OpenAI's 97.9% COBOL score revives the promise that knocked 13% off IBM in February
OpenAI is the second frontier lab in eight months to claim its models can translate decades-old bank COBOL almost perfectly. The first claim cost IBM its worst day since 2000; the market reversed it by June, and the damage that stuck came from IBM's own numbers.
Vincent Jiang · 3 min read
OpenAI$1.18T — OpenAI, private, latest valuation $1.18T published a benchmark on October 19, 2026, showing GPT-6.1 Sol rewriting COBOL, a business-programming language dating to 1959 2, into C++, Java, Kotlin, Python or Rust and passing 97.9% of hidden behavior tests that demand byte-for-byte identical output 1. Its predecessor, GPT-5.6 Sol, passed 5.4% 1.
The code still supports more than 40% of online banking systems, 80% of in-person card transactions and 95% of ATM transactions 1, and moving one bank system off it runs from millions to hundreds of millions of dollars 3. The programmers who keep it alive are aging out of the workforce, so banks still running COBOL must migrate within a few years 3.
This is the second frontier lab in eight months to make banks the same promise. Anthropic wrote on February 23, 2026, that its Claude Code tool could modernize a COBOL codebase "in quarters instead of years" 4. COBOL runs widely on IBM+1.68% — IBM, up 1.68 percent today mainframes, and IBM sank 13.2% that day, its steepest one-day drop since October 2000 4.
OpenAI's own COBOL migration test jumped from 5.4% to 97.9% in one model generation
Data
| Value | |
|---|---|
| GPT-5.6 Sol (previous frontier) | 5.4% |
| GPT-6.1 Sol, low reasoning | 86.6% |
| GPT-6.1 Sol, high reasoning | 97.9% |
IBM's lasting slide came from IBM's own numbers
The February trade did not stick. IBM closed at a record $329.23 on June 2, 2026, then fell 25% on July 14 after preliminary second-quarter results its chief executive, Arvind Krishna, called disappointing, its worst day on record 5. The filings show why: revenue growth ran 12.1% in Q4 2025, 9.5% in Q1 2026 and 1.1% in Q2 2026 6, and management cut its full-year growth forecast to 4–5% 5. IBM reports third-quarter results today, October 21 5.
IBM's revenue growth slid from 12.1% to 1.1% inside three quarters
Data
| Revenue growth, YoY | |
|---|---|
| Q3 '23 | 4.6% |
| Q4 '23 | 4.1% |
| Q1 '24 | 1.5% |
| Q2 '24 | 1.9% |
| Q3 '24 | 1.5% |
| Q4 '24 | 1% |
| Q1 '25 | 0.5% |
| Q2 '25 | 7.7% |
| Q3 '25 | 9.1% |
| Q4 '25 | 12.1% |
| Q1 '26 | 9.5% |
| Q2 '26 | 1.1% |
A 97.9% score on single files is not a bank migration
On a second OpenAI test, building a database engine from scratch in 20 languages, COBOL stayed the hardest: GPT-6 Astra passed 72.1% 1. The migration evaluation itself covers only single-file programs 7. Real mainframe estates share copybooks, shared data-definition files used by dozens of programs, and run inside job-control language, the scripts that sequence a mainframe's batch work; none of that is in the test 7. OpenAI itself says production use "still requires expert review and testing" against business rules 1.
The research firm Gartner expects more than 70% of mainframe exit projects started this year to miss their benefits, blaming overestimated AI tooling 7. Cognition, the company behind the Devin coding agent, reported five to six times speedups refactoring COBOL at Itaú Unibanco, Latin America's largest private bank, and still calls autonomous migration of transaction systems out of reach 7. IBM sells its own COBOL-to-Java assistant, Watsonx Code Assistant for Z: it is paid to host the code and paid again to help convert it 7.
What would settle it
OpenAI's post names no bank customer, citing only "feedback from developers and enterprise customers" 1, part of an enterprise push that has already been paying the vendors first. Watch today's IBM results for software growth, and watch for the first named bank to pay either lab for a finished migration. The market has priced this promise twice; one paying bank would settle it.
The daily newsletter
