OpenAI’s GPT‑6 Shines on Legacy Code: COBOL, Fortran, VB Boosted

Introduction

Legacy systems are the hidden backbone of global finance, government, and manufacturing. Despite the hype around Python and JavaScript, languages like COBOL, Fortran, and Visual Basic still power billions of transactions every day. OpenAI’s latest model, GPT‑6, claims a quantum leap in understanding and generating code for these “low‑resource” languages, potentially reshaping how enterprises modernize legacy stacks.

What Happened

On October 20, 2026, The Decoder reported that OpenAI released detailed benchmarks showing GPT‑6’s superiority over GPT‑5.6 on legacy code tasks. The company published two new evaluation suites:

• MultiLangBench – a 20‑language suite that asks models to build a SQLite‑compatible database engine from scratch.
• COBOL Migration Benchmark – translates COBOL into modern languages (C++, Java, Kotlin, Python, Rust).

Results: GPT‑6 Astra achieved an 87.4 % average pass rate on MultiLangBench, up from 38 % for GPT‑5.6 Terra. GPT‑6 Luna scored 40.3 %, matching GPT‑5.6 Terra but at a fraction of the compute cost. On the COBOL migration test, GPT‑6.1 Sol reached 86.6 % accuracy with low reasoning effort and 97.9 % when the model was allowed to think longer.

Key Details

• Low‑resource focus – OpenAI explicitly targeted languages with scarce training data, citing IBM’s estimate that COBOL runs 40 % of online banking and 95 % of ATM transactions.
• Two model families – Astra (high‑performance, higher cost) and Luna (cost‑optimized) give enterprises flexibility.
• Open‑source benchmark – MultiLangBench is hosted on GitHub, encouraging third‑party validation.
• Reasoning effort scaling – Accuracy climbs dramatically when the model is allowed more “thought cycles,” a new knob for developers.
• Cross‑language migration – GPT‑6.1 Sol can output idiomatic code in five modern languages, easing gradual rewrites.

Technical Analysis

OpenAI’s breakthrough stems from three engineering advances:

1. Synthetic data generation – The team created massive, high‑quality code corpora for COBOL, Fortran, and VB using rule‑based translators and human‑in‑the‑loop validation. This mitigated the scarcity problem that plagued earlier models.
2. Token‑level curriculum learning – Training started with simple arithmetic and file‑I/O snippets, progressively introducing complex control‑flow constructs native to legacy languages (e.g., COBOL’s PERFORM loops, Fortran’s DO statements).
3. Extended context windows – GPT‑6 supports up to 128 k tokens, allowing it to keep entire legacy modules in memory, reason about inter‑module dependencies, and generate coherent refactorings.

The reasoning‑effort parameter is a novel addition. By allocating more compute cycles per token, the model can perform deeper symbolic reasoning, akin to a programmer stepping through code line‑by‑line. This explains the jump from 86 % to 98 % accuracy on the COBOL benchmark.

Industry Impact

• Banking & finance – Faster, automated COBOL-to‑Java migrations reduce reliance on dwindling mainframe talent, cutting operational risk.
• Government & defense – Legacy Fortran simulations can be modernized without rewriting physics models from scratch, preserving scientific fidelity.
• Enterprise software – Companies using Visual Basic for internal tools can now auto‑generate .NET Core equivalents, extending product lifespans.
• AI vendor competition – Google DeepMind and Anthropic are likely to accelerate their own low‑resource language research to avoid losing market share in the lucrative legacy‑code segment.

Future Implications

OpenAI’s focus on “low‑resource” programming languages signals a strategic pivot: AI is no longer just a productivity booster for new projects; it becomes a maintenance engine for existing, mission‑critical code. This could spawn a new market for AI‑driven legacy‑code audit services, compliance tools, and automated security hardening. Moreover, the success of synthetic data pipelines may inspire similar approaches for niche domains such as embedded C, PLC ladder logic, or even legacy hardware description languages.

Conclusion

GPT‑6’s leap in legacy language competence is more than a technical footnote—it’s a potential catalyst for the largest, most risk‑averse segment of the software industry to finally embrace AI‑assisted modernization. Enterprises that act now can lock in cost savings, talent independence, and regulatory compliance, while competitors that ignore the shift may find themselves stranded on aging mainframes.