· Translated from the Spanish original

AI models

OpenAI says GPT-6.1 Sol migrates COBOL with 97.9% accuracy, against 5.4% before

OpenAI published figures on October 20 for COBOL and Fortran. They are its own measurements, and real migrations will still need expert review.

What happened

  • OpenAI published a blog post on October 20, 2026 about its performance in languages with little training material: COBOL, Fortran, Visual Basic and OCaml.
  • On MultiLangBench, which asks models to build a SQLite-compatible database engine in 20 languages, GPT-6 Astra averages 87.4% of tests passed, versus 26.8% for GPT-5.6 Sol.
  • In COBOL it goes from 10.9% to 72.1%. In Fortran, from 21.4% to 83.5%.
  • In a test of migrating COBOL programs to C++, Java, Kotlin, Python or Rust without changing behavior, GPT-6.1 Sol reaches 86.6% accuracy at low reasoning effort (GPT-5.6 Sol: 5.4%) and 97.9% at higher effort.
  • The figures come from OpenAI, according to BleepingComputer. There is no availability date or price.

Why it matters

  • OpenAI points out that COBOL underpins more than 40% of online banking, 80% of in-person card transactions and 95% of ATM transactions. It is the code that moves banks, and the people who maintain it are retiring.
  • For a financial institution in Chile or the region, the argument is no longer “migrating is impossible” but “migrating costs less.” The discussion moves from budget to verification.
  • The consequence the blog does not highlight: 97.9% on a test is not 97.9% at a bank. An error in 2.1% of the programs of a payments system is still an error in production.

The number

97.9%. Accuracy of GPT-6.1 Sol migrating COBOL at higher reasoning effort, according to OpenAI. At low effort, API cost was about 61% of GPT-5.6 Sol’s.

Context

The outlet itself warns that real migrations need expert review and testing. It is the same limit that appears in other experiments: Mamífero reported how an agent rewrote code in eight days. On the cost of running the GPT-6 family, see the savings from prompt caching.

What’s next

  • No availability date or price published for GPT-6.1 Sol.
  • Pending: independent evaluations on real banking code, not on benchmarks.
  • No announced dates.

Bottom line

Banks discover that their technical debt has a market price. The discussion used to be how much it costs to run the models; now it is how much it costs not to migrate.

Sources

Edited by Rodrigo Cornejo. How we select and verify: who writes these notes.

← All notes