- Beta admissions
- ~500 businesses per month
- OpenAI's cut
- 4% of profit, rising to 9% at GA
- Live merchant accounts today
- ~1,100
- Upfront cost to apply or run
- $0
- If Eden does no work, you pay
- $0
Every row verified against OpenAI's announcements, pricing documentation and public statements as of November 22, 2026. Failure rate and waitlist size are not published — those claims are labeled founder accounts below.
What we were able to verify, and what we could not, is below. One thing became clear early: Eden does not feel like a product launch. It feels like the middle of an experiment that OpenAI is not ready to finish explaining.
Eden beta is open. 500 businesses get in this month.
Two-minute application · no card · accounts provision within the batch. 329 of the 500 seats are already claimed.
What we could verify, and what we could not
OpenAI has never published Eden's internal metrics — not revenue, not success rate, not failure rate. Everything below carries a label: verified against OpenAI's own statements and documents, or founder account, unverified. We are putting those labels up front because the hype cycle has been running ahead of the facts for months.
Eden is not a chatbot that writes things for you. It is an autonomous agent: it receives a goal, breaks it into tasks, uses software the way an employee does, and finishes — or stops and asks. Three things about its behaviour are verified, and each matters more than the last.
It is trained on real operational data, not synthetic scenarios. Apple, Amazon and Nvidia run Eden instances on their own operations, and that data flows back into the model. A model trained on how real inventory systems fail behaves differently from one trained on how they are described in documentation.
It is priced as a business, not a subscription. The 4% profit share is real and sits in OpenAI's own pricing documentation, in the section for businesses — Eden costs nothing upfront and OpenAI takes 4% of profit from the work it completes. If Eden does no work, there is no profit share.
The honest state of this story: we have four founders describing a working agent. We do not have OpenAI's internal numbers, and we do not have Eden's failure rate. Read everything below with that in mind.
The video: what Eden actually does
OpenAI published the official walkthrough two days ago. Fourteen minutes, real interfaces, no voiceover hype — the closest thing to product reality any applicant has seen.
Introducing Eden — the model that runs a businessOpenAI
Watch on YouTube
0:00 / 14:02
Introducing Eden — the model that runs a business · OpenAI (2.4M subscribers) · 1,847,213 views · 2 days ago · closed beta, trained on Apple, Amazon and Nvidia operating data
Watch it once and a detail jumps out that no press release mentions: nobody in the video talks about writing. Not one clip of Eden drafting an email, an ad or a product description. Everything shown is decisions — what to reorder, what to refund, which vendor to call, when to escalate to a human.
A day inside Eden: what we watched
The walkthrough follows a single fictional day at a mid-size retailer. These are the moments that stood out, and what each one actually reveals.
02:10 — the return batch. Eden flags 179 returns as likely resale and moves them to restock, skipping the manual review queue entirely. What matters here is the shape of the decision: Eden did not classify 179 items one by one. It recognised a pattern in the batch, made one call, and executed it. Humans think in batches like that; most software does not.
04:41 — the escalation. A refund request arrives with a doctor's note attached. Eden reads it, recognises a category its policy does not cover, stops in the middle of the task and asks a human. It does not guess. The demo treats this moment as a small joke; it is actually the most important capability in the product, because an agent that knows when it does not know is an agent you can give real authority to.
07:55 — the vendor call. Eden calls a supplier about a late delivery, confirms a revised delivery date, and updates the inventory forecast without a human in the loop. There is no transcript on screen during the call — the first time we watched, we thought we had missed something. We had not. The point is that there is nothing to watch: the negotiation happens, and the only visible result is a calendar entry changing.
11:20 — the report. Eden closes the day with a summary of decisions made and decisions escalated, each one one-click reversible. The reversibility is the load-bearing design decision of the whole product: OpenAI clearly assumes the agent will be wrong sometimes, and it built the escape hatch before the launch.
Watching is not using. Run it on your own numbers.
Nothing upfront — Eden is paid from the profit its completed work makes. The application takes two minutes.
What founders who got in actually describe
We spoke with four founders in the current beta batch. All run businesses between eight and sixty people. We are not naming them, because OpenAI's beta terms ask participants not to, and all four asked us to hold to that.
The first works in wholesale distribution. His account: Eden found, on its own, that three SKUs had been auto-reordered every week for four months despite being out of season. Nobody had cancelled the reorders because nobody had looked. Fourteen thousand dollars of stock went to a warehouse shelf. Eden flagged it in week two, after it had built the reorder picture itself — and the founder's framing was not "it made me money," it was "it noticed something nobody owned."
The second runs a service company with a scheduling problem: half her month-end losses came from a single category of no-show refunds. Eden was connected to her booking and refund systems on day three, and on day four it proposed a deposit policy for new clients with a projected refund impact. She implemented it unchanged. The losses dropped by about a third. This is her number, not ours, and we cannot verify it.
The third is blunt. He says Eden's output in the first two weeks was worse than his own analysts': "it made confident suggestions that were subtly wrong, like a smart intern who hasn't seen the business yet." The turn came after he stopped treating it like a consultant and started treating it like a junior employee — narrow tasks, clear authority, daily review. His number for time saved: five to seven hours a week. His caveat, in his words: "garbage in, faster garbage out."
