Epoch AI reports signs of AI uplift in Codex contributions

Epoch AI reports signs of AI uplift in Codex contributions

Epoch AI’s analysis was published on July 7, 2026. It examines contributions to OpenAI’s Codex repository for evidence of software engineering uplift: whether engineers write more code per day over time. Epoch AI reports that a small but growing fraction of contributor-days produced outputs estimated to take an experienced developer more than 24 hours to accomplish without AI assistance. For each merged pull request, Epoch AI prompted an ensemble of three frontier AI models to estimate how long an experienced engineer working alone and without AI tools would need to reproduce the merged change. The analysis uses those estimates and the timing of contributors’ pull requests to assess possible uplift. It does not directly measure how much work contributors completed or how much AI assistance they used. The accuracy of the reproduction-time estimates, and whether the findings apply beyond this repository, remain unknown.

Publication: 17ecbd18-4e20-0fc6-2c49-49b89e734a28

attribution — pass: The page is a single Epoch AI data insight with its own analysis, not a listing. The reading attributes the analysis, finding and methodology to Epoch AI and keeps them as Epoch AI's reported findings. The methodology's credit to METR's Amy Deng is omitted, but this does not misattribute anything.

We analyzed contributions to OpenAI’s Codex repository for evidence of software engineering uplift.

summary — uncertain: The summary is faithful, and 'frontier model' is defined on the page. Several central terms are left unexplained for a reader without software background: 'pull request' (a proposed code change submitted for merging), 'contributor-days' (one contributor's output on one day), 'merged', and what OpenAI's Codex is. The 24-hour finding is the core result and depends on these units. Smallest repair: add brief plain-language glosses or term-library tooltips. The question is substantive and fair.

a small but growing fraction of contributor-days produced outputs which would take an experienced developer more than 24 hours to accomplish without AI assistance.

date — pass: evidence.originalPublished is 2026-07-07 (instant). The reading states July 7, 2026, which matches the UTC day and the CSV's 'Updated Jul. 7, 2026'. Nothing after that date is presented and there is no hindsight framing. The kind is publisher-page, so no archive disclosure is required.

CSV, Updated Jul. 7, 2026

support — uncertain: The three source-stated claims match the text. The summary also asserts as fact that the analysis 'does not directly measure how much work contributors completed or how much AI assistance they used.' The retrieved text does not state this. The page's Data, Analysis and 'Assumptions and limitations' sections are present only as headings, so the source's own scope statements cannot be checked. This is plausible from the described method but not shown by the evidence. Smallest repair: label it as Escape Meridian interpretation or say the evidence does not show how AI use was measured.

For each merged pull request, we prompt an ensemble of three frontier AI models to estimate the time one experienced engineer working alone and without AI tools would need to reproduce the net change that was merged.

media — pass: No media candidates were supplied, and the reading selects none (mediaId null). Nothing was invented.