5.4% Leapt to 86.6%: OpenAI Updates a Model That Translates the COBOL Nobody Will Touch into Modern Languages
On October 19, 2026, OpenAI released a model update that raised migration performance for legacy languages such as COBOL. Its COBOL migration evaluation leapt from the previous generation's 5.4% to 86.6%. While it addresses legacy modernization—one of DX's toughest challenges—observers also flag the
Behind the core banking systems and ATMs, COBOL written decades ago is still running. The engineers who can read and write it keep dwindling, and even when you want to rebuild, it's hard to "move it to another language while preserving the exact same behavior"—OpenAI claims that a model update released on October 19, 2026 pushes this migration barrier down a notch. It isn't a new product, but an improvement in the capabilities of existing models, delivered via the API.
Aiming at "low-resource languages" with little training data
What OpenAI targeted this time are languages like COBOL, Fortran, and Visual Basic—languages that remain deep inside financial and business systems yet have scant training material for models. These languages have been an area AI struggled with, where accuracy traditionally stayed in the single digits to around 40%.
To measure this weak spot, OpenAI says it prepared two kinds of evaluations.
- COBOL migration evaluation: scoring, via a private test, whether an existing COBOL program can be translated into C++, Java, Kotlin, Python, or Rust without changing its behavior.
- MultiLangBench: implementing a SQLite-compatible database engine from scratch in each of 20 languages. It measures per-language implementation ability side by side.
The gap in the numbers that appeared across generations
The main published scores are as follows. Note that all of them are figures based on OpenAI's own evaluations.
| Evaluation | Model / condition | Score |
|---|---|---|
| COBOL migration | GPT-5.6 Sol (previous generation) | 5.4% |
| COBOL migration | GPT-6.1 Sol (low reasoning effort) | 86.6% |
| COBOL migration | GPT-6.1 Sol (high reasoning effort) | 97.9% |
| MultiLangBench (20 languages) | GPT-6 Astra (average) | ~87% |
| MultiLangBench (COBOL only) | GPT-6 Astra | 72.1% (previously 10.9%) |
In COBOL migration, the score leapt from the previous generation's 5.4% to 86.6%. With more reasoning effort it is said to reach 97.9%, and moreover GPT-6.1 Sol produces this score at low reasoning effort while being described as roughly 39% cheaper than the previous generation. On MultiLangBench it showed high pass rates across many of the 20 languages, with only COBOL staying at 72.1%—but even that is a dramatic improvement from the previous 10.9%.
Why "migration" becomes a business battleground
What makes this matter weighty is that COBOL still underpins social infrastructure. Citing IBM's figures, COBOL is said to drive more than 40% of online banking, about 80% of in-person card payments, and roughly 95% of ATM transactions.
If AI can carry these frozen assets—ones you want to rebuild but have no people for—"to another language while keeping the same behavior," then legacy modernization (the overhaul of old systems), one of DX's toughest challenges, comes within reach. There are reports that AI agents can autonomously translate and optimize Fortran code at a cost of only a few dollars per unit of processing, and the assumptions behind migration costs are beginning to change. Coding agents like Codex, too, could become the ones carrying out migration work once their underlying models gain this capability.
The pitfall of "it compiles, but the result is different"
On the other hand, there are strong warnings that reading high benchmark scores directly as real-world safety is dangerous.
- All the figures come from OpenAI's internal, private tests, not from independent third-party evaluations.
- Passing a test is a different thing from working correctly on real production workloads and undocumented dependencies. It's entirely possible for a ported program to "run without errors yet produce an answer that differs from the original."
- In regulated industries like finance, even a slight difference in behavior ties directly to compliance problems. The sentiment is that the yardstick should be not "does it compile" but "does the output exactly match the original."
OpenAI itself positions this as a capability update rather than a product announcement, and it's reported that the viability of production operation and enterprise adoption remains unresolved. Even if migration is left to AI, the realistic reading is that a process in which humans verify behavioral equivalence can't be dropped for the time being.
References: OpenAI, "Advancing coding in low-resource programming languages" / Interesting Engineering / Crypto Briefing / RuntimeWire
