OpenAI Enterprise Signals: The Agentic Gap Is Now Measurable, and It Is Not About Models

OpenAI published two complementary research pieces on August 12: Enterprise Signals, a recurring dashboard of agentic AI adoption across its enterprise customer base, and a working paper on how organizational adoption grows across roles and seniority. Both are built on aggregated, de-identified usage data, a sample of more than 10 million messages, and an explicit statement that no employee reviewed customer messages.

The headline number is the widening gap between the firms OpenAI calls "frontier" (top 10% of usage each month) and "typical" (the middle decile). As of June, frontier firms generated 8.3x as many output tokens per active user as typical firms, up from 2.6x in January. That is a threefold widening in five months, across industries and company sizes, and it is the first time we have this kind of measurement at enterprise scale from a model vendor.

What the data says#

Five findings carry the report, and each is worth reading carefully:

• Agentic use now dominates enterprise output. As of June, Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers. The number is inflated by the nature of agentic work (longer, multi-step tasks produce more tokens), but the direction is unambiguous: delegated work has overtaken conversational use.
• The frontier gap is widening fast. 8.3x in June versus 2.6x in January, measured as output tokens per active user. OpenAI is explicit that the metric is a proxy for depth of use, not quality.
• Frontier firms adopt advanced capabilities at double the rate. Each week, 21% of active users at frontier firms use Plugins versus 9% at typical firms, and 19% use skills versus 3%. For calibration, 95% of OpenAI's own employees use Plugins weekly.
• Agents are spreading beyond engineering. Since February, weekly active enterprise Codex users grew 108x in legal, 41x in sales, 41x in recruiting, and 26x in marketing, versus 5x in engineering. Coding and system/agent operations account for nearly 75% of agentic messages, but recruiting (32%), sales (26%), policy (25%), and communications (24%) now spend between a quarter and a third of agentic messages on system operations. Coding is almost 60% of agentic messages in design.
• Early-career employees use AI most. Six months after adoption, early-career employees sent 13 more messages per week than executives, the opposite of what survey data usually reports.

What this means for developers#

The most important read is not the gap itself but what it measures. Output tokens per active user is a proxy, and it captures the spread of delegation, not the quality of outcomes. A team that runs agents on long tasks accumulates tokens without necessarily shipping more. Our cost-control analysis and the parallel-agent economics breakdown keep making the same point: token volume is an input metric, and treating it as a success metric is how budget blowouts happen. The report even concedes that frontier adoption at 21% Plugin usage is "a fraction of what is possible" when 95% is internally demonstrated.

Second, the report is a useful structural signal for the skills and plugins economy. OpenAI's own Plugin and skills numbers are the strongest evidence yet that packaging is where enterprise differentiation lands: the same models, the same base product, and a threefold difference in usage depth that correlates with Plugin and skill adoption. The Agent Plugins 1.0 standard, which OpenAI co-maintains, now has a data point attached to it: firms that wire packaged capabilities into agents use the platform roughly twice as much as firms that do not.

Third, the vertical growth numbers are the report's most underrated figure. Engineering grew 5x while legal grew 108x, and 26x in marketing. That matches the pattern we covered in how agentic AI spreads beyond developers: the build-vs-delegate boundary is moving, and the fastest growth is in functions that were not the early adopters. For developers, that means the users you will support next are not engineers. Tooling that assumes terminal-adjacent users will keep losing out to surfaces with permissions, review, and structured workflows, the exact things ChatGPT Work and the merged desktop app are designed around.

Where the methodology deserves scrutiny#

Three caveats before treating this as gospel. The data is OpenAI's own product telemetry, and the definitions are OpenAI's: "frontier" is a relative monthly ranking, so the 8.3x gap is partly a statement about the tail of the distribution by construction. The token proxy can be gamed by long-running agents, and the report does not separate agentic tokens from conversational tokens in the gap calculation. And the working paper's finding that enterprise adopters have stronger financial measures (more assets, more workers, higher R&D) is correlation, not causation; OpenAI says so itself.

Even with those caveats, this is the most concrete enterprise adoption dataset a model vendor has published. The practical takeaway for teams: usage depth is a function of context, tools, and packaging, not model choice. The firms pulling ahead are the ones giving agents Plugins, skills, and governed access, which is the same conclusion our enterprise budget analysis reached from the cost side.

FAQ#

What is Enterprise Signals?#

A recurring set of measures OpenAI publishes on how enterprises adopt AI, based on aggregated, de-identified usage data from its enterprise customer base. The August 12 edition is the first to include the frontier gap analysis.

How does OpenAI define a frontier firm?#

Frontier firms are enterprise customers in the top 10% of monthly AI usage, measured by output tokens per active user. Typical firms fall between the 45th and 55th percentiles.

Is the frontier gap measured in quality or usage?#

Usage only. The metric is output tokens per active user, which OpenAI describes as a proxy for depth of use. It does not measure outcome quality or code quality.

Continue Reading#

• ChatGPT Work and Codex Now Share One Desktop App: What Actually Changed - the product consolidation behind the agentic shift
• Agentic AI Is Spreading Beyond Developers: Platform Engineering's New Job - the non-engineering users in the report's growth numbers
• Agent Plugins 1.0.0: One Package Format for Agent Skills and MCP Servers - the packaging standard the frontier firms are adopting
• Enterprise AI Coding Budget Blowouts: Where the Money Actually Goes - the cost side of token-heavy agentic use
• What Parallel Claude Agents Actually Cost - token accounting discipline for agent fleets

Sources#

• From assistance to execution: How enterprises put AI to work - OpenAI, August 12, 2026. Accessed August 12, 2026.
• Enterprise Signals: What frontier firms are doing differently - OpenAI, updated August 12, 2026. Accessed August 12, 2026.
• How Organizations Use AI: Evidence from ChatGPT (working paper) - OpenAI, August 2026. Cited in the announcement.