Independent newsroom

Tuesday, October 20, 2026

The File

Every claim traces to a primary source. Uncertainty is labeled. Corrections are public.

AI · Tuesday, October 20, 2026

OpenAI says its newest models handle old code better

The File desk · Oct. 20, 2026, 6:44 a.m. ET

Partially confirmed2 primary sources

Status

Partially confirmed. OpenAI posted the results on Monday, and this desk re-opened the page. The scores are the company's claims from tests it built. No outside test was found, and no new model was released.
  • Confirmed
    OpenAI posted on Oct. 19 that its latest models are better at older languages such as COBOL, Fortran, and Visual Basic.
  • Partially confirmed
    The accuracy figures are OpenAI's, from two evaluations OpenAI built. This desk did not run them.
  • Confirmed
    The page does not announce a new model. It reports results for models already released.
OpenAI said Monday that its latest models are much better at older programming languages such as COBOL, the kind that still runs a lot of banking and government software. The scores are the company's, from tests the company built. No new model was released.1
On a test of moving COBOL programs into newer languages without changing what they do, OpenAI said GPT-6.1 Sol scored 86.6 percent at a low setting.
GPT-5.6 Sol scored 5.4 percent at that same setting.
At a higher setting, OpenAI said the newer score reached 97.9 percent. The company said that result came at 61 percent of the older model's cost, which an earlier line on the same page also calls 39 percent less. Applying this to a real system, OpenAI wrote, “still requires expert review and testing against the system's business rules and dependencies.”1
Why it matters (Estimate). Banks, governments, and insurers still run on old code that fewer people know how to fix. Benefit payments, account records, and a lot of ordinary back-office work sit in languages like COBOL. If a model can read that code and move it without changing what it does, those systems are easier to update. OpenAI's own numbers suggest a large jump, they are its own tests, and the company says people still have to check the work.1
A second test, which OpenAI calls MultiLangBench, asks a model to build a working database engine in 20 languages. OpenAI said GPT-6 Astra passed 87.4 percent of the tests on average, up from 26.8 percent for GPT-5.6 Sol, and that COBOL stayed the hardest at 72 percent. Those figures are company claims. The page doesn't name a bank or an agency using the tool.
How we know

Notes

  1. 1.OpenAI, “Advancing coding in low-resource programming languages,” Oct. 19, 2026, re-opened for this edition. https://openai.com/index/low-resource-programming-languages/ “GPT-6.1 Sol achieves 86.6% accuracy at low reasoning effort, compared with GPT-5.6 Sol's 5.4% at the same setting at just 61% of the API cost. At higher reasoning effort, accuracy reaches 97.9%.” An earlier paragraph on the same page says the comparison is “costing 39% less,” which is the same gap. “GPT-6 Astra achieves an average test pass rate of 87.4% across 20 programming languages, up from 26.8% for GPT-5.6 Sol,” with “COBOL remaining the hardest at 72%.” “still requires expert review and testing against the system's business rules and dependencies.” The page also prints COBOL at 72.1 percent in a language-by-language list. That extra tenth isn't used above. OpenAI's citations for shares of banking, card, and ATM systems weren't opened. Company post on X at 6:30 p.m. ET: https://x.com/OpenAI/status/2112310345863930168 No new model name appears as a launch. The models named are GPT-6 Astra, GPT-6.1 Sol, and GPT-5.6 Sol.

What is still unknown or disputed

  • No outside laboratory has published a check of these scores, as far as this desk found.
  • OpenAI names no bank, insurer, or agency using this on a live system.
  • The GitHub copy of the test wasn't opened.
  • Industry estimates of how much banking still runs on COBOL, which OpenAI cites, weren't opened and are not used as facts.

Primary sources

Every claim in this story is drawn from the documents below. If a fetch failed, that is recorded on the card.

  1. Source 1

    Advancing coding in low-resource programming languages

    OpenAI · October 19, 2026

    GPT-6.1 Sol achieves 86.6% accuracy at low reasoning effort, compared with GPT-5.6 Sol's 5.4% at the same setting at just 61% of the API cost. At higher reasoning effort, accuracy reaches 97.9%.

    https://openai.com/index/low-resource-programming-languages/

  2. Source 2

    OpenAI post on older programming languages

    OpenAI · October 19, 2026

    These languages power software that banks, governments, and businesses depend on.

    https://x.com/OpenAI/status/2112310345863930168

Ask this story

Answers come only from this story and its sources. Free test: limited questions per day.

0/300