"GPT-6.1 Sol achieves 86.6% accuracy at low reasoning effort, compared with GPT-5.6 Sol's 5.4% at the same setting at just 61% of the API cost," OpenAI reported.
MultiLangBench: a single test to stress multilingual coding
OpenAI measured progress using a bespoke benchmark it calls MultiLangBench, which "asks models to build a SQLite-compatible database engine from scratch across 20 programming languages." That single task is intended to stress a model's ability to reason about algorithms, manage low-level language details and produce fully working implementations in languages with widely varying ecosystems and documentation availability.
The results OpenAI published show sharp gains across the board. GPT-6 Astra reached an average test pass rate of 87.4%, compared with 26.8% for GPT-5.6 Sol. OpenAI presented these figures as evidence that newer models handle legacy and less-common languages far better than previous-generation models.
Large jumps in COBOL, Fortran, Visual Basic and OCaml
OpenAI highlighted especially large improvements in languages often described as "legacy" but still central to critical services. In COBOL, Astra jumped from a 10.9% pass rate for GPT-5.6 to 72.1%; Fortran improved from 21.4% to 83.5%; OCaml rose from 25.6% to 85.8%; and Visual Basic went from 27.5% to 89.4%.
OpenAI underscored why that matters. "These languages have less code and documentation available for models to learn from than languages like Python and JavaScript, yet support critical infrastructure and essential services around the world," the company wrote, and went on to note concrete dependencies: COBOL supports "more than 40% of online banking systems, 80% of in-person credit card transactions, and 95% of ATM transactions," while Fortran "is still used in areas such as weather forecasting and nuclear reactor safety analysis."
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleGPT-6.1 Sol and the practical task of migrating COBOL
OpenAI separately evaluated migration tasks — translating COBOL programs into modern languages without changing behavior — and reported striking gains for GPT-6.1 Sol. At a "low reasoning effort" setting, the model achieved 86.6% accuracy, versus GPT-5.6 Sol's 5.4% at the same setting, and at "just 61% of the API cost." Pushing the model to a higher reasoning effort raised accuracy to 97.9%, which OpenAI described as bringing the evaluation "close to saturation."
OpenAI framed these migration results as potentially useful for companies seeking to modernize older software, while also cautioning that "production migrations would still require expert review and testing." In other words, the models may automate large portions of translation work but are not presented as a drop-in replacement for human verification in safety- or money-critical systems.
What this means for technologists, affected enterprises, and security teams
- Technologists and migration teams: The tests suggest large language models may substantially reduce the manual effort needed to port code from COBOL, Fortran, Visual Basic and similar languages into modern targets like C++, Java, Kotlin, Python or Rust — especially as models hit pass rates above 80–90% in benchmarked tasks. Teams will still need to validate behavior and performance in production.
- Affected enterprises and procurement leaders: Organizations that rely on legacy stacks — notably financial institutions cited by OpenAI as heavy COBOL users — gain a new automation option for modernization projects, potentially lowering initial development cost (OpenAI reported a 61% API-cost baseline in one comparison). Procurement decisions will still need to factor in expert review costs and risk controls.
- Security teams: Because OpenAI emphasized the necessity of expert review for production migrations, security and assurance teams should treat model-produced translations as candidate artifacts requiring the same testing, verification and compliance checks applied to human-written migrations.
Closing observation
OpenAI's numbers show a rapid improvement in models' ability to write and port code in languages that power real-world systems. The MultiLangBench results and the migration accuracy figures for GPT-6.1 Sol and GPT-6 Astra provide concrete benchmarks that organizations can use to evaluate whether automated assistance meaningfully shifts effort, cost and risk in their legacy-modernization programs. What remains explicit in OpenAI's own presentation is that even with high accuracy on benchmark tasks, the path to production — and to trusted, auditable systems — still runs through human review, testing and validation.
