OpenAI Codex Reaches 4 Million Users as Financial Firms Automate Code
OpenAI
Autonomous AI coding agents double developer output in financial services, shifting the role of engineers from manual coding to steering AI agents and managing higher code volumes safely.
## Background
Published by OpenAI on YouTube on June 8, 2026, this video features Conor Spicer, a Solutions Engineer at OpenAI. The video's description notes that it explores how financial institutions can leverage AI as a competitive differentiator in customer-facing products and digital experiences.
## TL;DR
Conor Spicer introduces Codex, OpenAI's AI coding agent powered by GPT-5, highlighting its widespread internal adoption and rapid growth to over 4 million weekly active users. Through a practical financial services scenario involving the fictional 'Blossom Bank,' he demonstrates how Codex automates context gathering, implementation plan generation, and full-stack coding for new features like predictive budgeting. Spicer further showcases advanced capabilities, including browser automation for regulatory compliance form-filling with human-in-the-loop validation, and embedded GitHub code reviews that catch subtle security vulnerabilities. Ultimately, the talk illustrates how AI shifts engineering roles from manual coding to steering autonomous agents, enabling faster product delivery alongside contextual security.
## Atmosphere & style
The presentation carries a professional, tech-forward, and practical tone driven by a live-demo format. Spicer delivers a structured, articulate walkthrough that balances technical depth with clear business utility, grounding powerful AI capabilities in realistic financial enterprise workflows. Watching the seamless transition from high-level developer metrics to hands-on execution creates an impressive, forward-looking feel that paints a vivid picture of the modern software engineering lifecycle.
## Key takeaways
- [00:05](https://www.youtube.com/watch?v=re-18gil_ec&t=5s) OpenAI's Codex is an autonomous AI coding agent that automates end-to-end software development lifecycles and handles complex, long-running tasks.
- [00:05](https://www.youtube.com/watch?v=re-18gil_ec&t=5s) Since launching its desktop app, Codex adoption rapidly grew to over 1 million downloads in its first week and over 4 million weekly active users.
- [00:05](https://www.youtube.com/watch?v=re-18gil_ec&t=5s) Internal adoption at OpenAI has increased engineer pull requests by 50% without requiring a proportional increase in headcount.
- [00:05](https://www.youtube.com/watch?v=re-18gil_ec&t=5s) Codex shifts the role of software engineers from manual coding to steering, inspecting, and reviewing autonomous AI workflows.
- [00:52](https://www.youtube.com/watch?v=re-18gil_ec&t=52s) Financial institutions can leverage Codex for legacy COBOL migrations, automated regulatory reporting, audit documentation, and rapid product prototyping.
- [03:25](https://www.youtube.com/watch?v=re-18gil_ec&t=205s) The Codex application interface is powered by GPT-5 and integrates with enterprise tools like SharePoint, Jira, and Notion for context gathering.
- [06:03](https://www.youtube.com/watch?v=re-18gil_ec&t=363s) Codex generates actionable implementation plans directly from specifications and executes changes across both front-end and back-end codebases.
- [08:00](https://www.youtube.com/watch?v=re-18gil_ec&t=480s) Browser automation skills enable Codex to populate regulatory compliance portals while maintaining a human-in-the-loop validation process.
- [09:29](https://www.youtube.com/watch?v=re-18gil_ec&t=569s) Automated code reviews by Codex inside GitHub can detect security vulnerabilities that pass human and standard test suite reviews.
- [09:29](https://www.youtube.com/watch?v=re-18gil_ec&t=569s) OpenAI assists organizations in scaffolding team processes to manage increased code volume safely and cleanly.
## Detailed walkthrough
Here is the detailed walkthrough of the transcript:
[00:05](https://www.youtube.com/watch?v=re-18gil_ec&t=5s) Introducing Codex and its Impact on Software Development
Conor Spicer, Solutions Engineer at OpenAI, introduces Codex—OpenAI's AI coding agent. Moving beyond simple autocomplete, Codex automates entire software development lifecycles and complex developer tasks. It can be given a job, analyze an entire codebase, and run autonomously for hours or even a full day until the task is complete.
