OpenAI confirms transition to agential AI with Codex: usage increases 189 times among non-engineers
A study released by OpenAI in June 2026, titled “The Transition to Agential AI: Evidence from Codex”, showed a 97.9% usage rate for its agential artificial intelligence, Codex. Internally, the tool is already responsible for 99.8% of AI output tokens generated in the company, signaling a clear shift from the conversational model to autonomous systems. The significant growth in adoption by external users who are not developers highlights the migration towards automating complex tasks.
Data presented by OpenAI indicates that 97.9% of internal employees already employ the Codex autonomous artificial intelligence platform. This statistic shows a considerable jump compared to the approximately 40% measured in August 2025. Today, 99.8% of internally produced AI tokens are attributed to Codex, demonstrating the completion of the company’s shift from conversational interactions, such as ChatGPT, to agent-based systems that operate autonomously on complex demands. The rise of Codex is not restricted to software engineering professionals; Since August 2025, use by external non-developers has grown 137 times, and by internal employees without a development profile, an increase of 189 times.
In the OpenAI environment, token production expanded 13 times in the legal sector and 56 times in the research area. These numbers confirm the wide adoption of the tool across a variety of non-technical departments, including recruitment and finance teams.
Significant changes in system usage patterns were also identified. The vast majority, 80.6%, of activities attributed to Codex refer to operations that would require more than 30 minutes of manual execution by one person. Additionally, the proportion of tasks estimated to last more than 8 hours of human work increased more than tenfold, from 2.1% at the end of 2025 to reach 25.6%. Every week, more than 10% of users who actively access Codex operate three or more agents simultaneously. These results reinforce the new role of AI, which evolves from a mere answerer of specific questions to an autonomous processor of complex tasks that require a large volume of time.
Despite the progress, there is a notable difference in technology adoption rates both nationally and internationally. While internal usage at OpenAI almost reaches 100%, the percentage of usage by external organizations is just 17.3%, and by individual users outside the company, a mere 0.7%. However, even with the 17.3% active usage rate among organizations, Codex generates 63.3% of all outgoing tokens from these entities.
International press outlets, however, warned that the numbers presented derive from OpenAI’s self-declaration and do not have independent third-party verification. The release of this data comes at a crucial period, as OpenAI reorganizes its product strategy, migrating from the conversational model to the agent model, in anticipation of a possible initial public offering (IPO) in the last quarter of 2026. Therefore, some analysts defend the need for an external analysis to assess effective productivity gains.
