OpenAI published a customer story on October 8, 2026 describing how Oracle runs ChatGPT Work and Codex across recruiting, engineering and operations. It puts numbers on the rollout: 130,000 active ChatGPT users, 95,000 or more on Codex, and a 98% drop in the time its talent acquisition team spends on research.
What was actually released
No model and no feature shipped here. What OpenAI released is a document, four pages of Oracle executives describing how their teams built tools on top of ChatGPT Work and Codex, with the figures the company supplied. That distinction matters for anyone reading it as evidence: this is a sales asset that happens to carry useful numbers, not a study.
The page walks through three deployments, and each one is a workflow rather than a chat.
- Talent acquisition. The team built a talent market intelligence tool with ChatGPT Work. It takes a job description, researches comparable roles, benchmarks compensation and assesses the talent pool across locations. Collecting that information used to take two to four days; the page says prep now takes about 15 to 20 minutes.
- Oracle Applications Lab. The team built an ontology of the company’s objects, relationships and rules so a plain-language business question becomes a reliable SQL query through Codex. A user describes the outcome, Codex chooses the internal systems to call, gathers the data and returns an analysis, a report or an application.
- Production engineering. Site reliability engineers use Codex to pull context about an incident and surface the right playbook. A simple incident that used to take about an hour is now handled in minutes, according to the page.
The page also carries three leadership lessons worth reading past the figures: give the model correct guardrails, hand it prototypes instead of written specs, and stay in the code so what it produces remains maintainable. That third one is the only line on the page that prices a cost of the approach.
The four figures on the page
- 130,000 active ChatGPT users at Oracle.
- 95,000 or more active Codex users at Oracle.
- 98% decrease in the time talent acquisition research takes.
- 2 to 4 days of compiling down to 15 to 20 minutes of recruiter prep.
| What Oracle measures | Figure OpenAI publishes | What it actually covers |
|---|---|---|
| ChatGPT users at Oracle | 130,000 | Active users, no measurement window given |
| Codex users at Oracle | 95,000 or more | Active users, no measurement window given |
| Talent acquisition research | 98% less time | One team’s internal tool, reported by Oracle |
| Recruiter prep per role | 15 to 20 minutes | Down from 2 to 4 days of compiling |
| Simple SRE incident | Minutes | Down from about an hour, reported by Oracle |
What the claim rests on
Every number in the story is Oracle’s or OpenAI’s, and none of them is measured on the page. There is no study behind the 98%, no sample of roles, no baseline for the old process and no date range for the new one. The percentage is attached to one team’s internal tool, and the page explains it with a single recruiter’s account of moving from days of compiling to minutes of prep. That is a real workflow change. It is not a measurement, and the difference decides whether you can repeat it.
The user counts are softer still. Active users are not defined: no window, no way to tell a daily login from a licensed seat that opened the tool once. OpenAI’s own overview says only that Oracle uses the tools across recruiting, analytics and engineering, which is a description of intent, not of usage.
What the page does prove is that a company of Oracle’s size has put these tools into production across several functions. Three named executives, three workflows, one of them in recruiting, one in analytics and one on incident response. That is more than a pilot, and it is the part of the story a reader can trust without checking Oracle’s internal measurement.
What the page does not publish
No pricing, no seat count against licenses, no usage limits, no admin controls, no data-handling terms and no independent verification of any figure. The stories in circulation about enterprise AI rarely fold these in, and this one leaves all of them out.
On this page
What it changes for teams buying AI tools
For a buyer, the number that carries weight is 95,000, not 98%. A percentage from one team’s homemade tool tells you little about your own bill. A company running 95,000 Codex seats tells you the tool has cleared enterprise deployment, which is the harder test: identity, permissions, procurement and support at scale.
That shifts the questions worth asking a vendor. If Oracle’s peer group is standardizing this way, your account team has the references to quote, and you should ask for the same shape of evidence: seats in production, a defined active-user window, and the usage limits that come with them. The case study markets outcomes; your invoice is per seat and per token, and neither figure appears on this page.
The second lesson is technical and applies whether or not you use these tools. Oracle Applications Lab built a semantic layer, an ontology of objects, relationships and rules, before pointing Codex at its internal systems. The reason a plain-language question returns a reliable SQL query is that work, not the model on its own. A team that skips the ontology and asks the model to guess will get answers that look right and occasionally are not. Compare that discipline with the way vendors now ship cloud agents with hard spend caps, as Google did with its Gemini agent: the controls around the model are becoming the product.
The Neural Minute take
The honest line on the whole page is the one nobody will quote in a launch post. It comes from Oracle’s Richard Lam, and it is about maintenance, not speed.
If you don’t work alongside Codex, you’re going to get into a situation of having lots of code that is not going to be maintainable.
That is a customer telling you the cost of the thing his company is selling the benefits of. It is worth more than the 98%, because it is a decision rule: if your team generates code it cannot review, the days-to-minutes gain is borrowed against future work.
Our reservation is the same one we bring to every vendor case study. The 98% is the number Oracle and OpenAI will repeat, it is the least checkable figure on the page, and the page gives no basis for it. Ask any supplier quoting a percentage saving for the arithmetic behind it, in writing, exactly as you would for a discount on a license. Here the arithmetic is a single team’s account of its own tool.
Two things we would verify before treating this as a reference: an active-user definition and window, because 95,000 is the claim that matters and it is undefined; and the maintainability line, because a coding assistant that ships faster than your review process can absorb is a debt.
Sources
- How Oracle turns days of work into minutes with ChatGPT and Codex, published by OpenAI on October 8, 2026: https://openai.com/index/oracle
- The figures quoted here are read on that page on October 8, 2026, not from any third party
- The quotes from Ackerman and Lam appear in the OpenAI story; no independent measurement of the figures was published with it