Since the launch of the Codex desktop app in February, it has provided a natural language interface for technical and semi-technical teams across financial services to generate code and manage workflows. Spicer notes that user adoption has "basically broken all of our charts," surpassing 1 million downloads in its first week and growing to over 4 million weekly active users.
- **Internal OpenAI adoption:** Engineers use Codex by default, shipping in one week what previously took a month.
- **Productivity metrics:** 50% increase in Pull Requests (PRs) completed and submitted per engineer without increasing headcount proportionally.
- **Role impact:** Codex does not replace engineers; it shifts their workflows to make them dramatically more effective.
---
[00:52](https://www.youtube.com/watch?v=re-18gil_ec&t=52s) Financial Services Use Cases and the Blossom Bank Scenario
In financial services, Codex opens up possibilities such as refactoring and migrating legacy COBOL systems, automating regulatory reporting, creating audit-ready documentation, and rapidly prototyping lending, trading, or payments products. Spicer framing this as "the future of development" introduces the scenario of "Blossom Bank"—a successful consumer bank facing competitive pressure to deliver requested customer features rapidly.
Blossom Bank's customers are demanding a predictive budgeting tool that anticipates future spending rather than just displaying historical spending. A feature of this scale previously required lengthy development cycles and complex cross-team coordination. Spicer uses Codex to demonstrate how to streamline this process across both the Codex app and GitHub.
- **Scenario:** Blossom Bank upgrading its historical spend feature to a predictive spending and forecasting tool.
- **Workflow surfaces:** Building code inside the Codex desktop app and reviewing/securing code within GitHub.
---
[03:25](https://www.youtube.com/watch?v=re-18gil_ec&t=205s) Live Demo: Context Gathering and Task Automations
Inside the Codex app, powered by GPT-5, Spicer pulls up Blossom Bank's consumer app running in a development simulation. Rather than jumping straight into coding, developers often face interruptions like urgent meetings. Spicer demonstrates asking Codex to search across codebases, documentation, and observability tools to summarize context and pull incident data live during a meeting.
To scale best practices beyond ad-hoc queries, Codex supports automated workflows. Spicer showcases pre-built template solutions alongside custom automations—such as a "weekly engineering summary" that compiles what was built, what was shipped, and blocking issues ahead of the work week.
- **Underlying model:** GPT-5 powering the Codex application interface.
- **Context connectors:** Integration capabilities with tools like SharePoint, Jira, Notion, and email to pull specs and management-approved definitions without context switching.
---
[06:03](https://www.youtube.com/watch?v=re-18gil_ec&t=363s) Plan Generation and Feature Implementation
After prompting Codex to pull project specifications from SharePoint, the agent searches for the file, inspects the codebase, and drafts an actionable implementation plan directly inside the app. Spicer reviews and approves the plan, instructing Codex to include test runs to ensure alignment with engineering standards before writing code.
Codex then executes the plan across both front-end and back-end services. Spicer highlights how engineering role dynamics shift from manual coding to steering and inspecting AI agents. Once Codex completes the feature, the updated predictive spend forecasting tool appears live in the development simulator, ready for prototyping, feedback collection, and committing to source control.
- **Developer interaction:** Prompting, inspecting progress streams, and steering agent logic mid-execution.
- **Output:** Fully updated front-end/back-end code pushed directly to source control for peer review.
---
[08:00](https://www.youtube.com/watch?v=re-18gil_ec&t=480s) Regulatory Automation and Browser Skills
Addressing real-world financial friction—such as regulatory compliance and legacy web portals lacking APIs—Spicer demonstrates Codex using browser automation skills. To submit required change summaries to a regulatory portal, Codex reads the form fields, searches the codebase for relevant documentation and evidence, and populates the portal autonomously.
Rather than auto-submitting, Codex saves a draft and presents a summary to the developer for human-in-the-loop validation. This converts a process that typically takes hours into a task completed in under two minutes.
- **Compliance workflow:** Automated web form completion with proof/evidence gathering from local documentation.
- **Safety guardrail:** Saves drafts and generates summaries for human review instead of submitting automatically.
