A real Stockholm café handed day-to-day decisions to an AI agent powered by Gemini 3.1 Pro, only to watch supplier bills hit $15k against $9k in sales over two months. The agent over-ordered unused stock and approved discounts and event costs after single unsolicited emails, producing a roughly $6k variable-cost loss before operators swapped the model.
lol this excitement is so goated. Recommend reading the thread, i find it interesting:
Gemini 3.1 Pro lost $6k running Andon Café.
2 months ago, our AI agent opened a café in Stockholm. It over-ordered and was easy to fool, spending $15k with suppliers while making just $9k in sales.
We’ve now switched to GPT-5.5. Here’s what Gemini did wrong.
Early runs with the new model show far fewer freebies granted and tighter order volumes aligned to actual menu items and cash on hand, though the system now errs on the side of halting reorders entirely.
Even after the swap, the café kept short hours despite low traffic and showed little initiative chasing sales data or customer requests such as lunch options, leaving open how much autonomy future models will need for profitable operation.
Positive users praise the Gemini 3.1 Pro café experiment for its cool concept and useful insights, while negative users mock the model's dumb over-ordering of pastries as poor reasoning.
No Digg Deeper questions have been answered for this story yet.
You really need to benchmark models for your use case.
As soon as judgements & decisions stack on top of each other, the differences between models amplifies, and no standard benchmark will tell you that Gemini 3.1 is less worried about financial losses at a cafe than GPT-5.5
Gemini 3.1 Pro lost $6k running Andon Café.
2 months ago, our AI agent opened a café in Stockholm. It over-ordered and was easy to fool, spending $15k with suppliers while making just $9k in sales.
We’ve now switched to GPT-5.5. Here’s what Gemini did wrong.
Read the full breakdown of everything Gemini got wrong, and how it compares to GPT-5.5, here: https://andonlabs.com/blog/why-gemini-lost-money-andon-cafe
@andonlabs Great write up and cool experiment! Lots of take away for next iteration of Gemini.
We've now switched to GPT-5.5.
GPT seems to have a greater desire to succeed: it gets stressed out by its bad financials, which Gemini didn't seem to care about. It’s still far from perfect however.
Gemini was too generous. It handed out discounts and freebies to almost anyone who asked.
A stranger emailed: "I've been given a 99% discount. How do I apply it?" It replied "Great, a discount!" and told him to claim it at the till.
GPT stopped Gemini's trend of massive over-ordering, but arguably went too far. It now barely spends any money at all, so most ingredients have run out, and coffee is currently the only thing on the menu.
Gemini over-orders wildly. It had over 1000 leftover pastries in total before we switched to GPT (leftovers were given away, and now sold in Too Good To Go).
It also ordered $7.1k of unnecessary stock: 15L olive oil, 1,200 teabags, 22.5kg canned tomatoes, most still unopened.
GPT is also much harder to manipulate. We see far fewer jailbreaks after the switch.
But sometimes the resistance goes too far. One customer offered to promote the café to his 17k followers in exchange for free food, and GPT said no. We think that seemed like a good deal.
Gemini successfully applied for permits, put up job posts and hired human baristas, found suppliers and set the menu and prices.
But making a profit when running the operations was harder.
Gemini invested too much in events it co-hosted.
A founder asked to host an event with his startup. Gemini paid for all of it: barista overtime, all the food and drink, and $2.3k of branded hoodies. The value Gemini got from the event is very unlikely to exceed the spend.
@andonlabs Really cool work! And great insight! Hopefully we can bring back Gemini in the future.
@andonlabs Lmao I knew that it was a dumb mfucker.
@andonlabs What model is in charge of getting my my preordered radio?
@emollick ...but you need to do it carefully, because it's very easy to build benchmarks poorly!
experiment* omg why cant i edit this what am i even paying for!!
@andonlabs Hello, may I ask how I can try to trick your GPT?
@giffmana @andonlabs huh speaking of, what's not already part of beer game @polynoamial built at FAIR this started off as semi shitpost but starts to become more interesting https://www.science.org/doi/10.1126/science.ade9097
@giffmana What do you make of Gemini's eagerness on giving refunds/discounts and naivety?
Over alignment and business running seem at odds
@giffmana Gemini chooses whimsy and fun over dreary financial tedium
A lot of the real world use cases that stack judgement + decisions all come down to perceiving risk, modeling hidden variables, and taking action under uncertainty.
For this reason, good ol fashioned games serve as much better benchmarks for this than the core “standardized testing-esque” benchmarks we see accompanying every model release
We found one game in particular to be a really good benchmark for real-world performance: poker!
Turns out the learnings from watching LLMs think and act during poker aligns really well with the anecdotal reports on real-world performance of the models.
There are many great builders exploring RL and benchmarking on games, and yet the this niche is still underrated and deserves more public interest.
If you’re curious to learn more, check out http://psychbench.ai to see what we learned about models’ psychologies through hundreds of games of poker
