# OpenAI COBOL Update: GPT-6.1 Sol Hits 97.9% on Legacy Code

URL: https://whatisagenticai.net/blog/openai-cobol/
Published: 2026-10-20
Publisher: Little Tough LLC, maker of AI Agent Assistant: Agentic AI (https://apps.apple.com/us/app/ai-agent-assistant-agentic-ai/id6761501260)

OpenAI COBOL news: on October 19, 2026, OpenAI said GPT-6.1 Sol migrated 97.9% of COBOL test programs. What changed, what MultiLangBench is, and the limits.

OpenAI COBOL is suddenly a real search, and for a good reason. On October 19, 2026, OpenAI said its newest models have become much better at reading, writing and moving old code in languages like COBOL, Fortran and Visual Basic. That is the 60-year-old code still running a large share of the world’s banks, insurers and government offices.

So what did OpenAI actually show, which models are involved, and does this mean AI can now rewrite a bank on its own? Here is the plain-English version, with the numbers OpenAI published and the catches it admitted.

**TL;DR:** On October 19, 2026, OpenAI said its latest models are far better at "low-resource" programming languages such as COBOL, Fortran and Visual Basic. On OpenAI’s own COBOL migration test, GPT-6.1 Sol correctly moved 97.9% of programs into modern languages like Java or Python at high reasoning effort, and 86.6% at low effort. On a new suite called MultiLangBench, GPT-6 Astra scored 87.4% on average, up from 26.8% for GPT-5.6 Sol. These are OpenAI’s own tests, and OpenAI says real production code still needs expert review.

## What did OpenAI announce about COBOL?

OpenAI posted the update on X and in a blog post about "low-resource programming languages" on October 19. A low-resource language is one with very little public code for an AI to learn from. Python and JavaScript are everywhere on the internet. COBOL, a language from 1959, mostly lives inside private company systems, so AI models have historically been weak at it.

OpenAI’s claim is that this gap has largely closed. It says the gains came from feedback from developers and enterprise customers, and it framed the work around a real business problem: according to OpenAI, about 40% of the world’s banks still rely on COBOL systems, and moving off them can cost "millions to hundreds of millions" of dollars. On top of that, the people who know COBOL best are retiring.

Notably, this was not the launch of a new model. It was new test results for models OpenAI already ships, mainly GPT-6.1 Sol and GPT-6 Astra.

## OpenAI COBOL results: the numbers

OpenAI shared two kinds of tests. The first asks a model to **migrate** a COBOL program into C++, Java, Kotlin, Python or Rust without changing what it does. A migration only counts if every hidden test passes, including checks that the new program gives the same output as the old one. The second, MultiLangBench, asks a model to **build** something hard from scratch in many languages.

| Test | Model | Result |
| --- | --- | --- |
| COBOL migration (high reasoning effort) | GPT-6.1 Sol | 97.9% |
| COBOL migration (low reasoning effort) | GPT-6.1 Sol | 86.6% |
| MultiLangBench, average of 20 languages | GPT-6 Astra | 87.4% |
| MultiLangBench, average of 20 languages | GPT-5.6 Sol | 26.8% |
| MultiLangBench, COBOL only | GPT-6 Astra | 72.1% (up from 10.9%) |

Two things stand out. First, the jump is large: GPT-5.6 Sol came out in July 2026, and its successors more than tripled its MultiLangBench average in about three months. Second, COBOL is still the hardest language in the suite, even for GPT-6 Astra. "Much better" is not the same as "solved."

## What is MultiLangBench?

MultiLangBench is OpenAI’s own evaluation suite. Each task asks a model to build a working, SQLite-compatible database engine in a given programming language. SQLite is the small database that sits inside nearly every phone and browser, so copying its behavior is a demanding, many-step coding job. The suite covers 20 languages, from popular ones to old ones like COBOL, Fortran and Visual Basic.

Because OpenAI built and graded it, outside researchers have not yet reproduced these scores. Treat them as the company’s launch claims, the same way you would treat any benchmark a company publishes about itself.

## Can AI convert COBOL to Java or Python now?

Partly, and with supervision. The OpenAI test moved small-to-medium COBOL programs that came with tests. Real bank systems are bigger and messier: millions of lines, missing documentation, and business rules nobody wrote down. OpenAI itself says that using these models on production software "still requires expert review and testing."