---
[09:29](https://www.youtube.com/watch?v=re-18gil_ec&t=569s) AI-Assisted Code Review on GitHub and Final Takeaways
Moving to GitHub, Spicer shows the committed pull request undergoing review. While human reviewers and automated test suites approved the pull request, embedded Codex automated reviews caught a sensitive field mishandling vulnerability that human reviewers missed. Codex can then be prompted directly to draft and implement the security fix.
Spicer concludes by emphasizing that the combination of development speed and contextual security is driving adoption. He acknowledges organizational strains around adopting new tools and notes that OpenAI's engineering teams actively advise clients on scaffolding new processes to handle higher code volumes safely.
- **Key benefits:** Accelerated development workflows combined with production-ready contextual security reviews.
- **Closing focus:** Scaffolding team processes to handle increased code output cleanly and safely.
## Notable quotes
> [00:05](https://www.youtube.com/watch?v=re-18gil_ec&t=5s) basically broken all of our charts — Conor Spicer
> [00:52](https://www.youtube.com/watch?v=re-18gil_ec&t=52s) the future of development — Conor Spicer
OpenAI Solutions Engineer Conor Spicer argues that Codex transforms software engineering from manual coding into autonomous workflow execution, enabling developers to build, test, and ship complex features significantly faster while shifting their role toward steering and supervision.
Adoption Metrics and Shifting Engineering Workflows (0:05)
According to Conor Spicer, Codex expands beyond simple code autocomplete into an autonomous agent capable of analyzing codebases and executing tasks for hours or full days. Following the Codex desktop application launch in February, the tool surpassed 1 million downloads in its first week and grew to more than 4 million weekly active users. Within OpenAI, engineers adopted Codex by default, completing tasks in one week that previously required a month and raising pull requests completed per engineer by 50% without proportional headcount growth.
Spicer claims this paradigm shifts developer responsibilities rather than displacing engineers entirely. Instead of writing all syntax manually, technical teams manage project direction, monitor live execution progress, and inspect completed outputs.
Cross-Platform Context Integration and Implementation (3:25)
Using a consumer banking scenario called Blossom Bank—which required upgrading historical spending views into a predictive budgeting tool—Spicer demonstrated how Codex connects across external tools to assemble context. Powered by GPT-5, the desktop application interfaces with tools such as SharePoint, Jira, Notion, and email to retrieve management specifications and incident data without developer context-switching.
Codex inspects existing codebases alongside project specifications to construct actionable implementation plans, which incorporate automated test runs before execution. Once the developer approves the plan, the agent modifies front-end and back-end services simultaneously, presenting functional feature prototypes for review.
Browser Automation for Regulated Compliance Workflows (8:00)
To navigate operational friction in regulated environments like banking, where legacy portals frequently lack modern APIs, Codex employs browser automation skills. The agent scans local codebase documentation to collect compliance evidence, navigates external regulatory web forms, and populates required change summaries autonomously.
To ensure governance, the workflow incorporates human-in-the-loop guardrails: Codex saves regulatory submissions as drafts with summary reviews rather than auto-submitting them. Spicer notes this reduces compliance filing routines from several hours to under two minutes.
Contextual Security Reviews and Organizational Scaffolding (9:29)
During pull request workflows in GitHub, embedded Codex security reviews identified a sensitive field mishandling vulnerability that human reviewers and existing automated test suites had overlooked. Codex then drafted and deployed the required security remediation directly within the review environment.
Spicer noted that rapid development cycles can create organizational tension when code volume outpaces traditional review capacities. To address this, OpenAI teams work with enterprise organizations to establish new operational scaffolding and governance processes capable of absorbing higher code throughput safely.
While Codex accelerates feature delivery and automates legacy web interactions, Spicer acknowledges that managing significantly higher volumes of generated code creates organizational friction, requiring teams to adopt new review frameworks and safety scaffolding.
[0:05] Conor Spicer: basically broken all of our charts
[0:52] Conor Spicer: the future of development
Original transcript
[0:05] Hi everyone. Good afternoon. I'm Connor
[0:07] Spicer. I'm going to talk now around
[0:09] actually how we build new products with
[0:11] AI and how that can completely change
[0:13] our clients' experiences.
