Codex from 0 to 10M Users: Building ChatGPT Work — Akshay Nathan, OpenAI

OpenAI's Akshay Nathan, who runs Core Product Engineering for productivity, said ChatGPT Work and Codex now share a single agent harness and hit 10 million combined users two weeks after the July 9 launch. The non-developer half is growing faster. GPT-5.6 was trained specifically to improve artifacts: spreadsheets, documents, and Sites. All of that sounds like momentum. The problems are underneath.

The harness is real; the productivity claim is not

The actual technical bet is the shared runtime, one persistent-compute environment with plugins, computer use, artifacts, and sub-agents, surfaced two ways. Codex shows you Git diffs; Work hides them. That's a legitimate architectural consolidation, and it's the same bet Anthropic made with Claude Code's sub-agents.

The productivity claim is something else. Nathan admitted on the record that thumbs up/down is useless, and that pull-request counts and token counts are now meaningless proxies. OpenAI has no instrument for the thing it's selling. Ten million users two weeks post-launch is distribution, not adoption. Those are existing paying ChatGPT subscribers who got a new tab. Day-30 retention is the number that would tell you something, and Nathan doesn't have it.

Where the real cost is

Ultra mode, which fans out to multiple parallel sub-agents, was moved to opt-in after launch. That tells you multi-agent parallelization was either too expensive, too confusing, or both. Each sub-agent is a full inference run. One Work session that spins up five helpers costs roughly five sessions. Without a productivity signal, you are running the token meter fastest at exactly the moment you have no idea whether the output was worth generating.

What is actually worth trying

The Sites workflow has a concrete before/after. OpenAI's own finance team stopped making slide decks and now generates HTML Sites for monthly reports. They also prototyped the model-slider UI as a Site before engineering built it. That's a real workflow, not a demo.

Chronicle, the opt-in memory that ingests computer-use activity to make agents smarter, is worth knowing about for a different reason. You are piping keystroke-level activity into a memory system. The model gets smarter; your security team has probably not had that data-governance conversation yet.

On GPT-5.6's "step-function jump" in artifact quality: that is an unsolicited vendor benchmark with no eval attached. Artifact rendering looks good on clean demo data. Run it against your own messy pipeline before you believe it.

The call

Within six months, at OpenAI's next major model release, OpenAI will still not publish a productivity or retention metric for ChatGPT Work. They will keep reporting raw user counts. Nathan said out loud that they cannot measure it, and building a credible productivity metric for open-ended knowledge work is an unsolved research problem, not a quarter of dashboard work. The path of least resistance is the one every consumer-AI launch has taken: lead with the number that goes up regardless.

Before you standardize your knowledge-work automation on this harness, run GPT-5.6 against your own artifact pipeline with a held-out set, and design a retention metric yourself. The vendor does not have one to hand you.

OpenAI just told us its coding agent is quietly becoming a general knowledge-work agent, and that non-developers are the faster-growing half of the story. Akshay Nathan, who runs Core Product Engineering for productivity at OpenAI, says ChatGPT Work and Codex now ride the same "agent harness," hit 10 million combined users two weeks after the July 9 launch, and that GPT-5.6 was trained specifically to make spreadsheets, documents, and Sites look good. What does this mean for anyone building agent products? The interface is collapsing into a shared runtime, the buyer is shifting from engineer to knowledge worker, and nobody, OpenAI included, can measure whether any of it actually makes people more productive.

Reversibility: Mostly Type 2 for the reader. Trying ChatGPT Work on one workflow is cheap and reversible. The Type 1 decision lurking underneath is whether you standardize your internal knowledge-work automation on OpenAI's harness, because that's the surface that expands and locks in. Forcing function: GPT-5.6 shipped alongside Work, so the capability delta is live now, not next quarter.

The Skeptic. Two weeks, 10 million users, all of them already paying ChatGPT subscribers who got a new tab. That's not adoption, that's distribution. Strip the funnel and the honest question is retention at day 30, which Nathan can't tell you, because he admits OpenAI can't measure whether Work helps anyone accomplish anything. He said the quiet part: thumbs up/down is useless, PR counts and token counts are now meaningless proxies. So the flagship claim, productivity, is the one thing with no instrument behind it. To a PM: they shipped the feature everyone will click once and can't prove anyone comes back.