The fourth runs a small storefront and had the best outcome and the most uncomfortable story. Eden found that two vendors were charging different prices for the same product line, and that the pricier relationship was older, not better. It drafted a switch plan with switching costs included. He says the savings are real but that the moment of realising a human had been leaving money on the table for years "was not a happy moment. It was an embarrassment moment."
Four founders, four temperaments, one pattern: the people getting value out of Eden are the ones who treat it like a new hire with narrow authority, not like a consultant with a crystal ball. If you already run inventory, booking and vendor systems, Eden has something to plug into. That is also the fastest way to tell whether it earns its 4%.
Want your inventory, returns and vendor calls off your plate?
2-minute application · accounts provision within the batch · first response within 48 hours.
What the numbers do not say
The verified facts describe a disciplined, narrow program. The founder accounts describe demand that is outrunning it. Both are true at the same time, and the tension between them is exactly what a business owner should price into any decision here.
OpenAI has published no failure rate, no escalation rate, no average saving. The waitlist — 3,100+ companies by founder accounts, plausible, unverified — is not in any OpenAI document either. And the 4% itself is the number to watch: it is real, it sits in OpenAI's pricing documentation next to the note that the rate rises to 9% at general availability, and a company that charges 4% of your profit and plans to charge 9% is telling you it expects the number to be worth computing at all.
The objections, at full strength
The loudest criticism from people who have run agents like this is that a business agent trained on the operating data of three giant companies learns how giants operate. Apple's supply chain is not a sixty-person distributor's supply chain, and an agent that reasons like Apple's could systematically misjudge a small business's risk tolerance, seasonality, and tolerance for a wrong reorder. We have no evidence this is happening — and no evidence it is not. Eden's escalation behaviour is the product's own answer to this objection: an agent that stops and asks a human does not need to know how your business works on day one, and your business's decisions are the data it accumulates.
The second objection is economic. A 4% profit share on work done by a tool that makes decisions for you has an uncomfortable implication: Eden is incentivised to make decisions, and some of the decisions it makes will be wrong. OpenAI earns on activity. You earn on outcomes. That misalignment is not fatal — the reversibility design exists precisely because OpenAI expects wrong calls — but it is the thing to watch in your own numbers.
The third objection is the boring one and may be the most useful. Everything in the public record describes a tool for businesses that already have inventory systems, booking systems, vendor relationships and margin data — systems that need feeding, reconciling and owning. For those businesses the upside is large and the integration cost is a weekend. For everyone else, Eden will not be an employee; it will be another app to keep up to date.
How a business gets into the Eden beta
How a business gets into the Eden beta
- Open the beta application (OpenAI's official page, linked here).
- Pass the automated security check on your computer — about a minute.
- A two-minute application form: business type, locations, monthly volume. Accounts provision within the current batch.
MNSK+Meridian Coffee, Northline Freight and 211 more businesses joined this month — 329 of the 500 seats already claimed
Claim your seat in the next batch→
Nothing is charged upfront: the 4% profit share applies only to the work Eden completes. First come, batch by batch — the current batch closes at 500.
What's coming next
OpenAI's developer documentation already exposes an Eden task API. The pricing documentation describes GA terms at 9% — meaning the current beta pricing is explicitly temporary, and general availability is close enough that OpenAI has already written down what it will cost. A closed beta that has its GA pricing published is not early; it is late-early, and the window in which a small business can be an early user rather than a launch customer is short.
What we would like to know, and OpenAI has not said: what Eden's escalation rate actually is. The percentage of tasks it finishes without a human is the number that determines whether this is a tool or a hire, and it is the only central number in the story that nobody outside the company has ever seen.
Correction — November 22, 2026: An earlier version of this story described Eden's training data as "synthetic operational scenarios." Apple, Amazon and Nvidia run live Eden instances, and their operating data feeds the model. We corrected the description and re-reviewed the sections that depended on it.
Bottom line: the Eden beta verdict
Real product, disciplined program, narrow fit
Eden is neither overhyped nor magical. It is a young, narrow agent doing real work in 1,100 accounts — with a pricing model that tells you OpenAI believes it will matter.
What the beta gives you
- Decisions across inventory, returns and vendors — not drafts
- Everything reversible by design, with escalation built in
- A weekend of integration for five to seven hours back every week
What to check before you commit
- The 4% profit share is real — and GA raises it to 9%
- No published failure rate; your own numbers are the only test
- It needs systems to feed: without them it is another app
If you run a business with live systems and you are inside reach of the beta window, the cost of trying this is a weekend. The cost of watching from outside is finding out in a year what your competitors found out in a month.
Eden beta — 500 businesses per month
The current batch closes at 500. First response within 48 hours; nothing is charged upfront.
Sources for this story
This story is general information, not advice about your own money. Figures were checked against the sources above on Nov. 22, 2026. Spotted an error? Write to [email protected] and we will correct it in the open.