On September 16, OpenAI published a guide to the analytics built into the ChatGPT Admin Console[1]. Beyond usage and spend, the tools classify what work employees actually do with AI and track how much Codex contributes to shipped code. The post also walks through how to assemble the evidence for an investment decision, including an illustrative ROI calculation.

Adoption and Spend in a Single View

The Usage view combines active users, credits, and token usage across ChatGPT Work and Codex[1]. Filtering by group or user reveals where adoption has stalled. OpenAI suggests that when a group shows low usage, admins should treat it as a prompt to review starting workflows and training needs with that team's owner[1].

Seeing where spend concentrates is not unusual on its own, but being able to check with numbers whether a capacity request is justified has practical value.

Insights Classifies What the Work Actually Is

The task classifier in Insights groups a sample of messages into use cases and tasks[1]. Software engineering breaks down into feature development and code maintenance; sales and revenue into account research and planning. The Overview tab shows the mix of work at a glance, and the Use cases tab provides a detailed breakdown table.

Task details show the share of credits by Models, Reasoning, and Speed settings[1]. For a routine brief, that supports a decision to test a faster or lower-cost setup and compare quality against the time spent reviewing and correcting the output. Because this maps directly to cost, it is an unglamorous but effective lever.

The Plugin leaderboard and Skills view show which tools support a given task[1]. Low use of a relevant plugin points to an access or training gap, while a heavily used skill may need a named owner and regular updates.

Measuring Codex by Merged Code

The Outcomes view shows Codex contributions to merged commits and lines of code alongside code-review activity[1]. Group, user, and repository filters are available.

OpenAI argues that if Codex contributes to a growing share of merged code, that trend should be compared with review time, defects, and rework to confirm the team is actually shipping more effectively[1]. The point that more generated code means little if reviews back up is obvious enough, but having both in the same view changes how teams operate.

Reporting and the ROI Calculation

The Admin plugin in ChatGPT Work lets admins compare adoption, spend, and tasks, then turn the findings into reports for budget and rollout decisions[1]. It can also produce a leadership deck with charts, key findings, and recommended next steps. The Admin plugin itself was announced separately by OpenAI[2]. With the Admin API, teams can pull the data into their own dashboards and place it next to support metrics[1].

The ROI figures in the post are explicitly hypothetical[1]. Assume 20 sellers each prepare 2 account briefs per week and save 3 hours per brief: across 46 weeks that is 5,520 hours. If half of that time goes into productive work at a fully loaded cost of 75 USD (about 12,000 yen) per hour, the result is roughly 207,000 USD (about 32 million yen) in annual capacity value. Against a first-year cost of 60,000 USD (about 9.3 million yen), the calculation yields a 245 percent ROI[1].

※1 USD = 155 JPY (as of September 16, 2026)

What makes this useful is less the numbers than the fact that every assumption is disclosed alongside them. OpenAI states plainly that all figures are hypothetical and exclude gains from higher win rates or larger deals[1].

Customer Examples and Where to Start

The post also cites customer results. 1Password uses Codex to build features and internal tools, estimating a 553 percent ROI and about 800,000 USD (roughly 124 million yen) in annual engineering capacity value[1][3]. ATV Big Air Tour cut event listing reviews from 8 hours to 1 hour a week and inventory work from 2 to 3 days down to 2 to 3 hours[1]. Playco used GPT-6 Astra through the API to build playable game prototypes and reported 50 percent fewer manual fixes than with its previous model[1].

As a starting point, OpenAI recommends choosing one task in Insights that supports a business priority, agreeing on a baseline and the outcome to measure with a business owner, and setting a review date up front[1]. What would you like to improve, how does the process look today, what changes with AI, what does that make possible, and is the benefit worth the investment: working through those 5 questions in order assembles the case[1].

Summary

What distinguishes Admin Console analytics is that usage, spend, the substance of the work, and Codex outcomes all sit in one place. The design deliberately stops short of letting admins answer the value question alone; the post repeatedly stresses that business owners supply the context needed to measure it. For teams stuck explaining what their AI spend buys, it offers a usable template for measurement.

Source[1]: https://openai.com/index/how-to-connect-ai-usage-to-business-value

Source[2]: https://openai.com/index/introducing-admin-plugin/

Source[3]: https://openai.com/index/1password/