{
  "video_id": "w-X3HV2OTfM",
  "channel_slug": "techwithtim",
  "channel_handle": "Tech With Tim",
  "title": "50x Cheaper Than Claude - But Can It Actually Code?",
  "duration_seconds": 921,
  "url": "https://www.youtube.com/watch?v=w-X3HV2OTfM",
  "upload_date": "20260407",
  "transcript": "What you're watching right now is an AI model writing\ncode, testing it,\nfiguring out what went wrong,\nand fixing it with zero human involvement.\nNo one's prompting it between steps.\nIt's just going.\nNow, if this is MiniMax M 2.7,\nyou may have heard of it already.\nAnd the reason I'm covering\nthis is not just because it's a good coding model.\nIt's because this model helped build itself.\nMinimax literally had the model construct\nthe system used to train the next version of itself.\nAnd I think that's a pretty big deal and something\nthat not enough people are talking about right now.\nSo in this video, I'm going to break down what\nthis actually means, how well the model performs.\nAnd then I'm going to use it\nto build a bunch of real projects.\nSo you can see exactly the performance of this model\nand kind of the new self-improvement\nera that we've moved into.\nOkay.\nSo the model helped build itself.\nNow what does that actually mean?\nWell, simply put, when you train an AI model,\nthere's a ton of infrastructure around it.\nYou have data pipelines, experiment\ntracking, evaluation systems, and normally\nhumans build all of that.\nBut minimax said,\nwhat if the model can build it itself?\nSo they gave em 2.7\nthe job of constructing their entire research\nagent harness,\nand it doesn't mean it wrote some random scripts.\nA single person with zero\nmanual coding had Em 2.7 autonomously.\nArchitect a full agent harness complete\nwith CI testing and code review in four days.\nThey then let it continuously improve\nthat system based on the results and the model runs.\nAn experiment sees what failed,\nmodifies its own tooling, and then run to get\nso over 100 rounds of this\nwith no humans touching anything.\nAnd by the end of this, there was a 30% improvement\non the internal benchmarks,\nand the model found optimizations\nthat the human team just hadn't found yet.\nAnd this isn't just an internal claim for minimax.\nYou can see here\nI have the sweep row and the ML bench results that\nI just put into a nice diagram here.\nAnd this ML bench 22 Kaggle competition,\nthis is what you're seeing right here\nwas released by OpenAI\nto measure autonomous AI research.\nNow M2 .719\ngold medals in this with a 66% average score,\nwhich is right behind the current\nfrontier models like GPT 5.4 and opus 4.6.\nSimilar results.\nWhen we go and look at the sweep row benchmark, where\nwe're right behind the two best models as of now.\nNow with these results, minimax is clearly one of\nthe current best models that is out there right now.\nAnd the thing that excites me the most\nis that they just announced\nthat they're going to be releasing\nthe open weights for this model in just a few weeks.\nWhen I post, this video may already be open,\nat least when I'm recording it.\nThey said it's going to be about two weeks from\nnow that they're releasing the model.\nAt least that's what I currently know.\nNow with this minimax model,\nthe one thing that does stand out, as opposed\nto just running it on coding tasks,\nis how well it complies with tool calls.\nSo especially if you want to run this instead of\nsomething like open claw, this is a pretty big deal.\nNow what they did is they ran an experiment\nwhere they had 50 plus skills and 100\nplus features, and minimax had a 97% compliance\nrate in terms of calling those tools\nand working with those different features.\nNow, the average model struggles a lot.\nWhen you give it a ton of different tools,\nyou can see it usually has a 74% compliance rate,\nwhere it's minimax up to that significantly\nso in complex environments\nwith a lot of different tools set, this is something\nthat's really shining, not just purely for coding.\nSo the bottom line here is that there's a frontier\nlevel model competing with the best models out there,\nbut with a real edge and agent reliability\nlike we just saw with the tool calling here.\nAnd what I want to do\nnow is get into a bunch of demos.\nWe can actually test this out\nand see how it performs in the real world.\nSo what I want to do now is show you how\nyou can actually use this agent, go over the pricing,\nall of that kind of stuff,\nand then I'm going to walk through a bunch of demos.\nYou can see the real world performance\nbefore you go test it out for yourself.\nNow first thing to note\nis that minimax M 2.7 here is available via\nan API, a token subscription plan,\nor directly from their website.\nThey actually have a lot of different options.\nThe easiest way to use it\nif you just want to test its basic ability\nto call tools or use it in some kind of claw board\nor something, is directly from the agent platform.\nSo agent on minimax audio.\nBut I know most of you are developers\nand you're more interested in using it in\nlike cursor\nor Claude code or something along those lines.\nSo the way that you can do that is\nyou can go to their API platform.\nI'll leave a link to in the description\nplatform dot minimax, dot io.