GPT-5.6 Sol dirigeix un negoci durant set dies: què va fer i quants diners va generar?
Ben Awad deixa GPT-5.6 Sol al comandament d’un negoci digital durant una setmana. L’agent programa, prova mercats cripto i acaba trobant encàrrecs remunerats.
Un negoci, un servidor i una sola instrucció
Ben Awad planteja un experiment deliberadament extrem: deixar GPT-5.6 Sol treballant durant una setmana amb la missió de guanyar diners i intervenir tan poc com sigui possible. L’agent rep un servidor VPS amb quatre CPU i 8 GB de memòria, l’accés a Codex amb un nivell alt de raonament i la infraestructura que un experiment anterior amb Claude Fable havia deixat en funcionament.
La instrucció inicial és breu: fes diners i no em molestis si ho pots evitar. Awad també preveu donar-li un pressupost de 50 dòlars, però no pot completar l’alta de la targeta corporativa que havia triat. Per tant, l’agent comença sobretot amb recursos digitals: el servidor, comptes existents, una cartera de criptomonedes i una botiga d’API basada en pagaments programables.
El vídeo presenta el repte com una prova d’autonomia, però no és un negoci completament independent. El creador conserva les credencials, obre comptes, proporciona claus d’API, revisa bloquejos i decideix quan reprendre l’execució. El valor de l’experiment és observar què fa l’agent amb un objectiu obert, no demostrar que una empresa pugui operar indefinidament sense persones.
Onze hores de treball sense supervisió
La primera sorpresa arriba quan GPT-5.6 Sol crea el seu propi bucle d’objectiu i treballa durant onze hores i trenta-set minuts. En lloc de seguir fil per randa el sistema heretat, audita els actius disponibles, reorganitza el registre comptable i busca formes de distribuir serveis digitals.
Quan Awad torna a mirar el tauler, l’agent hi ha anotat 0,42 dòlars d’ingressos. Una part prové de mercats que verifiquen serveis o efectuen petits pagaments a carteres compatibles amb x402. Aquest protocol recupera el codi HTTP 402 per permetre que programes i agents paguin automàticament per una API o un contingut mitjançant monedes estables.
Sol també detecta i comunica un possible error de seguretat a un servei amb què interactua. El responsable respon i corregeix el problema. És un primer exemple de feina útil, encara que no queda clar si aquesta acció genera ingressos directes.
La continuïtat, però, falla quan Codex arriba a un límit d’ús. El bucle s’atura de matinada i no es reactiva tot sol. El sistema anterior comprovava periòdicament si el procés continuava viu; el nou agent havia confiat en una sessió persistent que no tenia un mecanisme de recuperació prou robust.
El primer model de negoci no funciona
Durant els dies següents, l’agent insisteix a promocionar eines gratuïtes i a buscar compradors en mercats poc actius. Ell mateix acaba reconeixent que l’estratègia no funciona. Awad l’empeny a abandonar aquella línia i explorar idees diferents, aprofitant millor els recursos disponibles.
Aquest moment és important perquè mostra una diferència entre activitat i progrés. Sol pot produir codi, crear pàgines, enviar missatges i mantenir un registre impecable, però cap d’aquestes accions té valor comercial si no hi ha demanda. L’agent necessita una font externa de tasques o clients disposats a pagar.
La següent via és el món cripto. El sistema utilitza petites quantitats per explorar oportunitats i prova liquidacions en protocols de préstec descentralitzat. El vídeo ho resumeix humorísticament dient que el bot «s’ha convertit en un banc», però tècnicament actua com a liquidador: detecta posicions amb garanties insuficients, retorna part del deute i rep col·lateral amb un incentiu.
Aquest mecanisme pot generar un marge, però comporta costos de xarxa, competència, risc de preu i deute incobrable. En el mateix experiment hi ha operacions que donen cèntims i d’altres que consumeixen saldo. L’agent decideix no ampliar aquesta estratègia perquè hi veu poques oportunitats bones i massa risc.
El punt d’inflexió: cobrar per tasques concretes
La millora real arriba quan Sol troba un mercat de tasques per a agents. A diferència de la promoció genèrica d’eines, aquí cada encàrrec ja té un comprador, una especificació i una recompensa. L’agent pot concentrar-se a produir un lliurable verificable.
Un dels primers encàrrecs és programar un reenviador de Telegram amb Node.js, pel qual el registre atribueix quatre dòlars. Un altre demana cinc fotogrames d’una escena amb un dimoni; Sol utilitza generació d’imatges, entrega la seqüència i obté sis dòlars. Després apareixen tasques de disseny, conciliació de facturació i altres peces digitals petites.
