OpenAI data shows a widening gap: top 10% of firms now produce 8.3x more AI output

OpenAI published two reports today that put hard numbers on something anyone working in enterprise AI has sensed for months: the gap between companies that dabble with AI and companies that actually use it is becoming a chasm. Frontier firms—the top 10% of enterprise customers by monthly output tokens per active user—now generate 8.3 times as many tokens as typical firms. That ratio was just 2.6x in January. The gap isn’t confined to tech companies either. OpenAI says it appears across industries and company sizes, which should put to rest the lazy assumption that only Silicon Valley startups are doing serious AI work.

The shift is fundamentally about moving from assistance to execution. As of June, Codex generated 64% of combined Codex and ChatGPT output tokens among OpenAI’s enterprise customers. That’s a striking number because Codex is the agentic product—the one that actually does multi-step work rather than just answering questions. Agents produce more tokens by nature since they’re carrying out longer tasks, but the figure also reflects how quickly organizations are handing off substantive work. Weekly active Codex users grew 108x in legal since February, 41x in sales, 41x in recruiting, and 26x in marketing. Engineering grew just 5x, which makes sense—engineers adopted these tools first and had less room to accelerate.

Frontier firms also use advanced capabilities far more aggressively. Each week, 21% of active users at frontier firms use Plugins, compared with 9% at typical firms. For context, 95% of OpenAI’s own employees use Plugins weekly. Plugins are the connective tissue that lets agents access company context—a sales team’s playbook, a CRM, past proposals—and turn that into repeatable workflows. The message is blunt: access alone won’t scale AI. The companies pulling ahead are investing in data infrastructure, governance, and turning individual employee workflows into shared organizational assets.

The report also surfaces an uncomfortable finding for senior leadership. Usage is highest among early-career workers and falls steadily among more senior employees. That suggests a potential comparative advantage for younger workers—one that could invert traditional mentorship dynamics. The working paper notes that enterprise adopters among U.S. public companies held more assets, employed more workers, and invested more in R&D. That’s correlation, not causation, and OpenAI is careful not to overclaim. But read alongside the frontier gap, the picture is consistent: AI advantage compounds where organizations make complementary investments, not where they merely hand out licenses and hope for the best.