OpenAI's GPT-6.1 Sol model, on October 19, 2026, demonstrated a 97.9% accuracy rate in migrating COBOL programs to modern programming languages [12, 14, 15]. This significantly surpasses its predecessor, GPT-5.6 Sol, which achieved only 5.4% accuracy in the same evaluation [12, 14, 15].

This progress could help modernize outdated software systems that underpin critical global infrastructure, such as banking operations, which still largely rely on languages like COBOL. The ability to update these systems more efficiently promises to reduce technical debt for businesses and enhance the resilience of essential services [12, 13, 14, 15].

Records Show Improved Code Migration

OpenAI published new coding evaluations on October 19, 2026, specifically for "low-resource" languages, including COBOL [12, 13, 14]. One evaluation measured a model's capacity to translate existing COBOL programs into modern languages like C++, Java, Kotlin, Python, or Rust, while strictly preserving their original behavior [12, 14]. A migration was considered successful only if the rewritten program's output exactly matched the original byte-for-byte, accounting for specifics like COBOL's fixed-point arithmetic [12, 14].

GPT-6.1 Sol achieved an accuracy of 97.9% at a high reasoning effort on this COBOL migration test [12, 14, 15]. At a low reasoning effort, the model still reached 86.6% accuracy, doing so at 61% of the API cost of GPT-5.6 Sol [12, 14].

Expanded Language Capabilities

OpenAI also developed MultiLangBench, a second evaluation that required models to build a SQLite-compatible database engine from scratch across 20 distinct programming languages [12, 14, 15]. On this benchmark, GPT-6 Astra recorded an average test pass rate of 87.4% across the 20 languages with maximum reasoning effort, a significant improvement over GPT-5.6 Sol's 26.8% [12, 13, 14]. GPT-6 Luna also improved on GPT-5.6 Sol, achieving a 40.3% pass rate at an estimated cost per run of $0.77, which is a 78% reduction compared to GPT-5.6 Sol's $3.48 [12, 14, 15]. COBOL proved to be the most difficult language for GPT-6 Astra on MultiLangBench, reaching a 72.1% pass rate [14, 15].

What Changed and For Whom

GPT-6.1 Sol became available to users of ChatGPT Work and Codex across Plus, Pro, Business, Enterprise, and Edu tiers, as well as through the OpenAI API [16, 17]. Its API pricing is $2 per million input tokens and $10 per million output tokens, with cached input costing $0.10 per million tokens [13, 14]. OpenAI stated that GPT-6.1 Sol offers intelligence comparable to GPT-6 Astra at approximately one-fifth of the price [7, 13].

OpenAI acknowledges that migrating production software "still requires expert review and testing against the system's business rules and dependencies" [12, 14]. This work is informed by feedback from developers and enterprise customers who highlighted the challenges coding models face with these low-resource languages [12, 14, 15].

What We Could Not Confirm

Whether external, independent verification of OpenAI's evaluation results has been conducted could not be confirmed.