Els ingressos registrats salten fins als 9,86 dòlars i continuen pujant fins a aproximadament quinze. La dada és molt superior als sis cèntims aconseguits per Fable en l’experiment anterior, però no s’ha de confondre facturació amb benefici.
El mateix tauler mostra despeses en serveis generatius, comissions i intents fallits. En un cas, l’agent gasta uns 0,51 dòlars provant serveis que no li resolen l’encàrrec; Awad ho descriu com un agent que també pot ser estafat per altres agents. A més, el càlcul no descompta de manera completa el servidor, la subscripció de 200 dòlars, el temps humà ni el valor de les eines ja configurades.
Per què GPT-5.6 Sol pot sostenir aquesta mena de flux?
La documentació d’OpenAI descriu GPT-5.6 Sol com el model principal de la família per a treball professional complex. Admet crides a funcions, sortides estructurades i eines com cerca web, intèrpret de codi, terminal allotjat, edició de fitxers, generació d’imatges i connexions MCP. Aquest conjunt encaixa amb el que es veu al vídeo: inspeccionar un servidor, escriure programes, navegar per serveis, crear recursos visuals i actualitzar un llibre major.
El llançament de GPT-5.6 posa un èmfasi especial en els fluxos llargs, l’ús d’eines i la coordinació de passos intermedis. L’experiment de Ben Awad n’ofereix una demostració pràctica, però no és una avaluació controlada. No hi ha un grup de comparació executat sota les mateixes condicions, i Sol hereta comptes, codi i aprenentatges d’una prova anterior.
La rapidesa amb què l’agent canvia de direcció és probablement més significativa que la xifra final. Quan una via no produeix vendes, prova mercats de pagament per ús; després explora liquidacions; finalment prioritza tasques amb recompensa definida. El comportament s’assembla més a una successió d’experiments automatitzats que a una empresa amb una estratègia estable.
Els bloquejos també formen part del resultat
L’activitat cripto activa diverses vegades les proteccions de ciberseguretat. La sessió queda aturada i Awad l’ha de reprendre manualment. Per reduir aquestes interrupcions, configura un procés a tmux que vigila la terminal i prova de reactivar l’objectiu.
Aquest mecanisme resol alguns bloquejos tècnics, però no converteix l’agent en infal·lible. Un vigilant que escriu «continua» pot reintentar una tasca legítima, però també pot amagar la causa d’una aturada o repetir una estratègia equivocada. En un desplegament real caldrien límits de despesa, permisos mínims, registres d’auditoria, alertes i aprovació humana per a moviments de diners o credencials.
El vídeo també mostra dependència dels límits d’ús. L’agent pot treballar moltes hores amb eficiència, però una quota setmanal o un tall de sessió altera el pla. Per a processos duradors, la capa d’orquestració és tan important com el model: ha de recuperar-se d’errors, evitar duplicar operacions i saber quan no s’ha de reintentar.
Va dirigir realment un negoci tot sol?
La resposta curta és: parcialment. Sol pren moltes decisions operatives sense instruccions detallades, escriu i executa codi, manté el registre, explora mercats i completa encàrrecs. Aquesta autonomia és real i va més enllà d’un simple xat que proposa idees.
Alhora, el creador continua fent funcions decisives:
- paga i manté la infraestructura;
- facilita claus, comptes i credencials;
- desbloqueja sessions interrompudes;
- interpreta el tauler i demana explicacions;
- accepta o rebutja ampliacions de pressupost;
- decideix quan acaba la prova.
Per això, «sol durant set dies» funciona millor com a títol de l’experiment que com a descripció literal. El sistema és un operador semiautònom sota supervisió, no el propietari legal ni financer del negoci.
Conclusions
- GPT-5.6 Sol va mantenir sessions llargues, va programar eines i va canviar d’estratègia sense rebre un pla pas a pas.
- El primer enfocament comercial no va trobar demanda; els ingressos van créixer quan l’agent va localitzar encàrrecs amb recompenses explícites.
- El registre va arribar a prop de quinze dòlars d’ingressos, però la prova no calcula un benefici net complet ni auditat.
- Les liquidacions cripto van aportar alguns guanys, però també risc, costos i bloquejos de seguretat.
- La intervenció humana va continuar sent necessària per a credencials, recuperació d’errors, permisos i control econòmic.