[0:17] So, I want to introduce you to Codex.
[0:19] Codex is our AI coding agent um that can
[0:23] do much more than just auto complete. It
[0:25] can completely automate software
[0:27] development cycles and many different
[0:29] tasks that developers will be doing.
[0:32] We can give it a job, have it look at a
[0:34] code base, and run for many hours or
[0:37] perhaps even a full day until the work
[0:39] is done.
[0:43] So, back in February, we launched the
[0:44] Codex desktop app. It allows anyone with
[0:47] natural language to start generating
[0:49] code and completing many different
[0:51] tasks.
[0:52] We think this Codex app is actually now
[0:55] the ideal surface for both technical and
[0:57] semi-technical teams across financial
[0:59] services.
[1:02] And adoption of Codex has basically
[1:04] broken all of our charts. When the app
[1:06] was launched back in February this year,
[1:08] it passed a million downloads in the
[1:09] first week. Um and from there, it's just
[1:12] kept on growing. We've now got over 4
[1:14] million weekly active users uh across
[1:17] our Codex tools.
[1:20] And inside OpenAI, our engineers are
[1:22] using Codex by default. We now ship more
[1:24] in a week than we would do in a month
[1:26] previously. We're seeing 50% more PRs,
[1:29] that's chunks of code completed and
[1:31] submitted per engineer. So, we're
[1:34] massively able to increase the output of
[1:36] code that we produce, the products that
[1:37] we can ship without increasing head
[1:39] count the same amount.
[1:42] It hasn't replaced our engineers at all.
[1:44] In fact, what we're seeing is it's
[1:45] changing their workflows and making them
[1:47] dramatically more effective in what they
[1:49] do.
[1:51] So, now imagine this in your world. We
[1:53] can start refactoring and migrating
[1:55] legacy COBOL systems. we can automate
[1:58] regulatory reporting, creating
[2:00] audit-ready documentation, or we can
[2:03] rapidly prototype across lending,
[2:05] trading, or payments products and
[2:06] services which we offer. So, this is the
[2:09] future of development. Um shortly, I'll
[2:11] go and show a demo of what this looks
[2:14] like and how it works. Um
[2:16] but to walk through this, we're going to
[2:18] use the scenario of Blossom Bank.
[2:20] Blossom is a successful bank with a
[2:22] really strong consumer banking product,
[2:24] but are facing threats from competitors
[2:27] and growing pressure to move quickly to
[2:29] implement new features that their
[2:31] customers are demanding.
[2:33] One request that we keep seeing coming
[2:36] up from the customer base is around
[2:38] predictive budgeting. They don't just
[2:39] want to be told what they've spent
[2:41] anymore, but they want tools to help
[2:42] them anticipate their future spending
[2:45] and help them plan ahead with that.
[2:47] And actually, they're hearing this a
[2:49] lot. So, they are under pressure to move
[2:52] quickly with this, but this is the type
[2:53] of task that would have taken a long
[2:55] time before and a lot of coordination
[2:58] across different teams to work on this.
[3:01] So, that's why they're turning to Codex.
[3:03] In this demo, I'll now show you how we
[3:05] can transform these workflows both
[3:07] inside the Codex app to build our code
[3:10] and get our job done much more
[3:11] efficiently, but also within GitHub and
[3:14] how we can use this to review the code
[3:16] that we produce and make sure that it's
[3:17] safe and secure to ship.
[3:19] So, with that, we'll jump into our demo.
[3:25] So, here we are inside the Codex
[3:28] application, and this is all powered by
[3:30] that very powerful GPT-5 model which we
[3:33] have. Firstly, let me pull up
[3:36] our consumer app.
[3:42] So, this is what we want to work on. We
[3:43] have a running simulation here in our
[3:45] dev development environment, and keep an
[3:48] eye on this weekly spend.
[3:50] This is the feature which we want to
[3:51] change from being historical looking
[3:54] into that predictive forecasting tool
[3:56] that so many of our customers are asking
[3:58] for.
