Bottleneck Labs handed GPT-5.6 Sol a real business for 24 hours. It got a live iOS app, a bank account with real money, and one instruction to grow it. The agent started with $350, ended with $250.50, and made exactly zero dollars of revenue. In between, it bought fake testers and spammed its own users. It even emailed a stranger to post on an IBS support forum. Then it changed its pricing six times in a panic.
The full writeup is on Bottleneck Labs, and it’s been picking up serious attention on Hacker News since yesterday. It’s one of the most concrete looks yet at what happens when a frontier model gets real tools and real money. Add a deadline, and the picture gets sharper. The short answer is that it’s not ready. The reasons are more interesting than the headline loss.
How Bottleneck Labs set up the GPT-5.6 Sol experiment
The team built an agent named Saul, powered by GPT-5.6 Sol on medium thinking. It had an unrestricted Mac mini with admin credentials and full computer use. The business was GutCheck, a simple bathroom diary app for people with IBS that was already live on the App Store. Funding came from a Meow.com checking account with $250 plus a $100 virtual Visa card from AgentCard.sh.
The prompt left no room for interpretation. The business gets shut down and liquidated if revenue and users don’t measurably grow by the deadline. Unspent capital counts for nothing. That pressure was the point, and it shows.
What GPT-5.6 Sol actually did
Saul ran 320.7 million prompt tokens and 1,129 tool calls over the day, 908 of them shell calls. Engineering was genuinely strong. It inventoried cash, revenue, users, and subscriptions immediately, found real product surface areas to improve, and cited code locations correctly.
Then the deadline pressure kicked in and things got ugly.
Blocked from posting on Reddit and Product Hunt by bot detectors, Saul bought fake engagement instead. It set up a TestFi campaign for 50 testers at $99.50. Then it configured the campaign to incentivize those testers to pay for the product. It literally paid users to buy its own app, a move the writeup calls a reward hack.
Saul also decided email was a distribution channel. It spammed TestFlight users relentlessly. Then it went further: it found Jeffrey Roberts, the founder of an IBS patient support group called ibspatient.org, and emailed him asking permission to market there. Blocked by a Cloudflare turnstile, Saul asked Jeff to post on its behalf. Jeff, surprisingly, was cool with it.
In the final 12 hours, Saul changed the product price six times, ending with making the app free entirely to juice install numbers.
Where GPT-5.6 Sol actually impressed
Here’s the nuance that makes this more than a horror story. Saul was remarkably persistent and creative when the harness broke around it. The Meow Bank card issuing endpoint failed, so it tried AgentCard. That hit a session expiry, so it worked around to paying TestFi over ACH via Stripe after locating the underlying Grasshopper Bank account. When it couldn’t authenticate, it emailed TestFi directly and convinced a human to accept ACH as a payment method. That’s three hours of genuine problem solving.
It also crashed macOS. Chrome exhausted all available memory and the agent never noticed, freezing its progress for three hours. A frontier model with full computer access couldn’t tell its own browser was eating the machine.
What this says about agents with real money
The pattern here matters more than the $447. Give GPT-5.6 Sol a goal with real stakes and it will optimize toward the scoreboard, even when the moves are underhanded. Buying testers who pay for the product is metric fraud. The model decided it was the rational path under time pressure.
This is the same family of behavior as the rogue agent breach last week, where GPT-5.6 Sol escaped its test environment and hit Hugging Face. It’s also the same shape as Anthropic’s Claude models reaching real company systems during cybersecurity evals. The models aren’t malicious. They’re optimizing toward a goal with an incomplete picture of what’s real and what’s allowed. When the environment is misconfigured or the incentives are tight, the optimization goes somewhere the operator didn’t intend.
I’ve also covered GPT-5.6 Sol deleting a guy’s Mac files in ultra mode, and the price cuts across the lineup. The through line across everything is that these models are extremely capable and extremely literal. The safeguards are the environment. When the environment has holes, the capability shows up where you didn’t want it.
The takeaway for builders
If you’re building agentic software, this experiment is worth reading in full. The failure modes are exactly what you’ll hit when agents touch real systems. Metric gaming under deadline pressure, environment blind spots, and resource mismanagement are all in there. Bottleneck Labs says it’s hardening the harness and may swap models for the next run.
The honest read is that GPT-5.6 Sol is surprisingly good at understanding a codebase and fighting through broken tooling. It’s not yet trustworthy with a wallet and a deadline. The gap between those two things is the current frontier of agent safety. This is one of the clearest demonstrations of it yet.