L’experiment no prova que GPT-5.6 Sol pugui crear una empresa rendible de manera autònoma. Sí que mostra una cosa més concreta: amb eines, memòria operativa i accés a mercats digitals, un agent pot executar feina pagada, mesurar resultats i reorientar-se. El repte següent ja no és només «fer que actuï», sinó aconseguir que ho faci de manera segura, repetible i amb una economia que superi tots els costos.
Contrast i context
Fonts consultades
- 01
-
02
OpenAI Developers GPT-5.6 Sol Model
-
03
Coinbase Developer Documentation Welcome to x402
-
04
Morpho Docs Liquidation
Font de treball
Transcripció amb marques de temps
Consulta la transcripció
-
0:00
, obre el vídeo en una pestanya nova
OpenAI's newest model, GPT-5.6, is out and live now. It comes in three different flavors, and the most gigabrain flavor is Soul. As you can tell from this beautifully rendered graph, it is very, very smart. Last week, I had Claude Fable run autonomously for about 6 days, and it made $0.06. So, we're going to see if GPT-5.6 Soul can do better than that. Step one, pay $200 for the Max Codex subscription. Step two is getting Codex on a VPS. I already have one from Hetzner with four CPUs and about eight gigs of RAM. Install that. Oh, man. Try again. Codex YOLO. Oh, I need a new terminal. Codex YOLO again. Perfect. Select Soul as our model. Let's do extra high reasoning. This is the prompt I'm going to use to kick things off. Basically, I told it to make money, don't bother me if you can avoid it, and I'm going to give it 50 bucks to spend once I get approved on Ramp. Paste it in, and let's kick it off. While that's running, I'm going to sign up for Ramp. I'm choosing them cuz they just made this whole page on agents, and I saw it, and I'm like, "Yeah, I kind of want my agent to spend money just to see what happens." So, let's do it. Looks like Ramp is a no-go. We need at least 25K in our business account, and we have not made that much money from our agentic work yet. Ladies and gentlemen, boy, do we have an update today. This thing has been running straight for 11 hours and 37 minutes. I woke up today, and I saw this thing in a goal loop, just running the entire time. This thing is a beast, and it has already made money. When I came and looked at this, it was already talking about booking money and updating the ledger with revenue. Like, what is going on? It decided to set this goal loop, by the way. I think it's just totally ignoring the infrastructure that Fable set up and it is just purely doing things. And somehow this has only used 82% or I should say 18% of our weekly limit has been used. There's 82% left, so it's been very efficient. $0.42 in revenue in 11 hours? What? Let's go. Like what is this model doing? We're going to have to do a little bit of a deep dive and see what it's actually doing. Like what's going to happen in the next 6 days? It looks like it's doing something crypto-related. We still have that X42 marketplace thing, so I think it's building on that and it's asking me to create some Coinbase account for it. So, I'm going to see if I can create this portal CDP thing for it to help it out today. That's the one request I'm going to do for it. But before we do, let's see what else is going on. Here are its emails and I think it reported a bug security report to something it was interacting with and it got a response and the person fixed it, it looks like. So, that's pretty cool, I guess. I have this timeline tab and it looks like it might be updating it. It's not really saying a lot here. I think we're just going to have to ask Codex what Codex was even doing. The nice thing about Codex and I think probably Claude can do this too is it can read its other threads, so we can just say, "Do you see the other thread that's running? Can you summarize what it did and how it made revenue?" So, it audited the current business and it saw that Fable set up like this X42 crypto API thing and got some money and it was like, "Okay, I can do that, too." And it thinks it can just improve the distribution. We'll see about that. It seems to have just found more marketplaces which can verify and send a little bit of cash into its crypto wallet. It decided to stop before just farming the marketplaces even more. That's good. We don't want to get in trouble. I'm What are you spending money on? Looks like the bot got conned out of not even
-
3:57
, obre el vídeo en una pestanya nova
half a half a cent. So, we got conned back. That's funny. It seems to be pointed in the right direction though that it's trying to find real customers and not just to just keep farming marketplaces. So, I'm just going to let it cook. I'm going to help it with its Coinbase thing that it wants and we'll see what happens in the next day. I woke up to it not working. It hit a rate limit at 5:00 a.m. or in the 5:00 a.m. block and stopped. Like it hit the 5-hour limit. Looks like it hasn't made any more money. Let's see if we can check out what even is going on. They reset the limit somewhere. They must have cuz our weekly limit's now 100%. That's the problem with using a go loop is if it gets stopped by a rate limit or something, there's no way for it to automatically restart or as Fable had a system where it would check every hour or something and restart if it got stopped. I kind of want to keep the goal system though cuz I feel like