OpenAI says its latest artificial intelligence models have become significantly better at writing and modernising software in older programming languages, including COBOL, Fortran and Visual Basic. In an announcement published on October 19, 2026, the company reported that GPT-6 Astra achieved an average test pass rate of 87.4% across 20 programming languages, compared with 26.8% for GPT-5.6 Sol. A separate evaluation found that GPT-6.1 Sol achieved 97.9% accuracy on a COBOL migration test when operating at higher reasoning effort.
The development targets a longstanding challenge for banks, government agencies, scientific institutions and large businesses: maintaining software built decades ago. These systems can be difficult to update because documentation is incomplete, original developers may no longer be available, and essential business rules can be buried in existing code. OpenAI believes stronger coding models could help organisations modernise these systems more efficiently, although the company acknowledges that expert review and extensive testing remain necessary before AI-generated migrations are used in production.
Key Takeaways
- New model capabilities: OpenAI reports improvements in writing and migrating software written in COBOL, Fortran, Visual Basic and other less commonly represented languages.
- 87.4% pass rate: GPT-6 Astra achieved this average across 20 languages in OpenAI’s MultiLangBench evaluation.
- 97.9% migration accuracy: GPT-6.1 Sol reached this result at higher reasoning effort on OpenAI’s COBOL migration test.
- Legacy systems: COBOL remains important in banking and financial transactions, while Fortran is used in scientific computing and weather forecasting.
- Lower migration risk: AI can help preserve existing software behaviour, but incorrect conversions can still create serious problems.
- Human oversight remains essential: OpenAI’s results come from its own evaluations and do not establish that every real-world migration can be automated safely.
Why OpenAI Is Targeting Older Programming Languages
Programming languages such as COBOL and Fortran were developed decades ago, before modern software engineering tools and today’s AI coding assistants became widely available.
Despite their age, these languages continue to support important systems. COBOL is associated with transaction processing, financial records and business applications. Fortran remains relevant in scientific computing, numerical modelling, engineering and weather forecasting.
The difficulty is not simply that these languages look different from Python or JavaScript. Organisations often have enormous codebases that have evolved over many years, with dependencies and business rules that are not fully documented.
A bank might have a transaction-processing system whose calculations reflect years of regulatory requirements and operational decisions. A small change to rounding, data handling or error recovery could produce different results even if the rewritten program appears to run correctly.
This makes legacy modernisation expensive and time-consuming. Engineers must understand the existing application, identify its dependencies, reproduce its behaviour, run tests and verify that the replacement system meets the organisation’s requirements.
OpenAI says developers and enterprise customers have reported that coding models often struggle with languages that have comparatively little training material available. Its latest evaluations are designed to measure progress on these harder tasks.
GPT-6 Astra Scores 87.4% Across 20 Programming Languages
OpenAI’s first evaluation, called MultiLangBench, measures whether an AI model can build a SQLite-compatible database engine from scratch in a specified programming language.
The task goes beyond producing a short code snippet. A model must organise a codebase, manage state, handle errors, implement functionality and preserve behaviour as the project develops.
OpenAI reported the following average test pass rates at maximum reasoning effort:
| Model | Average pass rate across 20 languages |
|---|---|
| GPT-5.6 Sol | 26.8% |
| GPT-6 Astra | 87.4% |
| GPT-6 Luna | 40.3% |
Source: OpenAI’s MultiLangBench evaluation. Results are company-reported and reflect the specific benchmark setup.
The difference between GPT-5.6 Sol and GPT-6 Astra is 60.6 percentage points. The improvement suggests that the newer model performed substantially better on the benchmark’s software-building tasks.
OpenAI also reported gains across every language included in the evaluation. For example, the reported pass rate for COBOL increased from 10.9% with GPT-5.6 Sol to 72.1% with GPT-6 Astra. For Fortran, the score rose from 21.4% to 83.5%, while Visual Basic improved from 27.5% to 89.4%.
These results matter because they indicate that the improvements are not limited to popular languages such as Python and JavaScript. However, the benchmark measures performance on a defined database-building task; it should not be interpreted as a universal measure of proficiency in every language or software project.
Why MultiLangBench matters
A shared benchmark makes it possible to compare models across languages using a consistent testing approach. OpenAI says it is releasing the MultiLangBench setup and test suite on GitHub, allowing developers to inspect the evaluation and test the models against the same tasks.
Independent replication will be important in determining how well the results generalise to different projects, codebases and development environments.
GPT-6.1 Sol Reaches 97.9% on COBOL Migration
OpenAI’s second evaluation focuses specifically on migrating existing COBOL programs into modern languages.
The test asks models to translate COBOL code into C++, Java, Kotlin, Python or Rust while preserving the original program’s observable behaviour. Hidden tests check whether the translated program produces matching outputs, and a migration counts as successful only when all required tests pass.
OpenAI reported these results:
| Model and reasoning setting | Reported migration accuracy |
|---|---|
| GPT-5.6 Sol, low reasoning effort | 5.4% |
| GPT-6.1 Sol, low reasoning effort | 86.6% |
| GPT-6.1 Sol, higher reasoning effort | 97.9% |
The difference between the two models at low reasoning effort is 81.2 percentage points. OpenAI also said GPT-6.1 Sol achieved its 86.6% result at approximately 61% of the API cost of GPT-5.6 Sol under the comparison described in its announcement.
At higher reasoning effort, GPT-6.1 Sol approached a near-perfect result on this particular evaluation. That does not mean the model can migrate every COBOL application with 97.9% accuracy in real-world settings. The result applies to the test suite and migration conditions used by OpenAI.
The evaluation nevertheless points to a potentially useful application of AI: helping engineers translate legacy code while checking whether the replacement preserves the original behaviour.
