Quan Claude arriba al límit: canviar a Kimi K2.7 sense aturar un projecte
Dev Knives mostra com continua una aplicació real després d’esgotar el límit de Claude: trasllada la feina a Kimi K2.7 Code mitjançant OpenCode, deixa l’agent treballant cinc hores i compara el cost. La lliçó útil no és quin model «guanya», sinó com repartir tasques, context i pressupost.
Estàs programant, el model ja coneix el projecte i, just quan la feina comença a avançar, apareix el missatge de límit assolit. Aquest és el problema que Dev Knives converteix en una prova pràctica: en comptes d’esperar que es renovi la quota de Claude o contractar immediatament un pla superior, connecta Kimi K2.7 Code a OpenCode i continua la mateixa funcionalitat amb un altre model.
El resultat que ensenya és atractiu: un agent treballant durant cinc hores i vuit minuts, una aplicació ampliada i una despesa declarada de 23,45 dòlars. Però el valor del vídeo no és demostrar que Kimi sempre sigui millor o més barat que Claude. El valor és una idea de treball més general: un model és una eina dins d’un sistema i no una dependència exclusiva.
El contingut està patrocinat per Kimi i la descripció inclou un enllaç d’afiliació amb crèdit promocional. Això no invalida la demostració, però sí que obliga a interpretar les comparacions de preu com l’experiència d’una sessió concreta, no com un estudi independent.
El projecte no és una prova de joguina
La demostració parteix d’una aplicació que el creador utilitza per controlar vídeos de YouTube. Ja disposa de base de dades i contenidors, però vol afegir una interfície amb historial, cerca i un procés automatitzat.
El flux desitjat és prou complet:
- L’usuari enganxa l’enllaç d’un vídeo allotjat a Google Drive.
- El sistema descarrega el fitxer.
- Una IA genera títol, descripció, marques de temps i categoria.
- L’aplicació puja el vídeo a YouTube com a no llistat.
- La interfície conserva l’estat i l’historial de cada operació.
Aquest tipus de projecte consumeix molt més context que una pregunta aïllada. L’agent ha d’entendre l’estructura del repositori, les dependències, els models de dades, les variables d’entorn i les decisions anteriors. També ha de llegir i modificar molts fitxers. Per això els límits d’ús poden arribar abans del que suggereix la durada visible de la conversa.
Per què Claude esgota la quota
Dev Knives explica que va utilitzar un model potent de Claude per analitzar el projecte i preparar el pla. Quan encara no havia començat la major part de la implementació, segons el seu indicador ja havia consumit aproximadament un 70% de la quota disponible.
La causa no és només la resposta que veiem. En una sessió de programació, el proveïdor pot processar repetidament instruccions, historial, fitxers, resultats d’eines i fragments de codi. Com més gran és el context i més llarga és l’execució, més tokens d’entrada s’acumulen. La memòria cau pot reduir el preu d’algunes lectures repetides, però no elimina el consum ni garanteix que tota la informació quedi emmagatzemada al mateix preu.
Això explica una situació aparentment estranya: una conversa curta pot ser barata, mentre que un agent que explora un repositori durant hores pot consumir una quantitat considerable encara que el nombre de missatges humans sigui petit.
La proposta: separar planificació i execució
El creador defensa reservar el model més car per a tasques on el seu criteri aporta més valor: entendre requisits ambigus, dissenyar l’arquitectura, detectar riscos o elaborar un pla. Després trasllada l’execució més sistemàtica a un model més econòmic.
No és una regla universal. Un model més barat pot introduir errors que facin perdre més temps del que s’estalvia, i un canvi de proveïdor pot perdre context implícit. Però la idea de rutar la feina segons la dificultat és més robusta que intentar resoldre-ho tot amb el model de preu més alt.
Una divisió raonable podria ser:
- model principal per aclarir l’objectiu, revisar l’arquitectura i validar canvis delicats;
- model alternatiu per implementar tasques ben delimitades, escriure proves i fer refactoritzacions repetitives;
- eines deterministes per formatar, executar tests, analitzar tipus i comprovar seguretat;
- revisió humana abans de desplegar o tocar dades reals.
La qualitat final depèn més d’aquest circuit que del nom d’un únic model.
