BlockBeats says OpenAI released GPT-6 series COBOL migration test results

AT A GLANCE

The report states that the highest COBOL migration-test success rate for OpenAI models reached 97.9%, but lacks official materials for verification.

Article

Anthropic shakes things up IBM Afterwards, OpenAI also posted a near-perfect COBOL migration result

Beating Insights AI Breaking news: This year,February after Anthropic promoted Claude Code's ability to modernize COBOL legacy systems,IBM's share price plunged in a single day. 13.2%。

8 months later, OpenAI also released the latest results for its GPT-6 series on legacy programming languages, with the highest success rate in COBOL code-migration tests reaching 97.9%。

COBOL was created in the 50 s and is still widely used in banking transactions,ATM and insurance systems. Many programs have been running for decades; their original developers have long since left, and some business rules exist only in the code. Rewriting them in modern languages requires not only making the code run, but also ensuring that calculation results are exactly identical. For example,COBOL in fixed-point arithmetic, if rounding is not explicitly required, excess decimal places are simply discarded. If a new program rounds them instead, it could cause errors in financial accounts.

OpenAI specifically tested AI's ability to rewrite COBOL programs into languages such as C++, Java, Python, and Rust, requiring the test outputs of the old and new programs to be byte-for-byte identical.GPT-6.1 Sol achieved a success rate of 86.6%at a low reasoning setting, while the previous-generation GPT-5.6 Sol achieved only 5.4%, with estimated API costs also reduced by 39%. At a higher reasoning setting, the success rate further reached 97.9%.

OpenAI also open-sourced MultiLangBench, requiring AI to each use 20 programming languages to build a SQLite-compatible database engine from scratch.GPT-6 Astra achieved an average test pass rate of 87.4%，GPT-5.6 Sol achieved 26.8%. Among them, the models respectively scored COBOL and 72.1%on 10.9%, which is also the most difficult of the 20 languages.

Original link https://m.theblockbeats.info/flash/372684

Key points

The report states that this year,February after Anthropic promoted Claude Code's ability to modernize COBOL legacy systems,IBM's share price fell in a single day 13.2%。
The report states that, 8 months later, OpenAI released GPT-6 series results, with the COBOL highest success rate in code-migration tests being 97.9%。
Test requirements AI Rewrite COBOL programs into languages such as C++, Java, Python, and Rust, with the test outputs of the old and new programs required to be byte-for-byte identical.
The report states that,GPT-6.1 Sol had a success rate of 86.6%，GPT-5.6 Sol had 5.4%, with estimated API costs reduced by 39%; at a higher reasoning setting, the success rate was 97.9%。
The report states that OpenAI open-sourced MultiLangBench, requiring AI to each use 20 programming languages to build a SQLite-compatible database engine from scratch.
The report states that,GPT-6 Astra and GPT-5.6 Sol achieved average test pass rates on MultiLangBench of 87.4%and 26.8%, respectively, and on COBOL achieved 72.1%and 10.9%。

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This test examines whether AI can rewrite legacy COBOL programs into other languages while keeping test outputs byte-for-byte identical. The original article illustrates that, in fixed-point processing, the difference between “directly discarding decimal places” and “rounding” may cause errors in financial accounts; therefore, code that runs does not necessarily mean the migration is correct.

Why it matters to readers

The original article states that COBOL is still used in banking transactions,ATM and insurance systems, and that some business rules exist only in legacy code. If the results can be reproduced, they may indicate that AI has made significant progress in legacy-system migration.

Beginners can understand the success rate as the proportion of tasks the model completed in the specified tests, rather than meaning that all real-world COBOL systems can be safely migrated at the same rate; particular attention should be paid to the different sets of results: low reasoning setting, higher reasoning setting, and MultiLangBench.

Risks and unknowns

The evidence consists only of a media news brief and does not include an official OpenAI announcement, complete testing methodology, or raw results.
GPT-6 Whether the series, MultiLangBench, and the various results can be replicated through independent evaluation cannot be confirmed from the evidence.
39%of API The cost reduction is an estimate, and the calculation methodology is not stated.

BlockBeats says OpenAI released GPT-6 series COBOL migration test results

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center