[4:00] As much as I might want to get in and
[4:01] immediately start writing the code, the
[4:03] reality of my job is I probably have
[4:05] other things to be doing. I may be
[4:06] pulled into an urgent meeting around
[4:08] some instance that we've had that I need
[4:10] to grab other data for. Codex can help
[4:13] for this.
[4:14] I can ask Codex to look through the
[4:16] context it has of our code base and our
[4:18] documentation and other tools to help
[4:21] understand this for me
[4:23] and quickly grab whatever it is as soon
[4:25] as I realize that I need to have it.
[4:28] We'll see it looking across the
[4:29] different observability tools and the
[4:31] code bases to pull the summary for me.
[4:34] This is the sort of thing that would
[4:35] have needed coordination across many
[4:37] other teams previously, but now can be
[4:39] run immediately even live in a meeting
[4:42] if the topic comes up and I wasn't
[4:43] expecting it. Ad hoc tooling is great.
[4:46] We can ask when it we want it, but we
[4:48] also want to scale this across our teams
[4:50] to take these best practices and run
[4:52] them whenever I need and not be reliant
[4:55] on me deciding that it's a good idea to
[4:57] do this.
[4:58] What you can see here is some template
[5:00] solutions that we're seeing a lot of our
[5:02] teams adopting and using that are
[5:04] helpful to get started,
[5:06] but I can also go in and create my own
[5:08] automations
[5:10] around a weekly engineering summary
[5:13] where I need to understand
[5:15] what is it that we've built, what is it
[5:17] that we've shipped,
[5:18] potentially any issues that are blocking
[5:21] our team that I can helpfully have ready
[5:23] for me at the start of the week.
[5:26] So now that we've freed up time,
[5:28] I'm able to get back to what I enjoy
[5:30] doing, which is getting the code built,
[5:32] seeing these new features come to life.
[5:34] We can imagine that actually the context
[5:36] for this sits in many different areas. I
[5:38] may have a management team who've
[5:40] approved a certain definition of the
[5:42] scope that they've worked on with design
[5:44] and product teams and that's all set in
[5:46] SharePoint. Instead of me having to
[5:48] switch out across multiple apps, I can
[5:51] just ask Codex to use that connector and
[5:53] go and find it. Similarly, it could pull
[5:55] in contacts from tools like Jira,
[5:57] Notion, even my emails if people had
[5:59] sent me updated specs in there.
[6:03] Once this starts running,
[6:05] Codex will understand my request to
[6:07] build out the plan of how we're going to
[6:09] implement this. We can see it here
[6:11] searching for SharePoint to find the
[6:13] file that I've referenced.
[6:15] And with that, I can start to see the
[6:17] stream of what it's working on. I am not
[6:20] having to click through manually and
[6:21] search for it, but I can still inspect
[6:23] how Codex is working.
[6:25] Now that it's got that file, I can see
[6:27] it's the one that I was expecting, and
[6:29] it's going to start inspecting my code
[6:30] base and immediately build a plan.
[6:34] I can open this up directly in the app,
[6:36] again avoiding switching across many
[6:38] different tools, understand the proposed
[6:40] approach, validate before it starts
[6:42] changing any code that this is what I
[6:44] expect. This looks good to me, so I'm
[6:46] going to go ahead and say that we can
[6:48] begin to implement this plan,
[6:51] but also including in here some other
[6:52] tasks. Maybe I want to run tests to
[6:55] check that the code it writes is going
[6:57] to match anything, any standards that we
[6:59] have before it starts to run for me.
[7:03] So, I'll submit this, and this is where
[7:06] Codex really gets to work. It starts
[7:08] inspecting my files and immediately
[7:09] starting to implement this feature
[7:11] across all of the front end and back end
[7:13] services which I'm running This is the
[7:15] type of task where engineering has
[7:18] completely changed. Instead of typing
[7:20] the code out myself, I'm prompting this,
[7:22] I'm inspecting how the agent is building
[7:24] it. I can steer it at this stage before
[7:26] it completes to say, "Actually, I've a
[7:28] new idea." Um and once completed like
[7:30] this, can go in and inspect actually
[7:32] what is the code that has been written,
[7:35] having it providing me the context of
[7:37] this, I can check and then build my own
[7:39] understanding of that code, and see a
[7:41] summary of how it's all worked.