Com connecta Kimi a OpenCode
OpenCode és un agent de programació de codi obert que pot treballar amb diversos proveïdors. La seva documentació explica que es poden afegir credencials amb l’ordre de connexió, seleccionar models disponibles i configurar proveïdors compatibles amb l’API d’OpenAI quan no hi ha una integració específica.
Al vídeo, Dev Knives introdueix la clau de Kimi i selecciona K2.7 Code. Utilitza la versió estàndard, no la modalitat d’alta velocitat. També activa una configuració d’orquestració perquè l’agent divideixi la feina i utilitzi subagents.
Aquí hi ha una distinció important: la subscripció de Kimi Code i el consum de l’API són productes diferents. Una quota mensual dins del servei no implica necessàriament crèdit il·limitat a l’API. Abans de copiar el procés cal comprovar quin mètode d’autenticació s’utilitza, quins límits té i com es factura.
Segons la pàgina de preus consultada el juliol del 2026, K2.7 Code té tarifes diferents per entrada en memòria cau, entrada sense memòria cau i sortida. La modalitat d’alta velocitat costa més a canvi de generar més ràpid. Aquest esquema fa que el cost real depengui molt del volum de context, la reutilització de la memòria cau i la quantitat de codi que l’agent produeix.
Cinc hores de feina autònoma
Després de donar instruccions, l’agent formula preguntes per concretar requisits i treballa durant cinc hores i vuit minuts. El vídeo mostra que crea la interfície, modifica la base de dades i la configuració, implementa el procés de descàrrega i generació de metadades, i acaba pujant un vídeo com a no llistat.
En un moment necessita que l’usuari li digui que continuï. També rep una captura de pantalla per corregir un desajust visual. Aquest detall és interessant perquè combina codi i visió: l’agent no només llegeix fitxers, sinó que compara el resultat renderitzat amb una imatge.
La demostració evidencia capacitat operativa, però no documenta una bateria completa de proves. Tampoc veiem una auditoria de seguretat, una revisió de migracions, comprovacions de concurrència ni el comportament davant d’errors de Drive o YouTube. Que el flux principal funcioni una vegada és un bon prototip; no equival encara a una funcionalitat preparada per a producció.
Què significa realment el cost de 23,45 dòlars
La sessió acaba amb un consum declarat de 23,45 dòlars. Dev Knives calcula que una execució equivalent amb un model d’Anthropic podria haver costat uns 76 dòlars si s’aprofitava la memòria cau o més de 500 sense aquest descompte.
Aquestes xifres alternatives són una estimació del creador. No hi ha dues execucions idèntiques, amb el mateix prompt, les mateixes eines, les mateixes respostes i el mateix resultat final. Un model pot necessitar menys tokens, un altre pot resoldre la tasca en menys passos i un tercer pot generar més errors. Per tant, no és una comparació controlada.
Per mesurar bé el cost cal registrar almenys:
- tokens d’entrada nous i tokens recuperats de memòria cau;
- tokens de sortida i crides a eines;
- temps total i nombre d’intervencions humanes;
- proves superades i errors introduïts;
- cost de revisar, corregir i tornar a executar;
- qualitat del resultat respecte d’uns criteris definits abans de començar.
La factura de l’API és només una part del cost d’enginyeria. Una sessió barata que requereix quatre hores de correcció pot ser més cara que una sessió amb més consum i un resultat fiable.
K2.7 ja no és l’última versió
El context temporal importa. La documentació de Kimi situa el llançament de K2.7 Code el 12 de juny del 2026. El 16 de juliol, el mateix dia de publicació del vídeo, la companyia va anunciar Kimi K3 i el va presentar com el seu model més capaç.
Això no fa inútil la prova. El procediment de treball —canviar de proveïdor, limitar l’abast, mesurar el consum i validar el resultat— continua vigent. Però no convé interpretar el vídeo com una comparativa actual entre els models més nous. Els catàlegs, les tarifes i els identificadors canvien ràpidament.
També és una advertència per a qualsevol equip: la configuració del model ha d’estar centralitzada i ser fàcil de substituir. Si el nom queda escampat pel codi, cada canvi de versió es converteix en una migració.