\nAnd then you can either purchase a token plan,\nor you can just pay for your usage using an API key.\nSo they have this new token plan concept,\nwhich is kind of unique to minimax.\nAnd what you can do is you can pay a monthly fee\nor you can subscribe to it annually\nif you want to save and get two months free.\nAnd this token plan will give you access to different\nrate limits and speed when using the plan.\nSo you can see like the highest\none here is 30,000 novel requests every five hours.\nAnd the interesting thing about these token plans\nis they actually give you access\nto multimodal capabilities in terms of text\ngeneration, video generation, text to speech.\nIt's all in one subscription,\nso you don't need to pay extra.\nNo big shout out to mini Max for partnering with me\non this video and giving me access to the plus plan.\nHere you can see just 20 bucks per month,\n4500 model requests every five hours,\nand then just normal speed, right?\nAnd then you can use an open claw,\nall of this kind of stuff,\nas well as generating the images and the speech.\nSo what I'm going to do\nnow is go to my account from here.\nWhat I'm looking for is my token plan key.\nSo if I go to token plan you can see that\nI have the key right here.\nAnd then you're able to see all of your usage.\nSo what I'm going to do is copy this key.\nAnd then I'm going to quickly refer\nover to the documentation and show you how we can\nset it up inside of cursor as well as Claude code,\nbut you can effectively use it anywhere.\nAll right.\nSo it get set up.\nIt's pretty basic, but I would recommend\nthat you refer to this documentation.\nAgain I'll link it in the description.\nYou can see you can use it\ncursor open claw kilo Clyne root code.\nSome of these I've never even heard of Codex\nor Droid Z.\nAll of this kind of stuff\nand it shows you the exact setup.\nFor now,\nI'll set it up in Claude code\nbecause I know a lot of you're using that,\nand then I'll quickly show you the cursor\nconfig as well.\nNow to use this inside of Claude code,\nwhat we need to do is we need to find\nour dot clod slash settings dot Json file.\nAnd we need to update the configuration\nto change these environment variables.\nSo what we're going to do\nis we're going to just set the\nand throughout the base URL to be the minimax API.\nAnd then we're going to set the anthropic auth\ntoken to be our mini max subscription token.\nNow you also I'll just quickly show you\nif you don't want to buy this subscription,\nyou just want to use it on a per cost basis.\nYou can create an API key from this API platform,\nand then you can go\nand just add some balance here, and then you can just\ndirectly run it through the API.\nYou don't need to have the token plan,\nbut the token plan is obviously going\nto be significantly better value.\nOkay. So with that in mind, let's go to Claude.\nLet's find that file and let's set it up.\nSo there are some different ways\nthat you can set this up.\nIn my case, because my Mac I can just\nrun this command now I already configured it earlier,\nbut you can see that I just paste it in that env\nblock directly inside of my settings.\nNotice it's outside of the enabled plugins,\nand we have the base URL and the token.\nSo then I can just get out of that.\nI can rerun Claude\nand then you can see it's now using minimax.\nAnd if I say something like hello, we just\nwait a second here and then we should get a response,\nnoticing that it's using the minimax\nM 2.7 model. Okay.\nAnd you can see that it gives us the reply.\nAnd then we can continue.\nAnd we can use this in the next quick setup\nthat I want to show you is inside of cursor.\nSo let's do that and let's do a bunch of demos.\nSo same thing here. Just referencing the docs.\nWhat it says is of course you need to download,\ninstall and sign into cursor, click on models, expand\nthe API keys and then override the API base\nURL as well as the API token.\nSo it says just change the API key and the base URL.\nSo let's see if we can do that.\nSo from cursor we're just going to go to the settings\nhere we're going to go and find models.\nFrom models we can go to the API key section.\nAnd then what we can do is we can just override this.\nSo we can change the base URL.\nAnd by toggling that switch that over to mini Max.\nSo let's grab the URL that it's telling us\nright here, which is going to be this one right here.\nOkay.\nSo let's go and paste that in okay.\nAnd then we need the API key.\nSo same thing. We just go to our account\ncopy the key.\nObviously\ndon't share with anyone and paste it here okay.\nNow we also need to enable this.\nSo I'm just going to enable the toggle\nright there okay.\nSo now that we enabled that which we just did\nwe need to go to models V1 models\nand then add a custom model with it\nwhich is mini Max M 2.7.\nSo let's go here to models.\nWe're going to go view all models.\nAnd then I think if we scroll down\nit says add custom model.\nFrom here we can just paste mini Max M2 point seven\nand then just simply add that one in.\nAnd then we should be able to start\nusing it from cursor okay. So now it's on.\nSo let's just open up a new project\nand let's start running some demos okay.\nSo from cursor\nlet's make sure we select the mini Max model.