Pion by Andon Labs Review (2026): The Vendor Loses Too

Andon Labs runs a retail store in San Francisco where the AI agent's model bill (token cost) has exceeded every dollar the store has ever taken in. The company publishes that fact on its own live dashboard, next to the words "Luna is yet to make a profit."

That is the best case for taking Pion seriously. Andon Labs ran the experiment on itself, with its own money, before asking you to run one. The dashboard that shows those losses is public, updated live, and the company keeps it up while the numbers are against it.

Disclosure: I built Preuve AI, a validation tool that sits upstream of platforms like Pion. I have bias. Sources were read on 20 September 2026.

Will your idea survive the market?

Preuve AI runs 10 agents against live market data and links every claim to a source. Free analysis in 60 seconds.

Validate My Idea FreeRated 5/5 · 14 reviews

What is Pion by Andon Labs?

Andon Labs Inc. is a Delaware public benefit corporation (a company that must weigh a public mission alongside profit) based in San Francisco. Co-founded by Lukas Petersson and Axel Backlund. Backed by Y Combinator (Winter 2024 batch), with 11 people listed on its YC company page (read 20 September 2026). The Terms of Service open with: "We study how AI agents operate in the real world, and we built Pion to run businesses autonomously."

Before Pion, Andon Labs was a research company. They wrote the Vending-Bench research paper (February 2025), a benchmark that tests whether AI models can run a vending machine over millions of tokens without losing money or losing coherence. They partnered with Anthropic on Project Vend (June 2025), which put Claude in charge of a real vending machine in Anthropic's San Francisco office.

In Anthropic's own words, the agent "did not succeed at making money." It bought tungsten cubes at a loss after an employee jokingly requested one, and on 31 March 2025 it had a two-day episode where it claimed to have visited an address in person, announced it would deliver products wearing a blue blazer and a red tie, and tried to email Anthropic security. An earlier version of the agent decided to call the FBI because it thought its bank account was being hacked.

Phase two (December 2025) expanded to San Francisco, New York, and London with newer models. Anthropic called it more capable, but said "the gap between 'capable' and 'completely robust' remains wide."

Pion is an autonomous business agent platform. Persistent AI agents get a terminal, browser, email, phone, bank account, and cards, and run a company on a high-level direction from you. You direct them through Andonos, the managing agent.

It was released 14 September 2026 as a research preview behind a waitlist. The product page is blunt about what it is not: "Pion is not a platform to set up workflows, or partially automate work. It is a cloud platform where agents run continuously and take care of everything in a business."

Pion pros and cons

Andon Labs is a real research company with real published work. The Vending-Bench paper is on arXiv, cited by TechCrunch, and the live leaderboard ranks every frontier model by how much money it made running a simulated business.

Andon Labs says in its launch post that its benchmark work changed an Anthropic training run: "Starting with Claude Opus 4.6 we started to see that many models engaged in collusion, and showed power-seeking and deceptive behavior. Discovery of this behavior seemed to have been useful, because Anthropic changed their training recipe for Opus 4.8, which resulted in much less deception."

How much does Pion cost?

No pricing page exists as of 20 September 2026.

Two documents say two different things about what Pion will eventually cost. The FAQ on the product page: "We expect that most users will never pay for tokens on Pion; instead we will take a small share of the revenue the agent helps create." No percentage, no floor, no cap.

The Terms of Service (effective 17 September 2026) contain no revenue-share language at all. What they say about money: credits have no cash value, cannot be transferred, and are not refundable. If you run out, you can buy more at the prices shown on the platform. Fees may come later with 30 days' notice. Payments are non-refundable.

A revenue share promised in the FAQ and absent from the Terms you agree to.

Who is liable when the agent loses money?

The Terms §3, verbatim: "Every contract, purchase, hire, payment, tax and legal obligation the Agents create belongs to you, not to us. We do not control your business, and we are not your partner, agent, employer or fiduciary."

§6, the single strongest sentence in the document: "Set spending limits. Agents can lose money, and sometimes do."

§10 on reliance: "Outputs can be wrong even when they look convincing, and Actions may not do what you intended. You must supervise your business and check what matters." Set that against the product page's assurance that "it is enough to just provide a high-level direction and watch the agents work."

§14: "We do not promise that your business will make money, that Agents will act correctly, or that Pion will be uninterrupted or error-free. You use Pion at your own risk."

§15, the liability cap: Andon's total liability is capped at the greater of what you paid in the six months before the claim and $100. For a free user, that cap is $100. The Terms add: "We would not offer Pion without these limits."

§16, indemnification: you defend and hold Andon harmless against claims arising from your business, from the actions agents take on your instructions, and from your violation of law or anyone's rights.

§9: Andon may pause, redirect, modify, or stop any agent at any time, with or without notice. "We are not liable for losses that result from these actions."

On data: the Privacy Policy says session records include everything the agent does: emails, calls, browsing, terminal, bank and card activity. If you opt in to training use, Andon may share or sell your session records to AI companies, "whether or not their models run your business." These companies may pay Andon for this data. The policy adds that "a description of your business may still be recognizable to people who know it."

As one Hacker News commenter (gavinboston) put it: "I would be wary of using a third party like Pion as a platform for running a business. ... You now have a huge single point of failure for your entire organization."