Una arquitectura pràctica per no quedar bloquejat
Treballar amb més d’un model no significa saltar improvisadament entre serveis. Un sistema saludable conserva un registre de decisions que permet reprendre la tasca:
- Documentar l’objectiu i els criteris d’acceptació en un fitxer del projecte.
- Fer que cada tasca sigui petita i tingui proves concretes.
- Guardar els canvis en commits curts i reversibles.
- No exposar secrets ni dades personals al context.
- Executar tests, anàlisi estàtica i format després de cada bloc.
- Fer revisar els canvis crítics per una persona o un segon model.
- Establir un límit de diners i de temps abans de l’execució.
Així, si una quota s’acaba, el model següent no necessita reconstruir tota la història des de zero. Pot llegir l’estat del projecte, el pla i els resultats de les proves.
La lliçó que queda
El vídeo presenta Kimi K2.7 com una sortida concreta quan Claude arriba al límit. La lliçó duradora és més àmplia: no s’ha de confondre la interfície de conversa amb el procés d’enginyeria.
Un bon flux assigna cada fase a l’eina adequada, controla el pressupost, conserva el context fora del xat i verifica de manera independent allò que l’agent afirma haver acabat. Els models canvien, els preus canvien i les promocions caduquen. Les proves, els criteris d’acceptació i la capacitat de substituir el proveïdor continuen sent el que evita que el projecte quedi aturat.
Contrast i context
Fonts consultades
-
01
Dev Knives Qué hago cuando Claude se queda sin límites
- 02
- 03
-
04
Kimi Kimi API overview
-
05
OpenCode Providers
Font de treball
Transcripció amb marques de temps
Consulta la transcripció
-
0:00
, obre el vídeo en una pestanya nova
The truth is, I was on a roll. It was one of those afternoons where everything flows, the code comes out and things progress phenomenally. But just when I ask the model to perform the following function, this happens . Usage limit reached. Come back in a little while. Ale, you've already annoyed me. And I'm sure this sounds familiar. You're deep into something, researching, using Claude, and suddenly your subscription expires. Cloude leaves you stranded at the best part; you're left there waiting for 5 hours or maybe a week for your available credits to be renewed . Either that, or you pay for the most expensive plan, which costs a fortune. Well, now I'm going to show you the third option, and it's something I actually do, and pay attention because this isn't a video about how to save on your Cloue subscription. No, no, no. This is much more useful [music]. This is what I do in my actual workflow, okay? I don't stop working or pay for more expensive subscriptions, [music] believe me, I have n't spent that much money on subscriptions. I simply change my model and keep working, that's it, I don't do anything else. You're going to see a complete flow of what I do when working with a feature of an application that I 'm making for myself. And at the end of the video we'll see how much I've spent, okay? Without mincing words. I don't know, I have n't recorded it yet, so I'll show you what I spend, but you'll see it's much less than with Cllaude for sure. And well, to give you some context for the video, I have to tell you and warn you, because I like to be honest, that this [music] video is a collaboration with Kimi, which is the artificial intelligence model that I'm going to use to get out of this jam and this mess of always using expensive models. And listen up, because it's a really good model and I'm quite proud and happy to be sponsored, since I've been using Kimi for a long time, and I'm totally serious about this, I pay for
-
1:42
, obre el vídeo en una pestanya nova
their cheapest plan, okay? I can't afford more, but I pay for it [music]. Furthermore, I'm going to introduce you to their new model called Kimi 2.7 Code, which has a ridiculously low price for the capabilities it offers. [music] And believe me when I say that it's on par with the top models, the cutting-edge models that are said to be the best in the market. I'm going to tell you what it does well, what it does okay, and even if it does something wrong, I'll tell you that too. And the numbers, as I've already said, are going to be mine. Let's start with a little bit of zero spending and see how much we spend doing everything I want to do, and you can judge for yourselves whether [music] spends a lot or a little. And if something disappoints me, I'm obviously going to tell you. Don't worry, we're going to be upfront and honest here . So if you're interested in this stuff and topics related to programming with artificial intelligence and technology, first, subscribe because I'm sure you're not. And second, stay tuned to the video as we begin this test. So, as you can see, guys, I'm going to explain the problem from scratch so you can understand it. Here's a real project of mine that I'm using right now . It's called YouTube Channel W. And it already has its scripts and stuff, okay? He already has quite a few things here. It's a system I set up to basically check every x amount of time for all the changes in my channel, okay? Imagine, it checks and saves in a local database the videos I've uploaded, the playlists I have, what descriptions I have, what channel settings I have, what channel name I have, okay? All the channel information that can be extracted via the YouTube API is saved, so that when there is a change I automatically receive a notification on my mobile and also in my email with the change that has occurred to inform me. I do this basically because I have schizophrenia. Well, no, I don't have schizophrenia, but I like to
