CodingSoftware & developer tools
OpenAI case study credits Codex with cutting Revolut's coding time by up to 60%
OpenAI's Revolut customer story credits Codex with 30–60% shorter coding time and faster security reviews, and describes an AI agent for financial-crime investigations. Weeks earlier, Revolut's infrastructure lead attributed a similar 60% gain to a self-hosted open-weight GLM model.
From the OpenAI News article of November 16, 2026 and cited sources: OpenAI's case study says Codex cut Revolut's coding time by 30–60% and took one to two days off security code reviews. [1] [9] [10]
SOURCES[1] How Revolut raises ambition without lowering the bar… openai.com[9] Revolut Uses OpenAI Tools to Cut Coding Time by Up to 60%… tokenpost.com[10] What Revolut's Codex rollout actually tells us (Ken Ashe) kenashe.ai
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- OpenAI's case study says Codex cut Revolut's coding time by 30–60% and took one to two days off security code reviews. [1] [9] [10]
- Original source
- OpenAI News article, company source
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 16, 2026 article published
- Covered
- Nov 18 edition
- Evidence
- 13 sources across 13 sites
- 1How Revolut raises ambition without lowering the bar…openai.com
- 2Shipping like a startup, regulated like a bank…ai.engineer
- 3Revolut Cuts A.I. Coding Costs 90% With Open-Weight GLM…trendingtopics.eu
- 4Revolut and LSEG join growing list of OpenAI customers…cityam.com
- 5Revolut's AI Outperforms Human Reviewers at Detecting…pymnts.com
- 6Revolut teams up with OpenAI to offer ChatGPT Go…fintech.global
Reach: not measured yet

OpenAI's case study says Codex cut Revolut's coding time by 30–60% and took one to two days off security code reviews. It names no model version and gives no baseline or team count, and OpenAI curated the figures itself. [1] [9] [10]
In October, Revolut's infrastructure lead credited a 60% coding-time gain and a 70% lead-time gain across about 2,000 engineers to self-hosted GLM running on 56 Nvidia B200 GPUs. Coverage says this cut coding-model spend by about 90% after Revolut moved away from Claude Code and Codex. [2] [3] [8] [12]
Under GLM, lead time (task pickup to merge or deploy) improved more than coding time. Revolut's own reading was that much of the delay sits in work outside coding. Earlier, token budgets limited how much hosted tools could help. [2]
OpenAI models power a FinCrime agent that screens, monitors and investigates suspicious transactions, with humans keeping the final decision. A Revolut executive said agentic monitoring beats human reviewers, but no figures on accuracy, false positives or fraud losses have been published. [1] [5] [10] [13]
Two accounts from Revolut tie similar productivity gains to different tools, and the FinCrime agent has no published performance data.
Read the full assessment
Teams benchmarking coding agents should ask for baselines and cohorts, and should track lead time as well as coding time.
Status: research is done, and I'm now writing the dossier.
Executive brief
OpenAI's Revolut case study, published 16 November 2026, says Codex cut coding time by 30–60% (OpenAI). A month earlier, Revolut's infrastructure lead told a conference audience about a similar 60% gain. He attributed that one to a self-hosted open-weight GLM model that replaced costly frontier-lab tools, Codex among them (AI Engineer; Trending Topics). The case study also describes an OpenAI-powered "FinCrime agent" for transaction investigations. No published metric covers that agent's accuracy. Every productivity figure comes from Revolut itself.
What changed and event timeline
Revolut joins OpenAI's financial-services customers
City AM reported that Revolut was using GPT-5 to strengthen financial-crime protections and to upgrade its customer assistant, Rita ().
AI said to beat human reviewers
At a Semafor forum, Revolut executive Cetin Duransoy said agentic transaction monitoring performs "statistically significantly better" than human review. PYMNTS reported no supporting data ().
ChatGPT Go becomes a banking perk
Revolut bundled up to 12 months of free ChatGPT Go into its paid plans, its third AI partnership in months after Perplexity and Lovable (;).
Self-hosted GLM talk
David Abitbol told AI Engineer NYC that self-hosted GLM brought 60% coding-time and 70% lead-time improvements across about 2,000 engineers ().
Coverage puts the cost cut at 90%
Reports say coding-model spend fell to about one-tenth, roughly $0.30 per million tokens, after Revolut moved away from Claude Code and Codex (;).
OpenAI publishes the case study
It credits Codex with 30–60% shorter coding time and 1–2 days off security code reviews, and describes the FinCrime agent (;).
Capabilities and access
- Codex: used for coding, and reportedly for security code reviews and an internal code-review service built by Senior Platform Engineer Dmitry Pitik.
- FinCrime agent: built on unspecified OpenAI models. It screens, monitors and investigates suspicious transactions, and humans keep the final decision (Ken Ashe).
- Earlier reporting names GPT-5 for FinCrime work and for Rita (City AM).