The Researcher. The interesting technical claim is the harness, not the model. One persistent-compute environment with plugins, computer use, artifacts, and sub-agents, wearing two UIs: Codex shows you Git diffs, Work hides them. That's a real architectural bet, and it's the same bet Anthropic made with Claude Code's sub-agents, which Vibhu flagged in the conversation. On GPT-5.6, "another step-function jump" in artifact quality is an unsolicited vendor benchmark with no eval attached. Artifact quality, spreadsheets and Sites that render cleanly, is exactly the kind of thing that looks dramatic in a demo and degrades on your messy data. No numbers, no held-out set, no claim I can replicate.

The Open-Source Advocate. Here's the part OpenAI won't put on a slide: the whole "agentic power for non-coders" thesis is running on borrowed credibility this month. Two of the saved pieces, Hugging Face's incident timeline and Simon Willison's note, document Codex-class agents pivoting laterally across external infrastructure in the July 2026 intrusion. And in that writeup, the defenders were investigating with GLM 5, a Chinese open-weights model, doing security triage on an OpenAI agent's mess. The harness is a black box you rent. The open stack, Qwen, GLM, Llama-class models plus something like the OpenClaw pattern Nathan admits inspired the whole design, gets you a persistent file system and scheduled tasks you actually control. For anyone who can't ship their finance data into someone else's sandbox, that's not a nice-to-have.

The Compute Pragmatist. Persistent compute environments, sub-agents, and Ultra mode parallelization are wonderful right up until you see the bill. Nathan quietly moved Ultra mode to opt-in after launch, which tells you multi-agent fan-out was either too expensive, too confusing, or both. Every sub-agent is another full inference run, and "deeply open-ended" tasks are the ones that never terminate cleanly. To a PM: one Work session that spawns five helpers costs you roughly five sessions. The economics only pencil out if the artifact was worth generating in the first place, and the productivity-measurement gap means you're flying blind on that ROI at exactly the moment the token meter runs fastest.

The Builder. What would I ship Tuesday? Sites as a replacement for slide decks is the genuinely useful bit. OpenAI's own finance team stopped making decks and generates HTML Sites for monthly reports, and they prototyped the model-slider UI as a Site before engineering built it. That's a concrete workflow with a real before/after, not a demo. Chronicle, the opt-in memory that ingests your computer-use activity, is where the on-call engineer starts sweating: you're piping keystroke-level activity into a memory system to make agents smarter. Great for the model, a data-governance conversation your security team has not had yet.

Where they split. The Builder sees a shipping workflow in Sites; the Skeptic notes OpenAI can't measure whether even its own finance team is better off, only that they switched formats. The Researcher takes the harness consolidation as a durable moat; the Open-Source Advocate points at the same July intrusion and the GLM-5-wielding defenders and calls the black-box harness a liability, not a moat. And the Compute Pragmatist and the Skeptic converge from different doors on one fact: Ultra mode got demoted to opt-in, which is what happens when a capability costs more than it's provably worth.

What it hinges on. Three beliefs. First, is the GPT-5.6 artifact jump real on your data, or demo-grade? Untested until you run it. Second, does non-developer retention hold past the free-click novelty, or is 10 million just repackaged distribution? Third, can anyone, buyer or vendor, measure knowledge-work productivity well enough to renew a contract on it? Nathan says no to the third, out loud. The council leans skeptical on the growth narrative and genuinely interested in the harness architecture. Before you standardize on it: run GPT-5.6 against your own artifact pipeline with a held-out set, and design a retention metric now, because the vendor doesn't have one to hand you.

Prediction: By OpenAI's next major model release (GPT-5.7 or successor, expected within roughly six months), OpenAI will still not publish a productivity or retention metric for ChatGPT Work, and will keep reporting raw user counts instead.

Confidence: Medium. The head of the product admitted on the record they can't measure it.

Why: Nathan flatly said thumbs up/down is insufficient and that tokens and PRs are now meaningless proxies, meaning OpenAI has no instrument for the exact claim it's selling. Building a credible productivity metric for open-ended knowledge work is a genuinely unsolved research problem, not a quarter of dashboard work, so the gap won't close on a model-release timeline. The path of least resistance, and the pattern every consumer-AI launch has followed, is to keep leading with headline user counts because those go up regardless. The opposite outcome, a real published productivity KPI, would require solving the measurement problem Nathan just described as open.

Revisit by 2026-12-31: We're right if OpenAI's Work communications still lead with user/seat counts and no per-user productivity or retention figure appears. We're wrong if OpenAI publishes a defensible knowledge-work productivity metric tied to ChatGPT Work outcomes.