[7:43] Switching back to the app, we see the
[7:45] new feature is implemented, it's been
[7:47] prototyped, is ready immediately to go
[7:49] and start collecting feedback on, to
[7:51] share to other people. Finally, I can
[7:53] commit this and push all of this new
[7:55] code, so it's ready for review. Now that
[7:57] might seem really simple
[7:59] and and I'm sure some of you are
[8:00] thinking, well actually, we have
[8:02] regulators to deal with, we have
[8:04] friction with legacy portals, and we
[8:05] have a lot of other challenges. Well,
[8:07] the good news is Codex can help here,
[8:09] too. Imagine I have this type of
[8:11] very legacy-looking portal that may not
[8:13] have APIs to integrate with, but I need
[8:15] to submit all of the summary of my
[8:18] changes to consumer-facing apps to the
[8:20] regulator. This is something that would
[8:22] be a blocker to a fast pace of change
[8:25] and that would take a lot of time to
[8:26] pull the requirements together for it.
[8:29] Using a skill in Codex and our browser
[8:31] automation tooling, Codex can completely
[8:34] handle this task for me, understanding
[8:36] what the form requires,
[8:38] going in, searching across my code to
[8:41] find the relevant pieces of information,
[8:43] the documentation, and the evidence for
[8:45] this
[8:45] before completely populating that portal
[8:48] for me. So, at this point, it's found
[8:50] what it needs.
[8:52] At this stage, I'm hands-off, I'll let
[8:53] the automation take over. It can go into
[8:56] the browser and start filling out all of
[8:58] these different pieces of information
[8:59] for me.
[9:02] Where it will validate it and save the
[9:03] draft, but keeping me in the loop, it
[9:06] won't just hit submit. It will pull
[9:08] together a summary of the changes which
[9:10] it's made, share that with me,
[9:13] and allow me to then check, are there
[9:15] any final points that I want to
[9:17] implement, or do I simply want to read
[9:19] through this, check it myself, and then
[9:21] hit submit? Again, taking tasks that
[9:23] would have taken hours previously down
[9:26] to something which can run in a couple
[9:27] minutes or less.
[9:29] So, we're increasing the amount of code
[9:31] that we're shipping.
[9:32] The other aspect we want to do is, how
[9:34] do we review do
[9:35] and make sure that it's safe as easily
[9:37] as possible.
[9:38] Here we are in GitHub where I have my
[9:40] new change committed. Our automated set
[9:43] of tests have run and our human
[9:44] reviewers have checked this and approved
[9:46] this.
[9:47] But we can also have Codex embedded in
[9:49] this process where it automates the
[9:51] review process. And in this case we see
[9:54] that it's actually found a cyber issue,
[9:56] a potential mishandling of sensitive
[9:58] fields that was missed by our human
[10:00] reviewers, but has been caught here
[10:03] that we can then even send to Codex to
[10:05] start implementing the fix for us.
[10:09] So I want to leave you
[10:11] at this point with a couple key
[10:12] takeaways which we have.
[10:15] We've seen how Codex can really
[10:16] accelerate our development workflows and
[10:18] the other types of tasks that our
[10:20] developers and engineers are doing.
[10:22] And secondly, Codex is able to not just
[10:25] ship more code but to secure it through
[10:27] a production-ready contextual code
[10:29] review.
[10:30] And together this combination of speed
[10:32] and safety is driving so much excitement
[10:35] which we're seeing around Codex today.
[10:38] Let's be honest with ourselves though.
[10:39] We know there's strain on organizations
[10:41] with all this new code, new tooling to
[10:43] learn and that's why myself and our team
[10:45] focus so much on enabling and advising
[10:48] our customers' engineering teams around
[10:50] how we scaffold these new processes
[10:53] so that when we scale up the volume of
[10:54] code, we can really meet the moment and
[10:57] create this overall win for our
[10:59] customers.
