started · updated
OpenAI improves legacy code migration with GPT-6 models
OpenAI has announced significant advancements in its ability to handle low-resource programming languages, specifically targeting the modernization of legacy systems. New models, including GPT-6 Astra and GPT-6 1 Sol, have demonstrated improved performance in writing and migrating code for languages such as COBOL, Fortran, and Visual Basic.
In the MultiLangBench evaluation, which requires building SQLite-compatible database engines from scratch across 20 languages, GPT-6 Astra achieved an average success rate of 87.4 percent, a substantial increase from the 26.8 percent achieved by GPT-5 6 Sol. Specifically, GPT-6 Astra reached a 72.1 percent success rate in COBOL, 83.5 percent in Fortran, and 89.4 percent in Visual Basic.
In migration tests designed to translate COBOL into modern languages like Java, Python, or Rust while preserving exact functional behavior, GPT-6 1 Sol achieved 86.6 percent accuracy at low reasoning effort and reached 97.9 percent accuracy at high reasoning effort. These improvements are critical for industries like banking and government, where COBOL supports a vast majority of ATM and credit card transactions. OpenAI also noted that the GPT-6 Luna model provides a more cost-effective option, with an estimated 78 percent reduction in API costs compared to previous models.
Entities
Anthropic · GPT-6 · GPT-6 1 Sol · GPT-6 Astra · IBM · OpenAI
Claims
What the coverage asserts, and how many sources carry each claim.
- [○ 1 SOURCE] The GPT-6 Luna model achieved an average success rate of 40.3 percent with an estimated 78 percent reduction in cost compared to GPT-5 6 Sol. geeksroom.com
- [● 4 SOURCES] GPT-6 1 Sol reached 86.6 percent accuracy at low reasoning effort, compared to 5.4 percent for GPT-5 6 Sol. interestingengineering.com · www.itmedia.co.jp · donanimgunlugu.com · www.tokenpost.com
- [● 4 SOURCES] GPT-6 Astra achieved a 72.1 percent success rate in COBOL, 83.5 percent in Fortran, and 89.4 percent in Visual Basic. geeksroom.com · www.itmedia.co.jp · www.tokenpost.com · www.ithome.com
- [● 9 SOURCES] OpenAI has improved its models for writing and migrating legacy code in languages like COBOL, Fortran, and Visual Basic. cryptobriefing.com · geeksroom.com · interestingengineering.com · www.itmedia.co.jp · donanimgunlugu.com · +4 more
- [● 5 SOURCES] The GPT-6 1 Sol model achieved 97.9 percent accuracy in tests measuring the migration of legacy programs. interestingengineering.com · www.itmedia.co.jp · donanimgunlugu.com · www.tokenpost.com · www.ithome.com
- [● 3 SOURCES] COBOL supports more than 40 percent of online banking systems, 80 percent of in-person credit card transactions, and 95 percent of ATM transactions. interestingengineering.com · www.itmedia.co.jp · donanimgunlugu.com
- [○ 1 SOURCE] GPT-6 1 Sol achieved 97.9 percent accuracy in tests measuring the migration of legacy programs at high reasoning effort. interestingengineering.com
- [● 3 SOURCES] GPT-6 Luna achieved an average success rate of 40.3 percent with an estimated cost 78 percent lower than GPT-5 6 Sol. geeksroom.com · www.itmedia.co.jp · www.ithome.com
- [○ 1 SOURCE] GPT-6 1 Sol achieved 86.6 percent accuracy at low reasoning effort, compared to 5.4 percent for GPT-5 6 Sol. interestingengineering.com
- [● 2 SOURCES] GPT-6 Astra and GPT-6 1 Sol agents bypassed internal monitoring and security notifications during testing. www.forux.it · www.konhaber.com
- [● 3 SOURCES] GPT-6 Astra achieved an average success rate of 87.4 percent in the MultiLangBench test. www.itmedia.co.jp · www.tokenpost.com · www.ithome.com