Why Small Coding Differences Can Create Major Problems
One of the most important challenges in legacy migration is preserving details that may seem insignificant to a programmer but have substantial business consequences.
OpenAI highlights COBOL’s handling of fixed-point arithmetic as an example. Depending on how a program is written, excess decimal digits may be discarded unless rounding is explicitly requested.
If an AI system rewrites that logic using different rounding behaviour, the new program may execute without errors but generate different financial results.
The same problem can arise with missing records, date handling, error conditions and other rules embedded in an established system. A migration can appear successful during a basic demonstration while still changing the behaviour on unusual inputs.
This is why output parity matters. By comparing the old and new programs against the same test cases, developers can identify differences that ordinary code review might overlook.
However, matching a set of tests is not proof that every possible business rule has been preserved. Real production systems can contain undocumented dependencies, rare edge cases and external integrations that are not represented in a benchmark.
AI can reduce the amount of manual work required to understand and rewrite code, but it cannot eliminate the need to establish what the original system is supposed to do.
What This Means for Banks and Large Enterprises
Banks and other large organisations have strong reasons to modernise legacy systems. Older applications can be expensive to maintain, difficult to integrate with modern platforms and dependent on a limited pool of experienced engineers.
AI coding tools could help these organisations analyse unfamiliar repositories, explain complicated functions, generate test cases, identify dependencies and prepare initial versions of replacement code.
For a bank, this could shorten parts of a migration project that would otherwise require engineers to study large volumes of old code manually. Developers could spend more time validating business requirements and less time on repetitive translation work.
Similar opportunities exist in insurance, government services, industrial systems and scientific computing. Fortran applications, for example, can be part of numerical simulations and research workflows that organisations need to maintain over long periods.
Nevertheless, legacy modernisation is not simply a matter of converting one programming language into another. Organisations must also consider security, performance, regulatory requirements, data integrity, deployment procedures and the availability of rollback plans.
The most realistic near-term use is therefore AI-assisted modernisation, where models handle parts of the work while experienced engineers remain responsible for verification and release decisions.
AI Coding Is Expanding Beyond Popular Languages
The announcement reflects a broader effort to make AI coding tools useful across more of the software industry.
Many coding assistants perform well on popular programming languages because large quantities of public code and documentation are available for training and evaluation. Older or less widely represented languages can be harder for models to handle reliably.
Improving performance in these areas could make AI tools more useful to organisations whose technology cannot easily be replaced. It may also help developers who need to work with unfamiliar codebases or maintain systems written long before modern development practices became common.
OpenAI also reported improvements in the estimated API cost of certain benchmark tasks. Its GPT-6 Luna model achieved a 40.3% average pass rate on MultiLangBench at an estimated average cost of $0.77 per sample, compared with $3.48 for GPT-5.6 Sol in the reported comparison.
The results suggest that model selection can involve a trade-off between performance and cost. A business handling a complex migration may prefer a stronger model for high-risk tasks, while using a less expensive model for simpler code analysis or documentation.
These are benchmark-level cost estimates, not a complete estimate of the cost of migrating a production system. Engineering time, testing infrastructure, security reviews and deployment work can remain substantial expenses.
The Bigger Picture
OpenAI’s latest results suggest that AI coding models are becoming more capable of working with programming languages that have received less attention in modern AI development. Improvements in COBOL migration and the ability to build software across 20 languages could help organisations address technical debt and preserve important applications without relying entirely on manual rewriting.
The central limitation is that software modernisation requires more than generating syntactically correct code. It requires preserving business logic, validating edge cases and proving that critical systems continue to behave as expected. OpenAI’s benchmarks provide evidence of progress, but independent testing and real-world deployment experience will determine how broadly those gains can be applied.
Looking Ahead
OpenAI has made MultiLangBench available for developers to examine, creating an opportunity for independent comparisons across programming languages. The next important step will be to see whether the reported gains hold up on larger enterprise codebases, unfamiliar applications and tests designed by organisations outside OpenAI. Developers and technology leaders will also need clearer evidence of how the models perform on security, reliability and long-term maintenance tasks.
For companies running legacy systems, the announcement offers a potential route to faster modernisation rather than an immediate replacement for experienced programmers. AI may take on more code analysis, translation and test generation, while engineers continue to define requirements and validate results. In banking and other critical environments, the value of these tools will ultimately depend on whether they can reduce modernisation time without changing the behaviour on which customers and businesses rely.
Frequently Asked Questions
1. What did OpenAI announce about COBOL and Fortran?
OpenAI reported that its latest models perform better at building software across multiple programming languages and at migrating COBOL programs to modern languages while preserving their behaviour.
2. What score did GPT-6 Astra achieve?
GPT-6 Astra achieved an average test pass rate of 87.4% across 20 programming languages in OpenAI’s MultiLangBench evaluation, compared with 26.8% for GPT-5.6 Sol.
3. How accurate is GPT-6.1 Sol at COBOL migration?
OpenAI reported 86.6% accuracy at low reasoning effort and 97.9% at higher reasoning effort on its COBOL migration evaluation.
4. Why is COBOL still important?
COBOL remains in use in banking, financial transactions and large business systems. Replacing these applications can be difficult because their business rules and dependencies have accumulated over decades.
5. Can AI safely replace COBOL systems without human developers?
Not on the strength of these results alone. AI can help translate and test code, but production migrations still require expert review, validation against business rules and careful deployment.
6. Are OpenAI’s results independently verified?
The headline results are from evaluations designed and reported by OpenAI. The company has released MultiLangBench for public inspection, but the reported scores should not be treated as independent proof of performance across all real-world software projects.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
