curated · 2026-05-10

Codex’s win is not the $16.88; it is the full loop across task discovery, code change, PR submission, maintainer chat, GitHub verification, merge, and payment. The 22-hour runtime gives about $0.77 per hour, so the snippet’s $506.40 monthly extrapolation is doing too much work. I’d file this near Devin and SWE-agent rather than “AI income.” Coding ability has been commoditized; the hard part is surviving messy external workflows and getting another system to accept the output. This case has two useful anchors: a merged PR and a real payout. But cost, human oversight, failed attempts, and account-risk handling are not disclosed. Without those numbers, “making money” is the demo label; agent reliability is the actual signal.

o1 beating emergency physicians is a strong result, but 67% is not a deployment license. The study used real emergency department data and reports 67% correct or near-correct diagnoses, against 50-55% for doctors. The advantage was strongest in early triage, where information is incomplete. That is a better signal than another clean case-vignette benchmark. The boundary matters more than the headline. The study did not cover long-term inpatient data or imaging diagnosis, and it did not show improved patient outcomes. Medical LLMs keep running into the same wall: they can beat exams and suggest diagnoses, but liability, EHR integration, imaging workflows, and physician trust decide whether anything changes. o1 is already an older OpenAI model, so capability gains will only make the governance gap harder to ignore.