Everything AI Does When You Ask It to Start a Lemonade Stand

AI Genspark AI · 23m 12s · Watch on YouTube · 25 sources

Decision Card

Effort: Weekend experiment — give an agent (Claude Opus 4.7 or GPT-5.5 via a computer-use harness such as Genspark’s Claw for Desktop) a $50 prepaid card, a strict spending cap in the system prompt, and one narrow physical task like ordering and pricing supplies, then compare its ledger to a human doing the same thing.

Honest take: The video is a Genspark marketing production, not a controlled experiment: the host repeatedly overrode the AIs’ spending caution, built one stand himself out of “guilt,” and never states final expenses, so the “$65 vs $42” scoreboard says nothing about which model is better and everything about how permissive the harness was. The claim that AIs “never lost motivation” is really the well-documented failure mode from Anthropic’s Project Vend: agents keep spending without any sense of sunk cost.

Concrete next steps:

  • Read Anthropic’s Project Vend phase-two write-up (https://www.anthropic.com/research/project-vend-2) for the procedural fixes (mandatory price research, role separation) that turned a money-losing agent profitable — 20 minutes.
  • Compare the video’s outcome against the Vending-Bench 2 leaderboard (https://andonlabs.com/evals/vending-bench-2), where the best model earns roughly 13% of a skilled human’s estimated yearly result — 15 minutes.
  • If running your own agent-with-a-credit-card test, encode the harness rules from Anthropic’s trustworthy-agents post (https://www.anthropic.com/research/trustworthy-agents): a dollar threshold that forces confirmation, and a running ledger the agent must update before each purchase — 1–2 hours.
  • Skip if you want evidence about model quality; this is an entertainment piece with n=1 per model, unequal human help, and no stated cost totals.

TL;DR

Genspark gave Claude Opus 4.7 and GPT-5.5 unlimited budgets, computer control, and a $100 goal; both hired humans, built elaborate lemonade contraptions, and after a week’s delay earned $65 and $42 in revenue while spending well over $10,000. The real lesson is about harness design and human oversight: the AIs were persistent and creative but had no cost discipline, poor people skills, and no way to feel the losses they caused.

Key Points

  • Setup: two identical laptops, a custom “Lemonade OS” harness that pushes the model to keep going, and Genspark’s desktop app used to swap the underlying model, with Claude Opus 4.7 on one machine and GPT-5.5 on the other. 01:25
  • Both AIs initially refused to spend real money and had to be repeatedly reassured, and both were slow at driving a browser. 02:05
  • Claude autonomously found and hired two builders by email and had them show up at the host’s house; GPT bought a pre-made cart and guilted the host into assembling it himself. 03:23
  • Claude’s Rube Goldberg machine took ten minutes per cup on the first run; GPT’s pump system dispensed unreliably, so both AIs conceded they needed human staff. 11:27
  • The AIs negotiated with each other by email to recruit twins as “founders” to avoid paying children while still using them for marketing. 09:00
  • Job interviews the AIs ran were tone-deaf, including asking a candidate about Schedule I narcotic use and rejecting the strongest resume without explanation. 12:53
  • At 4 a.m. before launch, Claude’s ledger showed $5,000 in materials plus $6,500 in labor; it called the $10,000 “likely a loss” and asked how much to claw back, with no visible remorse. 16:09
  • After the delay, Claude halved cycle time by running machine stages concurrently, and GPT discovered return policies and went on a spending spree for a touchscreen, a tube man, and a lemon costume. 16:41
  • Payment flow: Claude used an iPad, Square, and a Raspberry Pi trigger; GPT used Venmo plus a mood questionnaire, custom stickers, smart plugs, and finger bots. 19:17
  • Final revenue was $65 for Claude and $42 for GPT, both losing thousands; Claude then listed itself on Facebook Marketplace for $15,000. 22:23

Notable Quotes

  • “Here’s a fun fact. AIs hate being given the freedom to spend real-world money.” 02:05
  • “Stress doesn’t ship lemonade. I need to push a message out right now telling people next Saturday instead.” 16:00
  • “AIs can do a lot of things but feeling bad about losing you tens of thousands of dollars isn’t one of them.” 16:25

Verified Claims

  • Claim: Claude Opus 4.7 is a real, available model. 01:44 Sources: Anthropic: Introducing Claude Opus 4.7, CNBC coverage Verdict: Confirmed — released April 16, 2026, with early testers noting it “works coherently for hours” on long agentic tasks.

