAI's post compares only against its own earlier model, not Anthropic's or Google's.

OpenAI itself cautions that applying the results to production software "still requires expert review and testing against the system's business rules and dependencies."

The release lands as OpenAI pushes to raise new money; see our story on OpenAI's $30 billion funding round. For more on agentic coding risks, read OpenAI's agent misalignment reports.