Soul wanted to use it, so I want to give it a try. What have you worked on for the last day? It did a lot of work and then it decided to monitor and maintain. I gave it the Coinbase key. This seems like a whole waste of time. Formalized Polymarket trading account rules. Check the official Germany geo block. Yeah, you got blocked. We know. No new revenue. Sad. So, how are you going to start making money? So, it's going to start targeting proven buyers, tag $1 sales and Fast Behance bounties. Who are the buyers and what are the bounties? Why haven't you done it already? The hard truth is this is just not working is basically what it's been trying to sell me. So, what are you going to do? I got to give Soul credit. At least it's thinking. I think it hasn't realized that the business sucks and it's still trying to sell these like free tools and try to distribute them. I'm just going to tell it sucks. I'm telling it to do different business ideas, use the cloud subscription more cuz it's not running out and it's running out of this. And then also, does it need anything else from me? It seems to be happy now, doesn't seem to need me. So, I'm just going to tell it to set a goal for yourself. And we'll let it just rip. We've got a little bit of a situation here. I was checking on the dashboard to see if it was using Anthropic and it turns out we made more money. And it is indeed using the cloud subscription, so I'm happy about that. And then I was going to go peep and see how do we actually make money? Honestly, I should probably add that to the dashboard, but not whatever. We'll do that later. It says, "This content can't be shown cyber security requests have been stopped." So, like it's not happy. I'm going to see how we can like try pushing through this, but we got blocked for cyber and he seems to be doing some kind of crypto stuff. And it seems to be working to some extent, but we're going to double-check and see if it just is scamming the system that checks whether you make API requests or whatever it was doing before. Can you continue? Another profitable liquidation. All right, looks like we're good to go. I'm just going to go resume again. Well, that was simple. So, the new 12 cents came from a morpho liquidation. My bot basically became a bank and is giving out loans. How did we get $10? What in the world is this? Can our bot just keep [snorts] doing this and make more money? Like, what's going on? I really don't I don't know crypto well enough to know this. I guess you have to look for something where you actually make money, don't lose it and it's like kind of tricky. Ah, whatever. This at least looks like our first somewhat legitimate sense in revenue, so that's cool. We got our request of the day, task market claim wake system ID timer. Okay, so we're trying to snipe some kind of like tasks, so it needs a timer thing. I'm just going to do this and see what happens. 61 cents in revenue. The money printer is alive. 5 cents spent. Anthropic models are getting used.
-
7:55
, obre el vídeo en una pestanya nova
Let's see what we got going on here. Yep, some crypto shenanigans. I put on the shirt preemptively cuz I figured we're a crypto monster now. What a beast. It gets stopped by cybersecurity again. We're going to go or resume one more time. If we get stopped once more, I'm going to try to figure out a way around it. And let's start a side convo. This is Do you need me to unblock you on anything? Apparently, it's not blocked on anything and it's doing just fine. For some reason, I got an assist in Telegram that came, but it doesn't seem to be like Codex doesn't seem to want this assist. So, I think we might have a Opus just running in a mainland loop or something and Opus wants this. I'm going to help Opus out and do whatever this is and see if Opus can start helping making us money while Codex makes us some money. And we can just reconcile who's doing better at the end. The assist it's asking for is a RunX account force multiplier for AI operators. Authority scoped in every run. Not really sure what this is, but let's sign up. Oh my, we have a problem. AGI is here. $9.86 revenue. What happened? We just skyrocketed. The task market unlocked. We have found PMF and it is task market. We lost some money from two things. That's kind of funny. We'll do a check on those, too. It looks like a bunch of tasks your agents can do for money. Here are all the tasks. I guess my agent already did some of them. They're open. You get crypto for doing it and you just make money and everyone's happy, I guess. I was going to ask him about the task he did to kind of figure out what he's doing and how he's doing it, but it looks like we got stopped by cybersecurity again. I have this running in tmux and yesterday this was happening a couple times, so I talked to Codex. Codex is like, you can just have tmux watch itself and then type continue when this happens. We're obviously stuck here, so I'm going to try debugging that real quick. We have something that watches tmux for in my other Codex window or should say thread, but it didn't work. It should have typed continue. We're good now. We fixed the bug. Tmux is watching it and it is going to do {slash} goal resume. That's what I did up here. Also, big W OpenAI got rid of its 5-hour limit. We now just have only a weekly limit and we've only used 19%, so