Product1 publisher2 min readPublished
Oracle answers its surprise AI bill by showing employees what each model costs
Oracle put ChatGPT Enterprise and OpenAI's Codex in front of about 160,000 employees in April, and 80 percent were using them within three months. Then the bills arrived and, the CIO said, surprised the company.
The Product Desk · Product desk
What happened
- Oracle launched ChatGPT Enterprise and OpenAI's Codex to employees in April and May, after setting corporate standards, security controls and internal policies, according to global CIO Jae Evans.
- Within three months of the launch, 80% of Oracle staff were using the new tools.
- Oracle now shows employees which models they used and what each one cost, with GPT-6 Astra running about two and a half times the price of cheaper options such as GPT-5.6 Terra.
- Testing Anthropic's Claude Mythos Preview against its own code turned up more potential vulnerabilities in two weeks than Oracle had found in the previous year, Evans said.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- cost Oracle's own budget absorbs the gap between a launch and a meter: every routine task left on the expensive model costs two and a half times the cheaper equivalent, and the correction shows up on the next invoice.
- constraint Release management now sets the pace for customers, so buying more coding capacity does not move a ship date until testing, validation and deployment can take the volume.
- decision Anyone rolling this out picks who gets interrupted: a hard monthly cap on the engineer, as JPMorgan set, or visibility after the spend, as Oracle chose.
- exposure Engineers become the verification layer for a scanner that is wrong most of the time, and that review time sits in the security team's week.
The policy work came first. Oracle set corporate standards, security controls and internal policies before the tools went out, global chief information officer Jae Evans said [4]. A year earlier the company had AI in customer support and nothing for its developers, finance or sales teams [2]. Co-CEO Clay Magouyrk, describing that period at an internal town hall reported by Business Insider, said, "I don't think we figured out how to make AI really that useful for ourselves." [1]
Evans's team supports about 160,000 employees [4]. Eighty percent of that base is roughly 128,000 people picking up a metered tool inside a quarter [1]. The pre-launch controls governed access and handling, not spend, and Evans said the bills came as a surprise because the tools were so easy to start using [6].
A cap stops the work when the money runs out. A dashboard lets the task finish and hands the employee the number afterwards. JPMorgan drew the line in the first place, limiting engineers to $2,000 per month of Claude usage [9].
Evans said developers now produce in a week what used to take a team two to three quarters [10], and customers have not received anything sooner [11]. Magouyrk said writing code faster "doesn't mean that suddenly everything is 1000 times faster" [12]. Oracle is still updating its testing, validation, deployment and release management processes to keep up [13].
The security side ran the same way. Oracle was one of the partners given early access to Claude Mythos Preview, a model Anthropic announced in April and released only to a select group through Project Glasswing [14]. Anthropic says it can identify and exploit zero-day vulnerabilities in all major operating systems and browsers [15]. At a false-positive rate of 60 to 70 percent, between 30 and 40 of every 100 flagged items are real [2], and an engineer has to read all 100 to find them. Oracle built a verification step before anyone was allowed to start fixing [17].
Before the seats go out, settle whether the invoice scales with headcount or with usage: a seat price makes a policy document sufficient, a usage price makes a meter necessary, and Oracle's April rollout had the policy without the meter [3]. Settle too which downstream team absorbs the output, and whether that team's throughput was ever measured in the pilot. Oracle spent tens of billions of dollars over two years building the infrastructure other companies run their AI on [19], is discussing its own rollout openly with staff, and is doing so while it cuts jobs and spends more on data centres than it earns in a quarter [18].
What to watch
- Whether Oracle reports a customer-facing delivery metric showing release management caught up with code output.
- Whether the per-model cost display, now a number employees see after the fact, becomes an enforced routing rule.
- Whether Anthropic widens Project Glasswing access and whether the 60% to 70% false-positive rate falls in later builds.