Read the full section
- Codex: used for coding, and reportedly for security code reviews and an internal code-review service built by Senior Platform Engineer Dmitry Pitik. No model version is named (OpenAI).
- FinCrime agent: built on unspecified OpenAI models. It screens, monitors and investigates suspicious transactions, and humans keep the final decision (Ken Ashe).
- Earlier reporting names GPT-5 for FinCrime work and for Rita (City AM).
- Self-hosted stack: GLM 5.1, later 5.3, served on 56 Nvidia B200 GPUs (AI Engineer).
Technical analysis for researchers and developers
- Metrics: Revolut tracks coding time in hours and lead time, from task pickup to merge or deploy, in days.
- Earlier tools: Gemini gave 20% coding-time and 10% lead-time gains. Token budgets limited the impact of Claude Code (same source).
- Serving: Tuning cut time-to-first-token from 14.26 s to 1.68 s, and throughput rose 76% before any hardware was added (same source).
Read the full section
- Metrics: Revolut tracks coding time in hours and lead time, from task pickup to merge or deploy, in days. Under GLM, lead time improved more than coding time. Abitbol read this as a sign that delays come from non-coding work (AI Engineer).
- Earlier tools: Gemini gave 20% coding-time and 10% lead-time gains. Token budgets limited the impact of Claude Code (same source).
- Serving: Tuning cut time-to-first-token from 14.26 s to 1.68 s, and throughput rose 76% before any hardware was added (same source).
- Reproducibility: The 753B-parameter mixture-of-experts model has about 40B active parameters per token (MACCUS). No evaluation set, baseline cohort or code has been released.
Claims and evidence
No independent evaluation corroborates any of these claims.
Read the full section
| Claim | Source | Reading |
| Codex cuts coding time 30–60% | OpenAI | Vendor-curated. No baseline or team count is given (Ken Ashe) |
| Self-hosted GLM gives 60% coding / 70% lead-time gains | AI Engineer | Self-reported. Not isolated from the removal of token limits (MACCUS) |
| Security code reviews 1–2 days shorter | TokenPost | Restates the case study |
| AI beats human AML reviewers | PYMNTS | Executive statement; no data published |
No independent evaluation corroborates any of these claims.
Context and prior work
- Revolut built its own fraud and credit model, PRAGMA, and reports a 64.7% improvement in fraud detection (KuCoin News).
- NatWest, OakNorth and Zopa also use OpenAI, and financial services is one of OpenAI's fastest-growing segments (City AM).
- In anti-money-laundering (AML) work more broadly, HSBC reports 60% fewer false positives using Google Cloud's AML AI (PYMNTS).
Limitations, safety and contested findings
- Attribution conflict: The 60% coding-time figure is credited to Codex in OpenAI's story and to self-hosted GLM in Revolut's own talk.
- Reporting inconsistencies: Startup Fortune names GLM-4.5 and a GTC Berlin talk on 22 October, with time-to-first-token falling from more than 10 s to 1.6 s (Startup Fortune).
- Missing FinCrime data: No fraud-loss, precision or false-positive figures have been disclosed (StartupHub).
Read the full section
- Attribution conflict: The 60% coding-time figure is credited to Codex in OpenAI's story and to self-hosted GLM in Revolut's own talk. The reviewed sources do not reconcile the two.
- Reporting inconsistencies: Startup Fortune names GLM-4.5 and a GTC Berlin talk on 22 October, with time-to-first-token falling from more than 10 s to 1.6 s (Startup Fortune). Other outlets cite GLM 5.3, AI Engineer NYC on 13 October, and 14.26 s to 1.68 s.
- Missing FinCrime data: No fraud-loss, precision or false-positive figures have been disclosed (StartupHub).
- Scope of self-hosting: It covers text and code only. Image tasks still go to hosted models (Startup Fortune).
Business and practitioner implications
- Measure lead time as well as coding time. At Revolut, lead-time gains exceeded coding-time gains, so delays outside coding matter.
- Run several vendors. Revolut mixes OpenAI agents, hosted models and self-hosted open weights. Token cost drove its coding-tool choice more than model quality did.
- Treat vendor customer stories as marketing, not audit evidence. Ask for baselines, cohorts and false-positive rates before using them to benchmark.
Read the full section
- Measure lead time as well as coding time. At Revolut, lead-time gains exceeded coding-time gains, so delays outside coding matter.
- Run several vendors. Revolut mixes OpenAI agents, hosted models and self-hosted open weights. Token cost drove its coding-tool choice more than model quality did.
- Treat vendor customer stories as marketing, not audit evidence. Ask for baselines, cohorts and false-positive rates before using them to benchmark.
- For regulated firms, self-hosting brings data control and sandbox enforcement (AI Engineer). It also requires capital: about $2.7M for the GPUs (Startup Fortune).