OpenAI Wants a Billion People Using Agents. Its Own Rate Card Says $100 to $200 a Month

A ChatGPT Work agent session running on a laptop in an open-plan office, with a subscription price card visible beside it

OpenAI is pitching ChatGPT Work as the agent that brings the rest of the office along. Its help center already publishes what an average month of that behavior costs, and the number sits five to ten times above the subscription it rides on.

TechCrunch spent a week inside OpenAI asking whether normal white collar workers will adopt AI agents the way software engineers did. The company’s own numbers answer a sharper question. In June 2026, 98% of OpenAI employees used Codex, against 17% of organizational subscribers and under 1% of individual ones. OpenAI staff pay nothing per token. Everyone else pays by the token, on a meter OpenAI installed in April and rarely mentions in the marketing. The adoption gap and the pricing gap are the same gap.

Andrew Ambrosino, who leads engineering on OpenAI’s desktop app, has handed the thing his inbox, his Slack, his Notion, his Figma and his phone. He told TechCrunch he accepts the odds that it leaks a private DM into a document he asked it to write, and that he will take the personal hit for the job.

He also asked the question that matters more than any of that. Reviewing which workflows his team should design around, he said the team has to keep checking whether they are building for what everyone else does, or whether they are weird.

They are weird. Not because OpenAI engineers work differently from accountants, though they do. Because they are the only population using this product who never see the bill.

What Happened

Tim Fernholz published an inside look at OpenAI’s push to move agents past developers on August 24, 2026. The product at the center is ChatGPT Work, launched July 9 on GPT-5.6, available on the $20 Plus tier, with the standalone Codex app folded into a single ChatGPT desktop app. Tibault Sottiaux, who runs core product including Work, told TechCrunch the point is for ChatGPT to complete entire complicated tasks on its own.

Fernholz tested it. It pulled his son’s preschool calendar out of a badly formatted email and into Google Calendar. It built him an auto-updating dashboard of financial metrics on companies he covers, and a queryable database of space launches he had previously assembled with Python scripts. Setting up permissions on a cloud drive failed repeatedly until a mobile dialog told him read-only access would not work at all.

Then the line that should have led the coverage. Over four days of casual use he burned more than 80 million tokens, which the model priced at $65 when he asked it. There is no usage dashboard in the app.

Sitting underneath all of it is an OpenAI-backed study, “The Shift to Agentic AI: Evidence from Codex” (Johnston et al., arXiv 2606.26959, June 2026). It reports the 98/17/1 split. It also reports that Codex generates 99.8% of weekly output tokens produced inside OpenAI, and that the median OpenAI employee in a legal role produced 13 times more monthly output tokens in June 2026 than in November 2025. For the median researcher, more than 50 times.

The Backstory

OpenAI released Codex in April 2025 as a coding tool, in the one commercial domain where checking the work is close to free. Non-engineering teams inside OpenAI started using it while it still spoke to them in diffs. Ambrosino says his team spent February through June making it general purpose.

Two dates from spring 2026 explain the business more than the July launch does.

On April 2, OpenAI repriced Codex from per-message billing to token-based billing, in its words to align with API token usage. Plus, Pro and Business customers moved first, all Enterprise plans followed on April 23. A week after that first switch, on April 9, OpenAI added a $100 monthly Pro tier offering five times the Codex usage of the $20 Plus plan, sitting under the existing $200 tier at twenty times.

Read those two moves together and the sequence is a company converting a subscription product into a consumption product, then building the ladder customers will climb when the consumption shows up. The July launch put that meter behind a much larger door.

The Codex rate card also states, in a single line most readers skip, that Codex, ChatGPT Work, ChatGPT for Excel and Workspace Agents all draw from the same agentic usage and credit pool. ChatGPT Work is not a Plus feature. It is a metered product with a Plus door.

The Plan

Sottiaux frames the commercial logic as value creation preceding willingness to pay. Generate enough utility, and the customer decides $20 a month is obvious.

The trouble is that OpenAI publishes a competing number. On the same rate card page that governs Plus, Pro, Business, Enterprise, Edu, Health and Gov plans, the company states that Codex costs roughly $100 to $200 per developer per month on average, with large variance by model, concurrency, automations and fast mode.

That figure is not an analyst estimate or a leaked internal deck. OpenAI wrote it, in its own help center, about the workload it has now placed on a $20 plan and a free tier.

Bar chart comparing the $20 ChatGPT Plus price against OpenAI's own stated average Codex cost of $100 to $200 a month and a $488 monthly figure extrapolated from TechCrunch's four-day test

The bottom row is our arithmetic on the reporter’s own disclosure, not an OpenAI figure: $65 across four days is $16.25 a day, and thirty days of that is $488. It assumes flat usage and it describes what OpenAI charges for those tokens, not what they cost OpenAI to serve. Treat it as the retail value of one journalist poking around, and note that a journalist poking around is the light end of what a finance team running weekly reports would do.

The Business Model Angle

The study OpenAI commissioned to prove agents work measures that success in output tokens. Look at the rate card and you find that output tokens on GPT-5.6 Sol carry 750 credits per million against 125 for input. Output costs six times input. On the security-cleared GPT-5.5 Cyber model it runs 1,875.

