OpenAI Details Admin Analytics for ChatGPT Work and Codex
A new guide shows administrators how product activity, task classifications and engineering contribution views can support—but not prove—business value.
Edited by Tyronne Panaino
OpenAI published a September 16 guide explaining how administrators can use Analytics across ChatGPT Work and Codex. The available views bring together product usage, spending signals, sampled task classifications and engineering contribution data for organisations deciding where to support, expand or rework AI use.
The central lesson is a boundary rather than a promise: product activity can show where and how AI is being used, but it does not by itself establish improved quality, lower cost or business return. OpenAI's own guide says administrators need workflow context and outcome measures from business owners before treating activity as value.
Usage shows adoption and cost concentration
The Usage view combines active users, credits and token activity across ChatGPT Work and Codex. Administrators can filter by group or user to see where adoption is growing, where spending is concentrated and where low use may justify a closer look at access, training or the starting workflow.
That view answers operational questions about participation and consumption. It does not answer whether the work is better. High token use can reflect a valuable workflow, inefficient prompting, difficult tasks or simple experimentation. Low use can signal weak fit, missing access, poor training or a team that does not need the product. The metric becomes useful only when someone checks the surrounding process.
Insights classifies a sample of work
OpenAI says the Insights task classifier groups a sample of messages into use cases and tasks. Administrators can review the mix of work across categories and inspect task details that break down model, reasoning and speed choices. Plugin and Skills views show which tools are associated with a task.
These signals can help a team decide where to test a less expensive setup, offer model-selection guidance, improve access to a relevant plugin or assign ownership for a frequently used skill. They remain classifications of product activity, not direct measurements of accuracy, customer satisfaction or financial return. Sampling also means readers should not assume that every message or workflow appears in the same way.
Codex contribution views need outcome context
The Codex Outcomes view covers contributions to merged commits and lines of code alongside code-review activity. Filters for group, user or repository can help engineering leaders see where adoption is changing and where additional support might be warranted.
Contribution is not the same as causation. A growing share of merged work associated with Codex does not prove faster delivery or better software. OpenAI's guide recommends comparing those trends with review time, defects and rework. That is the more defensible measurement design: pair a product signal with an outcome the team already cares about and include the cost of checking and correcting AI-assisted work.
The plugin and API extend analysis outside the console
The Admin plugin in ChatGPT Work can compare adoption, spending and task trends, then help turn the findings into reports. The Admin API lets organisations automate reporting in their own dashboards and combine OpenAI analytics with data from business systems.
That connection is where outcome analysis becomes possible, but it also creates governance work. Teams need consistent definitions, access controls, a baseline period and an agreed review date. They should decide in advance what improvement would count, measure quality alongside speed and include setup, training and ongoing support costs.
This article deliberately excludes the guide's named-person testimonial, customer result claims and illustrative financial arithmetic. Those details are unnecessary for understanding the measurement surfaces, and publishing them would require a different review path.
Status
Learning. Internal confidence is medium because one official OpenAI guide documents the product views and recommended measurement process, while the practical value of those tools depends on each organisation's data, governance and independent outcome checks.
Sources
Update note: Last reviewed September 17, 2026. We will revise this explainer if OpenAI changes the Analytics views, task-classification method, Codex outcome fields or admin interfaces.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.