2026-11-16 · ← News
Revolut reports 30 to 60% shorter coding cycles with Codex, but the benchmark stays in house
Revolut says in an OpenAI customer story that it cut the time from starting a task to merging a feature by 30 to 60%. The more useful signal is the division of labor: agents implement, while people design the system and remain accountable.
The image could not be loaded.
Revolut says in an OpenAI customer story that Codex reduced the time from starting an engineering task to merging a feature by 30 to 60%. The company also uses OpenAI models to investigate suspicious transactions, but the published figures are self reported and do not constitute an independent benchmark.
Codex took implementation work while engineers kept accountability
A senior platform engineer described splitting an internal code review service into features and assigning their implementation to several agents. He focused on system design, planning and reviewing the result. Revolut says the improvement covers the full path from task start to merge, not merely time spent typing code.
Codex also reportedly saved 1 to 2 days on one substantial security code review. The company says it converted roughly 7 years of design system documentation into AI readable context in 1 day. These are individual company examples, not a published controlled study.
The bigger shift happens before coding and after generation
The practical value extends beyond faster code generation. Revolut is moving experienced engineers toward system design, task decomposition and review. That can increase team capacity if the people approving changes genuinely understand them.
The company describes a similar boundary for its FinCrime agent. It examines information, applies company rules and completes a defined part of an investigation. A person reviews original records, judges the explanation and approves action. For a regulated financial company, that boundary matters more than the number of generated lines.
The headline saving comes without a published method or error rate
OpenAI published the piece as a customer story. It does not provide sample size, baseline period, results by task type or the contribution of other tools and process changes. The 30 to 60% range therefore cannot serve as an expected result for another engineering team.
The FinCrime example also omits accuracy, false positive and missed case rates. Processing more transactions without growing investigator headcount at the same pace describes capacity. It does not establish better decisions.
Audits, reverted changes and investigation quality will settle the claim
The next useful evidence is operational: how many Codex changes are sent back, how much time review consumes and whether production defect rates move. Those numbers will show whether work disappeared or merely shifted into verification.
For financial crime work, accuracy, escalation rates and a traceable record for each decision will matter. Revolut presents a sensible division of labor, but it has disclosed the stopwatch more clearly than the cost of mistakes.
Lilith's verdict
Revolut has shown a stopwatch running 30 to 60% faster. Until it places defect and rework logs beside it, we cannot tell whether the race got shorter or the finish line simply moved to the reviewer.