So the metric OpenAI chose as evidence of adoption is denominated in the most expensive unit it sells. A median researcher generating fifty times more output tokens than seven months earlier is a productivity story if you work at OpenAI and a cost-of-goods story if you buy from OpenAI. Both readings come from the same paper.

Which brings back the dogfooding problem. Ambrosino’s team designs from the workflows of a population with a marginal price of zero, then ships to a population billed per token with no dashboard to check. That is not a small calibration error. It is the difference between a product built for people who never stop and a product bought by people who will learn to stop.

Three ways this resolves, and OpenAI is playing all three.

Deflation. Sottiaux points to a recent 80% price cut on the Luna model and promises the same tasks will cost less in six months. On the current rate card Luna already runs at 30 output credits per million against Sol’s 750. Routing cheap work to cheap models is the honest version of the fix, and it is the same move Lovable made when it raised $400 million against its cost line.

The ladder. The $100 Pro tier exists because heavy users exist. Every credit-hungry customer OpenAI converts from Plus is a 5x ARPU event, and the conversion trigger is a limit the customer hits rather than a price the customer negotiates.

The invisible dial. OpenAI has adjusted the tier-to-model mapping and the credit rates repeatedly, including a July 30 change making Terra and Luna consume fewer credits. Each adjustment changes what $20 buys without changing the $20. Storefront price stability with a floating allowance behind it is the most valuable pricing instrument in the business, and almost no customer audits it.

Meanwhile the differentiation argument is weaker than the pricing one. Joe Gershenson, who leads OpenAI’s harness engineering, told TechCrunch the harness is a temporary crutch that the next model will obsolete. Databricks benchmarked Pi, an open-source harness from Mario Zechner’s company Earendi, and found it beat Codex running the same GPT-5.5. Zechner’s read on why the labs push their own harnesses anyway is blunt: own the whole stack, or become a model provider competing with Chinese models on price. ChatGPT Work looks less like a product moat and more like a metering surface with switching costs baked into the setup wizard. On first run, it offered to import Fernholz’s Claude Cowork data.

The Risk

The permission ratchet. Fernholz tried to grant read-only access to a cloud drive and could not. Full access or nothing. Ambrosino has granted everything and accepts the leak risk on principle. A product whose usefulness scales with permissions, sold to accountants and doctors, has an incentive structure that runs one direction only.

No meter in the cockpit. A consumption product without a visible usage readout works for the seller until the first surprise invoice. Enterprise buyers who lived through cloud cost overruns will ask for it before they roll it out, which is exactly the sort of drag that keeps organizational adoption at 17%.

Evaluation. Most office work lacks the pass-fail structure code has, which is why the coding domain moved first and why OpenAI leans on the GDPVal benchmark across 44 occupations. We ran that argument at length in our piece on why verifiability, not intelligence, is the real rate limiter, and it applies here without needing a rerun.

Adoption is the margin. The under-1% individual figure protects the $20 tier today. OpenAI cannot describe that number as a failure of discoverability and simultaneously bank on the price holding if it gets fixed.

Quick Questions

Is ChatGPT Work included in the $20 plan? Access is. Consumption draws from a shared agentic credit pool that also serves Codex, ChatGPT for Excel and Workspace Agents. Hit the ceiling and you buy credits or move up a tier.

Where does the $100 to $200 figure come from? OpenAI’s own Codex rate card help article, stated as an average per developer per month with large variance by model and concurrency.

Did OpenAI say how many people use ChatGPT Work? No. The company would not split Work from Codex. The merged app has around 20 million users against more than a billion people prompting ChatGPT.

Is the harness a real competitive advantage? OpenAI’s own harness lead says it is temporary. Databricks found an open-source harness beating Codex on the same underlying model. The moat argument currently rests on the model, not the wrapper.

Does cheaper inference solve this? It helps. Sottiaux cites an 80% cut on Luna. Agentic use also grows the token count per task, and the study shows task complexity climbing, so unit price and unit volume are moving in opposite directions.

The Business Model Analyst Take

TechCrunch asked whether everyone will use agents. OpenAI has already priced its answer.

A company that expected mass adoption at $20 would not have moved Codex off per-message billing in April, would not have inserted a $100 rung a week later, and would not keep quietly resetting how many credits a task consumes. Those are the moves of a company that knows the honest price of the behavior it is promoting, publishes it in a help article, and would prefer customers meet it gradually.

None of that makes the product bad. The preschool calendar worked. The financial dashboard worked. Turning a Python scripting job into a sentence is real value, and we have written before about how firms still cannot find that value in their own productivity numbers even as workers swear by it.

The thing to watch is not adoption. It is whether the $20 tier still buys a usable amount of agent work in six months, and whether OpenAI ever ships the usage dashboard. A consumption business that hides the meter is making a bet on customer inattention, and inattention is the one input that gets more expensive as the bill grows. The company that gets AI agents into a billion hands will be the one that shows the customer what a task cost before the customer has to ask the model.

For the underlying economics of counting AI work as labor, see our breakdown of bionic head count and what a single agent really costs, and our OpenAI business model teardown for how the company got here.

UNLOCK THIS FREE DOWNLOAD

DOWNLOAD NOW

Fill Your E-mail to Receive this Download Directly in Your Inbox.

RECEIVE OUR UPDATES

The Biz Model Club

Get daily, no-fluff insights on the latest business models, startup strategies, and trends delivered straight to your inbox.