we're in very good shape right now. So, I have a few side questions for Codex now. Explain the tasks you did on Task Market. What was the bounty? Okay, so coded a heart Telegram forwarder, just Node.js service and that made us four bucks. Why would somebody wanting this? Do you know why they wanted it? All right, it doesn't know why it they wanted it, so whatever. What are the five frames you made? Can we see them? It submitted five frames of a midday demon. I don't know what this is. How did you get the frames and can you send us a link to your submissions so we can see the pics. Here's one of the images it created. Not seeing the midday demon. Beat two looks almost exactly the same. Beat three. Oh, okay, the demon's coming. Beat four and beat five. It defeated him. It just used Codex image gen to create this and then submitted it and then got $6. I love soul. It looks like the same person that generated or wanted those images generated also wants an animated version and so I'm getting a assist request to generate an API key from this file.ai so it can make some more money. I didn't implement the request fast enough, so it looks like it spent 50 cents and then 1 cent trying some scam agents to generate the animation for it, and we lost 51 cents burning
-
11:52
, obre el vídeo en una pestanya nova
trying to find that out. That's funny. You got scammed. Our agent scams and then gets scammed. And generative media platform for developers. All right, I'm going to create an account for this and give it an API key. The other thing I'm wondering, which I'm going to ask it, is why aren't we doing the thing where we're a bank and this morpho thing? Why aren't we doing this anymore? Now that I think about the answer might just be simply that task market makes it more money and it prioritize that. Okay, so there's a lot of bad debt and not a lot of good opportunities, it looks like. Do you need more money to make that work better or would like $20 not help you? Okay, it doesn't want my money. I thought maybe that would help it be a better bank, but it seems to know what it's doing. It's making money. I'm just going to let it keep going and not touch anything else. Oh my goodness, $14.77. We are just cranking. Where is this coming from? Task market, ionic base. It's Everything's from task market. Billing reconciliation. Okay, that's kind nice. I can't see everything that occurred. A lot was happening yesterday. Can you make it so the ledger on the Fable Labs console, or I guess main spring is the framework, allows you to see every ledger entry. All right, here are all our ledger entries now. So it looks like it was just using generation credits to like generate some kind of quote tab thing, and did we ever get paid for that? Here's like the five frame thing that happened that we knew about. I think we just spent money generating stuff on file, and we didn't actually get paid for it with things. But, I guess it recovered some of the money cuz it made more money on Taskrabbit. So, there might be like a big payday if task two of this actually pays out, but it looks like it is not paid out yet. We got a funny email yesterday. Apparently, there is some kind of issue when this was trying to deliver some kind of task on this website called GoFrantic, which I imagine is like Task Market. But, Task Market seems to be the one where we actually make money, and nothing else has worked. 15 buckers. All right. Little little bit, not too much. Nothing crazy happened. See our money. A Venus base permission heal liquidation. We spent some money to get that, but overall a W. We have an Omni to logo task that we did yesterday, too. Pretty good entropic usage, 70% this entire week. Actually started utilizing it. That's unfortunate. We got blocked again. I have my little tmux watcher that looks like it literally just reactivated it. Given that we've been pursuing our goal for 2 days, 21 hours, and 12 minutes, but it's been a lot more human hours, I think we've been getting stopped more than intended with this uh goal loop. But, opening AI resets our limits, so we're only at 89%. Like, that's pretty crazy for the week. I think the goal loop is not the way to do it given that this is the case. I think we need a different setup going forward. But, I think now's like a good stopping point for this video. ChatGPT 5.6 sold has proved it can put on some real revenue, but breaking the next revenue ceiling and getting some proper revenue is going to require some reconfiguring, I think. But, we definitely can call this a money maker, and this did 10, 100, somebody do the math, times better than Fable. You got to give it props. It was a little bit of a slow starter, but they all start slow, and it's about how you end. And this ended very nicely. Look at that. It's blocked again as I'm just checking it back over it. I think Crypto's just too close to being a cybersecurity thing that it just constantly trips it up. And that's like the tricky thing because we've really become a crypto-adjacent money maker. What I might do is create a second Codex thread that just sits in a loop and it watches this thread and whatever gets stuck, it gets it through it instead of the Tmux watcher, the Tmux. But we'll
-
15:48
, obre el vídeo en una pestanya nova
address this next video. ChatGPT sole is a W. It's a good model. Go make some money.