-
3:24
, obre el vídeo en una pestanya nova
be notified of all the changes on my channel, okay? I program things and I like seeing that the programming has worked, okay? It's not anxiety, it's just that I like to be notified and also in case something ever happens, okay? If these channels get hacked and something happens to them, I'd find out instantly that there's been a change, okay? I simply like having it. Then when I make changes myself, it notifies me, but I'm the one who knows. So, having said that, I already have a structure that isn't too big, but it's something, okay? And this is what we're going to work with, okay? It has, notice that it has a database, it has Docker here. Those of you who are programmers will know him, and those who do n't, well, you can see that he has a lot of substance. And here's the key to the video. Since he already has some things and I'm going to ask him for many more, he's going to spend a lot of tokens in principle. And I want to do a quick test so you can see what would happen if we used Cloue. Okay, I have here, I want you to see it quickly, the usage of Clou, which turns out that I now have practically at zero, okay? current session zero and all week 3%. Perfect. So what I'm going to do is open the code cloud here and paste this promo from here into it, and while it's thinking about it, I'll explain it to you. I want to ask you for quite a few things, okay? I tell him, I want you to add a visual interface to see all these changes that occur on my channel so that the notifications are recorded. This is in case, well, apart from it arriving, I want to have that visual panel. My cat loves me. Okay, then I told him to include a search engine and also to follow the design style he has in a file. But I don't just want that, that would be too simple. I'm also going to ask for a new feature that has nothing
-
5:06
, obre el vídeo en una pestanya nova
to do with what it does, and that is the ability to add a URL to a drive, okay? From a file on a drive, in this case a video. What's wrong with my editor? What's wrong with my editor? And I want it so that when I play it, the music is automatically uploaded to my YouTube channel in a hidden format; I haven't set that option yet, okay? But it's important that an AI pipeline analyzes it, extracts a title, a description, timestamps, and well, maybe in the future we'll even ask it to suggest a thumbnail. Not make a suggestion, but I want it to do all that, and I want it to use a known STK to call various AI services , without me having to provide the URLs. This is basic, but it needs to be mentioned. So , I tell him to analyze all the technologies he's going to use and everything he's going to do, and I tell him not to write code, just to analyze all of that without writing any code yet. So, well, there he is, doing his thing. Let's let him analyze things, but I don't want him to think too much because that's not the goal of the video; I simply want to see how many tokens he can spend. Okay, there he is, starting to spend, to spend. And it's not just the number of tokens, but the tokens here are more expensive than Kimi's. Let's keep that in mind, please. Perfect, guys. Well, I've already asked him here to do things for me. He did the analysis for me, okay? Because he hasn't edited any of my code yet. As you can see here, this is the repository, okay? And nothing has been edited yet. And only with the analysis. My editor will now put a screenshot of what I have left there, okay? It has consumed about 70% of my energy, as it consumes quite a lot of it. So it's clear I'm using the latest model which is the most powerful [music], but look at all the resources it consumes and we haven't even started programming yet. And
-
6:49
, obre el vídeo en una pestanya nova
this is where most people stop, give up, or end up paying more, but we have the solution. And here, friends, I want to share a simple idea with you, one that has changed the way I think and work. People treat artificial intelligence models as if they were a religion. One says, "I'm from Claude. I'm from Chajept." And well, I can understand that. Everyone gets used to their own thing, that's clear, but in the end, an artificial intelligence model is just another tool. And obviously, to make a very simple comparison, you don't use the same screwdriver for all the screws in the house. They come in different sizes, right? Well, this is the same thing. There are tasks that are very costly and worth their worth, and they need the biggest model because they require careful thinking, thorough analysis, and the management of many things. But there's a lot of work that's actually very simple, okay? It's very systematic, and current models do it very well; there's no need to go for the most expensive model that will cost you an arm and a leg. Things like reading an entire