AI news ·
New models show major gains in low-resource programming languages like COBOL and Fortran
GPT-6 Astra builds a working SQLite database engine across 20 languages with an 87.4% average pass rate, up from 26.8%. COBOL alone supports 95% of ATM transactions worldwide and saw its score jump from 10.9% to 72.1%.
GPT-6 Astra now builds a working SQLite-compatible database engine in 20 programming languages with an average test pass rate of 87.4%, up from 26.8% for the previous generation. The jump in performance covers languages that power critical infrastructure - COBOL, Fortran, and Visual Basic - where training data is scarce but operational stakes are high. COBOL alone supports more than 40% of online banking systems, 80% of in-person credit card transactions, and 95% of ATM transactions worldwide.
The improvements arrive alongside a separate model, GPT-6.1 Sol, that migrates COBOL programs to modern languages with 86.6% accuracy at low reasoning effort and 97.9% at higher effort. Both advances target a persistent problem for enterprise teams: maintaining and modernizing software when documentation is thin, original developers have left, and business rules exist only in the source code.
These results were published October 19, 2026, alongside two new evaluations designed to measure capabilities that matter for legacy system work. The company said the work was guided by feedback from developers and enterprise customers who reported that coding models often struggle with low-resource languages.
Building SQLite from scratch across 20 languages
MultiLangBench tests whether a model can construct a SQLite-compatible database engine in a specified language. The model must organize a codebase, manage state, handle errors, and add features while preserving existing functionality. Database files serve as a shared interface: the candidate program receives an initial database and a SQL query, produces a modified database, and the reference SQLite implementation checks whether the result matches expected contents. This allows the same tests to apply regardless of the language used.
At maximum reasoning effort, GPT-6 Astra scored above 83% in 19 of 20 languages. COBOL remained the hardest at 72.1%, though that represents a leap from GPT-5.6 Sol's 10.9%. Other gains include OCaml (25.6% to 85.8%), Fortran (21.4% to 83.5%), and Visual Basic (27.5% to 89.4%).
GPT-6 Luna offers a lower-cost alternative that still outperforms GPT-5.6 Sol in every language. It reaches an average pass rate of 40.3% at an estimated API cost of $0.77 per sample, compared with $3.48 for GPT-5.6 Sol - roughly 78% less. The evaluation setup and test suite are available on GitHub for developers to inspect performance in the languages they use.
Migrating COBOL while preserving business logic
The second evaluation asks models to migrate COBOL programs to C++, Java, Kotlin, Python, or Rust while maintaining byte-for-byte parity on observable outputs. COBOL programs often process financial transactions where details like rounding rules encode business requirements. For example, COBOL's fixed-point arithmetic discards excess decimal digits unless rounding is explicitly requested. A rewrite that rounds those values instead can run without errors while producing different financial results.
GPT-6.1 Sol achieved 86.6% accuracy at low reasoning effort, compared with 5.4% for GPT-5.6 Sol at the same setting, while costing 39% less. At higher reasoning effort, accuracy reached 97.9%, nearly saturating the evaluation. The company emphasized that applying these capabilities to production software still requires expert review and testing against system-specific business rules and dependencies. For teams working to understand unfamiliar repositories and plan changes across languages, AI Coding Courses can help developers build the skills to evaluate and integrate these tools effectively.
Why this matters for developers, educators, and technical writers
The results change the practical economics of legacy modernization. A migration task that previously required extensive manual effort - and still produced correct results only 5% of the time with automated assistance - now succeeds at rates above 86% at lower cost. For developers maintaining Fortran-based weather forecasting systems or COBOL banking backends, this means fewer hours spent deciphering undocumented code. For educators and technical writers, the publicly released MultiLangBench test suite provides a concrete benchmark to discuss what AI coding tools can and cannot do across language ecosystems. The company said it will use direct feedback on where models struggle to shape future training priorities.