Science & Tech · AI · 1 hr ago
OpenAI reports stronger results migrating legacy software
OpenAI says its latest models are better at working with older programming languages such as COBOL and Fortran.
These languages still support important systems, including banking and scientific computing.
Replacing their software is risky because rewritten programs must preserve the original calculations and business rules.
On OpenAI’s COBOL migration test, GPT-6.1 Sol reached 97.9% accuracy at higher reasoning effort; that score applies to the test, not every real-world migration.
In a separate test across 20 programming languages, GPT-6 Astra achieved an average pass rate of 87.4% for building a database engine compatible with SQLite.
OpenAI is releasing that test, MultiLangBench, on GitHub, while developers still need to check migrated systems against real workloads and dependencies.
OpenAI reported improved results from models tested on migrating legacy software and writing code in older programming languages.
GPT-6.1 Sol achieved 97.9 percent accuracy on OpenAI’s COBOL migration evaluation at high reasoning effort.
At low reasoning effort, GPT-6.1 Sol scored 86.6 percent, compared with 5.4 percent for GPT-5.6 Sol, while OpenAI reported a 39 percent reduction in API costs.
On MultiLangBench, GPT-6 Astra achieved an average test pass rate of 87.4 percent across 20 programming languages, up from 26.8 percent for GPT-5.6 Sol.
The reports say the benchmark results do not guarantee error-free migrations, and developers still need to validate modernized systems.
- Who
- OpenAI reported the results for its GPT-6.1 Sol, GPT-6 Astra and GPT-6 Luna models.
- What
- The models were tested on migrating COBOL programs and building a SQLite-compatible database engine across multiple programming languages.
- When
- The reports were published on October 20, 2026.
- Where
- The evaluations covered COBOL migration and 20 programming languages on MultiLangBench; the test location is not stated.
- Why
- The evaluations assess models’ ability to modernize legacy software and work with older programming languages.
This story does not have two clearly opposing sides.
This story does not have a timeline yet.
- COBOL migration accuracy
- GPT-6.1 Sol scored 97.9 percent at high reasoning effort.
- Low-effort COBOL migration accuracy
- GPT-6.1 Sol scored 86.6 percent; GPT-5.6 Sol scored 5.4 percent.
- Low-effort API costs
- OpenAI reported a 39 percent reduction.
- MultiLangBench average pass rate
- GPT-6 Astra scored 87.4 percent across 20 languages, up from 26.8 percent.
- COBOL on MultiLangBench
- GPT-6 Astra scored 72.1 percent, compared with 10.9 percent for GPT-5.6 Sol.