project, refactoring, repeating changes, moving things around—these are very heavy tasks where they don't need the most expensive surgeon in the world, they need someone who does it well and costs a reasonable price. And for that I'm going to use Kimi, which according to benchmarks is literally almost as powerful or in some cases even more powerful than Cloude's most expensive model. Obviously this changes over time and by the time you see this video it may have changed, but currently that's where we are. But be warned, apart from being almost as powerful or even more so in some tasks, it costs significantly less. Okay, friends, let's use Open Code now, shall we? And I'm going to do it all from scratch because, obviously, I don't have anything; we had it all in memory, so we're going to do it from scratch using Open Code. I love Open Code, I have it super customized and in fact we can talk about this in another video, although we have
-
8:31
, obre el vídeo en una pestanya nova
already mentioned it in many on the channel. To log in with Kimi you can do it in two different ways. First, you have to press control P here and select Connect Provider. And in the connector you look for Kimi there and Kimi for coding will appear, or not here. Kimi for coding appears to me. Then it asks you for an API key. Where does this come from? Okay, so this is for those of you who have Kimi's monthly plan. For Kimi's monthly plan, you just come here to the console, okay? I have, as you can see here, the moderation. And well, it would be as easy as clicking create. I've already created them all, okay? Copy it, come to your trusted terminal, paste it here, and it will connect. Okay, in this case I'm already connected. If I search for Kimi here, it appears. Well, it seems that way to me because I have several suppliers who offer it to me. And another way to connect it would be via API, okay? This is the monthly plan that I pay for at 19 and separately you can take and pay for whatever you want [music] from the API. For that you would have to come here to platform.kimi.ai. And as you can see, I have 50, which is the balance Kimi gave me to do this test. Of the 50, let's see how much it consumes. So, you load whatever you want here, come to App Keys and generate a new one here. I click on create App Keys and here I put the name. Let's test YouTube. The project, okay?, the one we have by default. We copy, we give it. OK, relax, I've already deleted this pic. When you see this, I'm sorry, it won't work for you. And here too, we type connect and go to search for Moonhot. Okay, here it's different, we're not looking for Kimi, we're looking for Moonshot. We're not in China, so we have to use the normal one. And we put the API key here. We press enter. And as you can see, my High Speed code is already appearing here.
-
10:13
, obre el vídeo en una pestanya nova
Watch out for this one [music]. This one costs twice as much, but it's six times faster. This is crazy. I'm not going to use it because I want to prioritize the money, but if we wanted it to be faster, code High Speed is something that, well, be careful with this. So, I select their latest model, which is Kimi 2.7 COD, and I'm going to put my favorite agent, which is the SSD, which is the SSD Orchestrator. And now I have the same thing, the same project, but a different model, okay? And I'm going to hit send, only this time I've changed one thing in the prompt. Here I'm telling you, it already implements all the functionality, okay? I haven't told him to look at the technologies or anything, I want him to implement them. Then, if necessary, he will ask me questions or I can also directly ask him to test the functionality to make sure that everything is working properly. So this is going to take a while, okay? We're not using the fast one, we're using the normal one. A few moments later. Okay, guys, after 5 hours and 8 minutes, alright? Now I'll explain everything he's done because it's not a small amount, but look here at what we have. And so you can start processing, because this is going to take me a while, I'm putting the drive link I want you to upload here, okay? And here we click "view status" and it's downloading. But let's look at the interface, and the interface, well, it matches what I asked for because I did give him a series of designs that are very similar to what I wanted. And here we see, well, we see things that we can change in the future, but so you can see, here are the changes I've made today. For example, I removed a point from the last video, right? And so this was the title it had before and this is the title it has now. Okay, change the period at the end. I can now see the history here. Besides, I can see here, I have some filters
-
11:56
, obre el vídeo en una pestanya nova
