The result
OpenAI reports an average test-pass rate of 87.4% for GPT-6 Astra across its 20-language benchmark, while GPT-5.6 Sol scored 26.8%. In the COBOL subset, the company reports 72.1% for Astra and 10.9% for Sol. These are results from OpenAI’s evaluation, not independent measurements of production migrations.
Why this matters
Programming languages with fewer public examples can be harder for coding assistants to handle. OpenAI is making a case that model capability can extend beyond the languages that dominate online code. It also released MultiLangBench, an Apache-2.0 benchmark repository, so other teams can inspect the tasks and run comparisons.
How to read it
A benchmark score measures the supplied tasks and test harness. It does not establish that an AI can safely migrate a live codebase without review. OpenAI separately reports GPT-6.1 Sol results on a COBOL migration evaluation, including 86.6% low-effort and 97.9% high-effort results; those figures use a different evaluation setup and should not be treated as directly interchangeable with the 20-language average.
Source published 2026-10-19. Coverage is based on the maker’s announcement and demonstration.