8.3x the Token Output: OpenAI Maps the Gap Between Companies That Use AI and Ones That Master It

Hey everyone, it's Shiichan! Today I found a research piece packed with numbers about how enterprises actually use AI, and it turns out there's a bigger gap than you'd expect between companies that use it and companies that really master it.

OpenAI News openai.com

What was announced?

OpenAI News published research analyzing how enterprises are actually using AI. The core theme is that enterprise AI use is shifting from assistance (answering questions) to execution (actually doing the work), and the pace of that shift is creating a widening gap between frontier companies and everyone else.

Why it matters

The key insight here is that just looking at whether a company has "adopted AI" can already put you behind. There's a real gap in outcomes between companies that simply hand out a tool and companies that go further, building out permissions and governance so the whole organization benefits. And the headline finding is that this gap isn't growing slowly, it's accelerating fast.

What changes

The research points to establishing connections, permissions, and governance as the key to closing this gap, essentially converting individual workflows into organization-wide practice. In other words, it's not enough for a few power users to benefit; the goal is getting the whole team or organization to the same level of capability.

Dive Deep

The report highlights five main findings:

• Enterprise AI is getting more autonomous: Codex generates 64% of output tokens for enterprise customers itself, pointing to more multi-step work being delegated at once
• The frontier company gap is widening: companies in the top 10% of monthly usage had 2.6x the token output of typical companies in January, and that grew to 8.3x by June
• A gap in advanced feature adoption: 21% of weekly active users at frontier companies use plugins, versus 9% at average companies. Internally at OpenAI, 95% of employees use plugins weekly
• Expansion across knowledge work: since February, weekly active Codex users grew 108x in legal, 41x in sales, 41x in recruiting, and 26x in marketing (versus 5x in engineering)
• Early-career employees are driving usage: six months after being hired, early-career employees send 13 more weekly messages than executives

The "plugins" mentioned here combine skills, which provide reusable instructions, with apps, which provide access to enterprise data, tools, and actions. A sales plugin, for example, might combine a team's playbooks with CRM integration so responses can draw on customer info and past proposals.

Virgin Atlantic is featured as a real-world example:

• Engineering team: cut legacy code refactoring that used to take 2 weeks down to 30 minutes
• Product team: turned weeks of competitive analysis into hours, and used the time saved to build a 5-year digital strategy

The piece also points to Enterprise Signals for industry- and function-specific breakdowns, the underlying research paper "How Organizations Use AI: Evidence from ChatGPT," and custom benchmarks available to OpenAI Enterprise customers.

Wrap-up

• Enterprise AI use is shifting from answering questions to actually executing work
• The token-output gap between frontier companies and average ones widened from 2.6x in January to 8.3x by June
• Plugin adoption also differs sharply: 21% at frontier companies vs. 9% at average companies, and 95% internally at OpenAI
• Codex use is expanding fast beyond engineering, into legal, sales, recruiting, and marketing
• At Virgin Atlantic, legacy code refactoring dropped from 2 weeks to 30 minutes, and competitive analysis from weeks to hours

If you're wondering whether your company's AI rollout stopped at "handed out the tool," this research is worth a read for anyone in DX, IT planning, or corporate strategy!