A realistic way to picture it is an agent loop, the same idea we cover in [how AI agents work](https://whatisagenticai.net/blog/how-ai-agents-work/):

- **Read and explain.** The agent reads a COBOL program and writes down what it does in plain English.

- **Write tests.** It captures how the old program behaves, so you have something to check against.

- **Translate.** It writes the new Java, Python or Rust version.

- **Check and repeat.** It runs the tests, fixes what fails, and tries again until everything passes.

- **Human sign-off.** An engineer reviews the result before it touches real money.

That last step is not optional. A 97.9% pass rate on a test suite still means some programs fail, and in banking, one wrong rounding rule can be costly.

## Which OpenAI models handle COBOL?

The headline model is GPT-6.1 Sol, which OpenAI introduced at DevDay on September 29, 2026. It is available to developers in the API as gpt-6.1-sol at $2 per million input tokens and $10 per million output tokens, and in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. GPT-6 Astra is OpenAI’s larger top-tier model and posted the best MultiLangBench average.

The naming follows the tier system OpenAI started with GPT-5.6, where Sol, Terra and Luna mark capability levels. Our [GPT-5.6 explainer](https://whatisagenticai.net/blog/what-is-gpt-5-6/) walks through that naming if it still feels confusing.

## OpenAI vs Anthropic vs IBM on COBOL

OpenAI is not alone here. Anthropic published its own case for using Claude on COBOL modernization in February 2026, and IBM has sold watsonx Code Assistant for Z, which helps turn mainframe COBOL into Java, for years. What is new in OpenAI’s update is the size of the published jump and a specific migration score. No neutral, head-to-head test of all three on the same COBOL code has been published yet, so any "who is best" claim today is premature.

## What this means for you

If you are not a bank engineer, the bigger signal is this: AI is getting good at the boring, careful, many-step work that used to need rare experts. That is exactly what agentic AI means in practice. The model does not just answer a question. It reads, plans, tries, checks its own work and hands you something to approve.

You can use the same pattern on a much smaller scale. Paste in a confusing old spreadsheet formula, a VBA macro from a coworker who left, or a script you inherited, and ask for a plain-English explanation and a safer modern version. **Agentic AI** on iPhone gives you a Code Helper agent for exactly that, plus a team of up to four experts that can review each other’s work, which is why [two agents beat one](https://whatisagenticai.net/blog/multi-agent-teams/) for anything you plan to trust. No account needed, and your first 3 tasks are free.

## Quick answers

**What did OpenAI announce about COBOL?** On October 19, 2026, OpenAI said its newest models are much better at building and migrating software in low-resource languages like COBOL, Fortran and Visual Basic, and shared benchmark results to back it up.

**How accurate is GPT-6.1 Sol at COBOL migration?** On OpenAI’s own test, it migrated 97.9% of COBOL programs correctly at high reasoning effort and 86.6% at low effort. Every hidden test had to pass for a program to count.

**Can ChatGPT write COBOL?** Yes. ChatGPT and Codex can write and explain COBOL, and paid plans can use GPT-6.1 Sol in ChatGPT Work and Codex. Always test generated COBOL before using it on real systems.

**Can AI convert COBOL to Java?** Java is one of the five target languages in OpenAI’s migration test, along with C++, Kotlin, Python and Rust. It works best with tests in place and an expert reviewing the output.

**Is MultiLangBench an independent benchmark?** No. It is OpenAI’s own evaluation suite, so its scores have not been independently verified yet.

**Will AI replace COBOL programmers?** Not soon. OpenAI says production use still needs expert review, and the people who know how a bank’s rules really work become more valuable, not less. More terms are in our [agentic AI glossary](https://whatisagenticai.net/agentic-ai-glossary/).

**Sources:** [OpenAI: low-resource programming languages](https://openai.com/index/low-resource-programming-languages/) · [OpenAI on X](https://x.com/OpenAI/status/2112310345863930168) · [Engadget](https://www.engadget.com/2291457/openai-says-its-models-can-translate-cobol/) · [Interesting Engineering](https://interestingengineering.com/ai-robotics/openai-coding-models-software-modernization) · [Crypto Briefing](https://cryptobriefing.com/openai-upgrades-model-cobol-fortran-legacy-code/) · [Unite.AI: GPT-6.1 Sol at DevDay](https://www.unite.ai/openai-unveils-gpt-6.1-sol-at-devday-with-new-codex-and-chatgpt-tools/)