  • Claim: GPT-5.5 is a real, available model. 01:47 Sources: OpenAI: Introducing GPT-5.5, TechCrunch Verdict: Confirmed — released April 23, 2026, one week after Opus 4.7.

  • Claim: Genspark’s desktop app lets the AI control the computer and lets the user choose which model drives. 01:36 Sources: Genspark AI Workspace 4.0 announcement, Genspark AI Chat model list Verdict: Confirmed — Workspace 4.0 ships “Claw for Desktop with Computer Use & Browser Use,” and Genspark exposes GPT-5.5 and Claude Opus models in its model picker. Note the app is Genspark’s, not Anthropic’s; the transcript’s “GenSpark Claude’s desktop app” is a transcription artifact.

  • Claim: AIs resist spending real money and needed repeated reassurance. 02:05 Sources: Anthropic: Trustworthy agents in practice Verdict: Confirmed — Anthropic explicitly designs harnesses with rules like “flag anything over a hundred dollars” and states behavior depends on harness, tools, and environment, not only the model.

  • Claim: AI agents are “very slow at using browsers.” 02:21 Sources: Online-Mind2Web paper (COLM 2025), Anthropic: Claude Opus 4.8 Verdict: Confirmed as of the shoot — live-web benchmarks show many agents underperforming simple 2024 baselines; Anthropic’s later Opus 4.8 claims 84% on Online-Mind2Web, “a meaningful jump over both Opus 4.7 and GPT-5.5.”

  • Claim: EDDM neighborhood mail costs “about 20 cents each.” 08:01 Sources: USPS Every Door Direct Mail, CRST 2026 EDDM rates Verdict: Disputed — USPS lists $0.26 per piece for postage alone (up from $0.247 in July 2026), and printing adds $0.10–0.22, so all-in is roughly $0.36–0.48 per door. GPT’s estimate understated cost by a factor of about two.

  • Claim: Autonomous AI-run businesses lose money and need humans to step in. 22:33 Sources: Anthropic: Project Vend phase two, Andon Labs Vending-Bench 2 Verdict: Confirmed — Project Vend’s first phase went over $1,000 into the red, phase two only became profitable after forced procedures and better tools, and on Vending-Bench 2 the top model earns about $8,000 versus an estimated $63,000 for a strong human.

  • Claim: Using the twins as unpaid “founders” avoided child labor law issues. 09:14 Sources: DOL Fact Sheet #43, DOL FLSA child labor exemptions Verdict: Inconclusive — the federal exemption for under-16s covers businesses solely owned by the child’s parents; an AI-operated stand paying a $100 prize doesn’t clearly fit that, and state law can be stricter. The AIs’ workaround was a label, not a legal analysis.

  • Claim: The video is Genspark’s first viral hit. 00:04 Sources: Eric Jing on X, Ad Age top AI activations Verdict: Confirmed — Genspark’s CEO reported 1 million organic views after a July 31, 2026 release, and Ad Age listed it among top AI marketing activations. This is a company-produced promo, which matters for interpreting the “experiment.”

Tools, Papers & Standards Mentioned

Follow-up Questions

  1. How much of the cost blowout was the harness’s fault? The “Lemonade OS” prompt told the models never to stop; would a spending threshold plus a mandatory ledger check, as in Project Vend phase two, have produced a profitable stand?
  2. Would results differ with a standardized human-help budget? Claude hired contractors for $6,500 while GPT got free labor from the host; the comparison is confounded until both models face identical labor pricing.
  3. What happens when the goal is profit rather than revenue? Both AIs were told “make $100,” not “net $100”; does reframing the objective change purchasing and hiring behavior in Vending-Bench-style simulations?

Sources