OpenAI Says GPT-6.1 Sol Hits 97.9% on COBOL Migration

OpenAI on October 19, 2026 published research on advancing coding in low-resource programming languages, reporting that its latest models are substantially better at building and migrating software written in COBOL, Fortran, Visual Basic, OCaml and similar languages that have far less public training data than Python or JavaScript. The company paired benchmark scores with a plan to release evaluation tooling on GitHub and to offer the models through its API.

Key facts

What OpenAI announced

The post defines low-resource languages as those with less public code and documentation than Python or JavaScript, even though they still underpin critical systems. OpenAI says enterprise developers asked for better support where documentation is incomplete, original authors have left, and business rules live only in source.

OpenAI describes two evaluations—building a database engine from scratch across 20 languages, and migrating COBOL into modern languages with strict behavioral checks—and says it will release MultiLangBench on GitHub while offering the latest models through its API.

Why COBOL and Fortran still matter

In its post, OpenAI cites outside research on how deeply COBOL and Fortran remain embedded in regulated and safety-critical work. OpenAI cites these third-party estimates, linking to external sources: COBOL supports more than 40% of online banking systems, 80% of in-person credit card transactions, and 95% of ATM transactions; Fortran appears in weather forecasting and nuclear reactor safety analysis.

Independent reporting adds workforce pressure. Engadget notes that migrating large COBOL estates can cost banks millions to hundreds of millions of dollars and considerable calendar time, while the pool of experienced COBOL programmers is aging. OpenAI's announcement does not promise instant replacement of those programs; it argues models can help teams understand, extend, and translate code that is expensive to maintain by hand.

How the two tests work

MultiLangBench, as OpenAI describes it, requires a model to build a SQLite-compatible database engine from scratch in a specified language. Twenty languages are in scope. Database files serve as the shared interface, and a reference SQLite implementation checks results. Tests are adapted from SQLite's own test suite. The benchmark measures test pass rate and API cost per sample.

The COBOL migration evaluation asks models to move COBOL programs into C++, Java, Kotlin, Python, or Rust while preserving behavior. Hidden tests require byte-for-byte parity on observable outputs. A migration counts only if all tests pass—partial rewrites do not score.

OpenAI illustrates why parity matters with COBOL fixed-point arithmetic: excess decimal digits may be discarded unless rounding is explicitly requested. A rewrite that rounds instead can run without errors yet produce different financial results—exactly the class of subtle defect migration projects fear.

The numbers

All figures below come from OpenAI's own testing as reported in its October 19 post, unless noted.

OpenAI reports per-language MultiLangBench gains for GPT-6 Astra at maximum effort, including COBOL 10.9% → 72.1%, OCaml 25.6% → 85.8%, Fortran 21.4% → 83.5%, and Visual Basic 27.5% → 89.4%. Astra scores above 83% in 19 of 20 languages; COBOL remains the hardest at 72%.

For GPT-6 Luna at maximum effort, OpenAI lists smaller lifts on the same four languages—for example COBOL 10.9% → 17.1% and Visual Basic 27.5% → 41.6%—at an estimated average API cost of $0.77 per sample versus $3.48 for GPT-5.6 Sol (about 78% lower).

What it costs

On COBOL migration at low reasoning effort, OpenAI says GPT-6.1 Sol reaches 86.6% at about 61% of the API cost of GPT-5.6 Sol (or 39% less). Luna's MultiLangBench average of 40.3% comes with an estimated $0.77 per sample versus $3.48 for GPT-5.6 Sol.

The caveats

OpenAI explicitly warns that applying these tools to production software still requires expert review and testing against a system's business rules and dependencies. High benchmark scores do not remove compliance, audit, or operational sign-off.

Analysis (AImagazine). AImagazine notes that OpenAI's COBOL migration benchmark is not public, so the 86.6% and 97.9% migration results cannot be independently reproduced. The public MultiLangBench repository (Apache 2.0) README says evaluations should use 272 tests by default, and AImagazine reports that OpenAI's post does not specify which test set its scores used. When AImagazine checked, the repo contained tests and scripts but not per-model results.

What it means for banks and developers

Analysis. Strong benchmark scores could help teams draft migrations and surface hidden rules before experts sign off, but OpenAI's fixed-point example shows that running code is not the same as matching decades of ledger behavior. Engadget's reporting suggests AI may complement—not replace—costly multi-year rewrites as COBOL expertise shrinks.

The update also lands amid broader OpenAI platform changes, including the GPT-4o and GPT-4.1 API shutdown on Jan. 14, 2027 and GPT-6 Astra Ultrafast on NVIDIA Blackwell.

What to watch next

• GitHub release — whether MultiLangBench artifacts and per-language results appear as OpenAI promised, and whether third parties replicate Astra's MultiLangBench curves on the 272-test default.
• COBOL migration benchmark — any move to open the hidden migration suite so the 97.9% figure can be verified outside OpenAI's lab.
• Enterprise pilots — banks and government agencies testing migration assistants under audit controls, not just demo repositories.
• Model access — API pricing and reasoning-effort defaults for GPT-6.1 Sol versus Luna on long COBOL files.

Sources

• OpenAI: Advancing coding in low-resource programming languages — Oct. 19, 2026
• Engadget: OpenAI says its models can help banks migrate from old COBOL programs — Oct. 20, 2026
• AImagazine: OpenAI says GPT-6.1 Sol went from 5.4% to 86.6% on COBOL migration, but that test isn't public and its open benchmark grades just 272 tests — Oct. 20, 2026