OpenAI’s heaviest users generate 60+ hours of AI agent work per day. 80.6% of users now assign tasks that would take a human 30+ minutes. 85% of output tokens at OpenAI flow through Codex. This isn’t a productivity tool report — it’s the first data on what work looks like when agents do most of it.
OpenAI published “How Agents Are Transforming Work” — a report using its own internal data showing how work patterns have shifted since agents became the primary interface.
60 Hours of Agent Work Per Day
At the 99th percentile, OpenAI’s heaviest users generate more than 60 hours of Codex agent turns per day — distributed across multiple parallel agents. One person, running dozens of agents simultaneously, producing 60 hours of work output in a single workday. That’s the Harness Theory measured in runtime hours.
From December 2025 to May 2026, the share of users assigning tasks that would take a human 30+ minutes rose to 80.6%. Tasks taking 1+ hour: 70.2%. People aren’t using agents for quick questions anymore — they’re delegating multi-hour projects.
85% of Output Through Codex
For the average OpenAI employee, Codex now accounts for 85% of all output tokens. Not ChatGPT. Not the API playground. Codex — the agentic coding tool. The chat window is dead at OpenAI. The agent is the interface.
The key insight: 60 hours of agent work per day from one person. That’s not productivity improvement — it’s a categorically different way of working. The human doesn’t do 60 hours of work. The human frames what 60 hours of agent work gets pointed at. This is the principal/operator split in operational data: the human is the principal, the agents are the operators.
The Structural Read
THE CHAT ERA IS OVER — INSIDE OPENAI
85% of output tokens through Codex, not ChatGPT. The company that built ChatGPT has moved past it internally. The interface of the future isn’t conversation — it’s delegation. You don’t chat with AI. You assign work to agents.
60 HOURS = THE HARNESS IN RUNTIME
“One person with judgment plus a harness produces what used to take a team.” We wrote that in the Harness Trilogy. OpenAI just quantified it: one person, 60 hours of parallel agent runtime, distributed across multiple agents. The team didn’t shrink — it was replaced by an agent swarm the principal directs.
GPT-5.6 SOL’S ULTRA MODE MAKES THIS THE DEFAULT
GPT-5.6 Sol has an “ultra” mode that spawns subagents. That’s this exact pattern — parallel agent work — built into the model for everyone. Today it’s OpenAI’s top 1%. With Sol’s ultra mode, it becomes accessible to every enterprise customer.
The Bottom Line
OpenAI published what might be the most important productivity data of 2026: 60 hours of agent work per day per person. 85% of output through Codex. 80% of users assigning 30+ minute tasks. This is what Stripe’s solopreneur data looks like from the inside — the mechanism that makes one person produce the output of a department. The work didn’t get faster. The work got delegated. And the delegation infrastructure now runs 60 hours of parallel computation for every hour the human is awake.
Business Engineer
The Harness Trilogy — 60 Hours of Agent Work Per Day
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
