OpenAI says GPT-6 nearly doubles test pass rates on legacy languages

OpenAI reports that GPT-6 Astra hits an 87.4 percent test pass rate across 20 programming languages on the company's new MultiLangBench benchmark, up from roughly 38 percent for its predecessor, GPT-5.6 Terra. The gains are concentrated in what the company calls low-resource languages: COBOL, Fortran, and Visual Basic.

The second number matters even more for one specific job. OpenAI says GPT-6.1 Sol reaches up to 97.9 percent accuracy on COBOL migration tasks. That is the kind of result that could matter for modernizing critical financial infrastructure, where COBOL still runs core systems.

Why do legacy languages trip up AI models?

Large language models learn from training data, and that data skews heavily toward languages with a large living developer base. Python, JavaScript, and C++ generate enormous amounts of public code, documentation, and discussion. COBOL and Fortran do not.

That imbalance has a practical consequence. A model can be excellent at Python yet unreliable when asked to read, explain, or rewrite a 40-year-old COBOL program. Benchmarks historically reflected the same bias, because they tested what most developers use. OpenAI's MultiLangBench, covering 20 languages, is an attempt to measure performance in the places where mainstream benchmarks go quiet.

The gap between 38 percent and 87.4 percent tells you how large that weakness used to be. GPT-5.6 Terra failed most of the tests in this broader evaluation. GPT-6 Astra passes the large majority of them.

The takeaway: OpenAI has substantially closed a known blind spot, at least by its own measurement.

What does the COBOL number actually mean?

The 97.9 percent figure from GPT-6.1 Sol applies specifically to COBOL migration. Migration here means translating aging COBOL code into modern languages, a task banks, insurers, and government agencies have struggled with for decades.

The stakes are real. The source report notes that this result could matter for modernizing critical financial infrastructure. COBOL systems still sit underneath payment processing, ledger management, and other operations where an incorrect translation is not a cosmetic bug — it is a financial error.

That is also why the number deserves scrutiny. A 97.9 percent accuracy claim on migration tasks, measured on a benchmark the model maker itself introduced, is promising but not independently verified. Buyers of AI-assisted migration will want third-party evaluation before trusting the remaining 2.1 percent.

The takeaway: near-perfect COBOL migration accuracy is the single most commercially significant claim in OpenAI's report.

How do the model names map to the results?

The announcement covers three named models:

• GPT-6 Astra: 87.4 percent test pass rate across 20 languages on MultiLangBench, up from about 38 percent for GPT-5.6 Terra.
• GPT-6.1 Sol: up to 97.9 percent accuracy on COBOL migration specifically.
• GPT-5.6 Terra: the prior generation, setting the baseline on the same benchmark.

The structure suggests specialization. Astra is the general performer across the language spread; Sol is the targeted tool for the highest-stakes legacy task. OpenAI has not, in this report, published comparable Sol figures for other languages.

The takeaway: OpenAI is packaging legacy-language capability both broadly and in a migration-specific model.

What happens next?

If these numbers hold up outside OpenAI's own benchmark, the practical effect lands on IT departments that have postponed mainframe modernization because skilled COBOL programmers are scarce and rewrite projects are slow and risky. A model that reliably reads and migrates COBOL changes the cost calculation for those projects.

Expect the pressure to move to verification: independent audits, real-world migration pilots, and comparisons against competing models on MultiLangBench. The 87.4 and 97.9 percent figures are OpenAI's claims today; the market will decide what they are worth tomorrow.