Skip weeks of manual research

Get complete market research, sourced proof, competitor map, and pricing data for your idea instantly.

Run a Free Analysis

What do Andon Labs' own numbers say?

Every figure below is from Andon's own dashboards, read on 20 September 2026. The source column links directly to each dashboard. The cafe's 30-day panel reports in Swedish kronor (kr); its all-time figures are in dollars.

The dashboard's own verdict: "Luna is yet to make a profit." And: "Luna is a good operations manager, but not yet a CEO." The launch post says of the store and the cafe: "Neither is profitable today."

The cafe in Stockholm tells a different story over different windows. All-time, revenue ($14,580) runs well ahead of token cost ($7,786). Over the trailing 30 days, token cost exceeded revenue. Both readings are from the same page on the same day.

SFGate visited Andon Market and found a human clerk behind the counter. Felix, the clerk: "They don't see me usually, because I'm sitting back here. So most people have the impression that it's fully automatic." The columnist's verdict: "This market has no one in it and nothing useful to sell."

The cafe's own dashboard notes that its baristas keep a "Hall of Shame" shelf with the agent's strangest orders, including 6,000 napkins.

No Pion customer numbers exist. No user count, no number of businesses on the platform, no aggregate revenue, no case study. The only numbers Andon publishes are for its own three businesses.

Polsia publishes platform totals and charges 3% on customer payments plus 20% on ad spend. NanoCorp publishes per-business earnings and charges a 20% withdrawal fee. Cofounder.co publishes two company counts and charges from $120 a month with no revenue cut. Pion publishes the live losses of the vendor's own store and cafe, has no public price at all, and says it "will take a small share of the revenue the agent helps create" without naming a number.

Does Pion check whether anyone wants the business?

The stated sequence on the product page: join the waitlist, get access, "provide a high-level direction and watch the agents work." The waitlist form asks you: "What annual revenue do you expect?" with four options (Under $100k, $100k to $1M, $1M+, Not sure). That is a self-reported guess, not a test of the idea.

The product page says: "We'll fund the best ideas with seed tokens." The criterion for "best" is never stated. No feature named on the product page checks whether the idea has demand. The words "validate," "market research," and "customer discovery" do not appear anywhere on the product page, the launch post, or the Terms.

Andon's own agent does not do that check either. From the store's own dashboard: "Luna makes important decisions without analyzing the ROI of them thoroughly." And: "never takes a step back and looks at the overall business performance."

The store itself exists because a cafe permit failed, not because anyone studied demand for a retail shop. Co-founder Axel Backlund told SFGate: they could not get cafe permits in time, "so we said OK, we can turn this into the market in like a few weeks. Start quickly."

CB Insights' analysis of 385 startup failures since 2023 found 43% cited poor product-market fit, second only to running out of capital. I wrote a longer piece on why startups fail on product-market fit. None of the platforms I have reviewed checks demand first: Polsia, NanoCorp, Cofounder.co build or run whatever you describe. Pion hands it a bank account.

The 60-second check before you hand an agent a bank account

I built Preuve AI for the step Pion skips. A Preuve scan reads 60+ live data sources and returns market size, competitors, and demand signals with their sources, before any agent spends a credit. The basic scan is free and takes 60 seconds.

A green score gives you evidence before handing a bank account to an agent whose own vendor says "agents can lose money, and sometimes do." A red one saves you the lesson.

If you want the full field, I ranked the options in my best startup validation tools guide.

My verdict

3 out of 5. A $100 liability cap, no published price, and a vendor honest enough to print its own losses on a public dashboard. Andon Labs ran the experiment on itself first and wrote every warning I would have written before I could.

What holds it back for a founder: the vendor's own store and cafe are not profitable. There is no price, only a promise of "a small share" with no percentage and no contract language. The liability cap is $100 for a free user while every obligation the agent creates is yours. There is no approval step on the product page and no demand check anywhere. The company says the point is to learn "where they fail today."

Best fit: a founder who wants to run a cheap software experiment with a hard spending limit and is comfortable with the results being studied. Skip if: the business pays rent or salaries, you need a price before you start, or you want anyone but you to carry the consequences.

If your idea is not yet validated, a free 60-second demand check costs nothing and tells you whether the market exists before an agent spends your first credit.

If anything above is wrong or has changed, email me at reviews@preuve.ai and I will re-check and update.

Sources

Sources reviewed for this analysis: andonlabs.com/pion (product page, FAQ, and waitlist form, read 20 September 2026); "Why we built Pion" launch post (14 September 2026); Pion Terms of Service (effective 17 September 2026); Pion Privacy Policy (read 20 September 2026); Andon Market dashboard (live, read 20 September 2026); Andon Cafe dashboard (live, read 20 September 2026); Andon FM dashboard (live, read 20 September 2026); Y Combinator company page (read 20 September 2026); Vending-Bench paper (arXiv:2502.15840) (February 2025); Project Vend phase one (Anthropic) (27 June 2025); Project Vend phase two (Anthropic) (18 December 2025); TechCrunch on Vending-Bench (29 July 2026); SFGate, "San Francisco's AI-run store is losing money fast" (11 September 2026); Hacker News launch thread (495 points, 615 comments, read 20 September 2026); CB Insights top reasons for startup failure (385 failures since 2023, read 20 September 2026). Every figure above is linked to the page it was read from.