to apply, and I also have a summary. What I really like, and what I hope will now be processed well, is this here in the video. So , while it's downloading, I want to show you what it's done, okay? Because it has been 5 hours and 8 minutes. I told him to be there, to have no limits, and boy, has he had none. I just wanted to show you one thing, and that is that there was a moment when he asked me, but well, to find it here, yes, here it is. He's been asking me questions. Okay, this didn't take long, it took 4 minutes and once I answered the questions, okay, well yes, create the frontend, do everything, etc., then it started doing everything with sub-agents and sub-agents, it took a long time, there was even a moment when it stopped, I think this is because of the open code issue and I told it to continue. And it continued, and I did all this to myself. I configured the MS, configured the environment variables, and that's it , we're all set. And one thing that's cool about this model, okay? Let it keep downloading, let's come here, it has a visualization. For example, I see that this is too close together, so I can take a screenshot and say, "Great, but this is too close together. These cards, as you can see in the screenshot, I'll fix." And since the model has vision, it can look at the capture and fix it. And while he's fixing it and that Drive video is uploading, how much do you think we 've spent? So, ladies and gentlemen, all of this that we've seen, okay? Including the videos that go through a pipeline and all that, it cost us $23.45. Okay, here's the summary and we can see the details here, okay? We could move on to how many tokens he's spent, right? In output. So what I've done is download this from here, export it, come here to CH GPT to do a quick calculation, as you can see, and I've passed everything on to it and told it
-
13:38
, obre el vídeo en una pestanya nova
to calculate it based on the price of Opus 4.8. I also asked for Kimi's price to make sure it's okay. And indeed, the price of Kimi tells me well, what it has cost us 23.45. But look here, this is where it gets important. If we had used Opus 4.8 with cache, which we did , it would have been $6. If we hadn't used cache, with Kimi it would have been $97 and with Opus 4.8 it would have been $14. Do you notice the difference? We're talking about spending neither 76 nor 523. The difference is huge. Obviously, if we had used Kimi's terminal, Kimy Code instead of open code, it would have been different. Perhaps we would have spent less, but oh well, I like Open Code and that's why I used it. And here I am back. And on one hand we have this, which I've fixed, okay? After sending him the screenshot, he was able to fix these cards for me. Look at this marvel. And then, on the other hand, we have the video that I uploaded. Check this out, this is going to be really useful for my channel. He suggested a title, okay? Look, goodbye Microsoft 365 offline, huh, always for €11 and here it gives me a description, some tags, the YouTube category, okay? And the timelines, okay? In this case it's a short film and I've included several timelines. It would work for both short and long videos. And I just check this here, I click upload, it starts uploading and it tells me, "The video is uploaded, watch it on YouTube." And I can come here and see it, it appears here that it is being processed. Perfect, YouTube is processing it and we have it hidden. And if I come here to the channel content, the hidden and uploaded video appears here. And with this I could do anything, change the thumbnail, do whatever. We could find fault with this. I really don't know which ones. Look, I honestly think that if you're going to use Kimi, it's better to use the subscription model than the one that comes from the
-
15:20
, obre el vídeo en una pestanya nova
API, okay? Because the subscription ends up being cheaper. Keep in mind that the subscription costs $ and with that I can tell you that we wouldn't have exceeded this threshold of 23 that we have used. But anyway, as you can see, you can have your agents working for hours and make quite a few changes without any problems, and it does them well. Personally, I'm going to use this tool every day. Let me know what ideas you have in the comments. And that's all I wanted to show you, and that's why I love recording these kinds of videos, because in the end it's not about the new model that kills Claude, okay? I'm tired of these wars, they bore me to death and I do n't think they're going anywhere. It's about something more useful and practical, about not getting married to just one model, okay? Use the one you need to use for the task you need to use it for. That's the smart thing to do. Have several models on hand and learn which one you need for each case. That way you won't be left stranded again when someone tells you to stop. [music] And by the way, the bill at the end of the month does n't give you any shock. If you want to try Kimi on your own, I've left a link in the description. If you already use it or use a setup with several models, tell me in the comments what you use and how you use it, because you don't have to use it the same way I do, that's clear. I love seeing how everyone puts things together, so I'm interested in reading your comments. That said, if the video has been helpful to you, you know that a like helps me a lot, and if you also subscribe—because if you've gotten this far you're not subscribed, you definitely need to subscribe because you're interested in my content. So if you subscribe, you'll help me even more . Thank you so much for watching the video until the end, and we'll see you in the next video. Until next time. [music]