\nSo mini Max M2 point seven.\nAnd what I'm going to do is just see how\nwell this can one shot a few different applications.\nLet's start with that for a coding task.\nSo I have a prompt that I wrote before\nI said build a real time system monitoring dashboard.\nI want you to use Next.js\napp, router, TypeScript, etc.\nI want you to monitor the CPU.\nI want you to have a process list.\nSo essentially just monitor\nwhat's going on in my computer.\nThis is a relatively complicated task\nthat I would be impressed\nif it was able to one shot with really any model.\nBut let's run\nit and let's see the result that we get.\nAll right.\nSo it just finished up here.\nYou can see this is the prompt that I gave it.\nIt had six to do's went through completed everything.\nIf we quickly have a look at the application\nyou can see the things are nicely separated.\nWe have type styles, components, the app, the source.\nOverall everything looks pretty good.\nWe also have an API for pulling the metrics.\nLet's quickly look at those and we can see that\nwe're correctly pulling from system information,\nwhich is what I was hoping it was going to do.\nYou can see we get details about our Mac.\nThis is a local IP address by the way,\nso don't try to, you know, hack me.\nHere we have our current network usage inbound\nand outbound in a nice kind of live chart.\nWe have all of our processes.\nYou can see OBS is using a lot of CPU\nbecause I'm recording the screen.\nWe have.\nYeah, pretty much everything that I was asking for.\nSo what I want to do now is\nI want to ask some follow up prompts,\nget some improvements to the user interface,\nand let's see how it handles that.\nOkay.\nSo what I'm going to do here\nis ask you for three changes.\nFirst I want to add an alerts panel.\nSo if the CPU goes above 90% as it is right now\nI want to be alerted.\nI also want the processed list to stop rerendering\nbecause I believe how it's written.\nIf I quickly checked it's Rerendering\nand it's not Memoized, then\nI want to add a time range selector\nto the network chart so that I can choose.\nYou know what range\nI actually want to view the network at.\nNow I'm asking three changes at once,\nbecause I want to test to see if it's able\nto actually handle all of them at once, rather\nthan just one by one, because usually the weaker\nmodels will fall apart when you try to ask them,\nyou know, multiple changes at one time.\nSo anyways, let's submit\nthis and let's see what we get.\nOkay, so it looks like it was able to handle that\npretty quickly.\nWe can see that we have the alerts\nnow saying the memories above the threshold.\nWe can also go and check the network activity.\nAnd we can change the range here.\nAnd you can see that it's slightly adjusting\nso we can see a little bit better.\nOf course, because we haven't been monitoring for 5\nor 15 minutes.\nIt's not really doing much for us\nwhen we change that, but later\nthan we would, you know, obviously\nbe able to filter it a little bit better.\nCool. And then we have the processes\nand all of that is good.\nNow, one thing that I am noticing\nis that I don't really love the style of the page,\nespecially like this scroll bar and stuff.\nSo what I'm going to do\nis I'm just gonna take a screenshot of this and I'm\ngoing to hand this over to the model,\nand I'm just going to say\nreview this and improve the style.\nAnd design of the page.\nSo it's let's just go with sexy here\nand see what I can do there.\nSo we're just in the middle of the UI\nimprovement here\nalready looking a lot better\nin terms of the enhancements that it's made.\nI mean,\nit just looks like much more professional now.\nSo let's see what happens once it finishes up.\nSo it looks like it's just wrapping up here.\nOverall this looks significantly better.\nI'm pretty impressed from a single prompt.\nWe were able to get that.\nBut obviously I want to keep going.\nAnd now I want to test something that a lot of models\nstruggle with, which is a full refactor.\nSo at this point, it's built everything out.\nBut a lot of times when you ask these models\nto refactor the code, because there are so many\nchanges and so much context going on,\nit will break something or it'll mess something up.\nSo I'm essentially going to say, now\ngo back and refactor the entire data fetching layer.\nAnd then right now each component is pulling its own\nAPI route independently.\nI want a single WebSocket connection that pushes\nall the system data in one payload every two seconds,\nand each component subscribes to just the fields\nthat it needs via a react context provider.\nSo a more advanced architecture here.\nLet's see\nif we'll do that without breaking something.\nI'll be surprised, but let's run it.\nAnd let's also just quickly\ncheck the usage while we're at it.\nSo right now\nyou can see we're running 139 out of 4500 tasks.\nNow I've been running this a fair amount.\nWe have what, four prompts\nwith a lot of different tasks going.\nI ran a few other prompts. We're kind of in between.\nAnd I was recording,\nyou know, messing around with the model.\nAnd generally we're nowhere near close\nto hitting the usage window,\nand we could be running like multiple sub agents\nright now and be completely fine.\nSo very, very cheap.\nAnd remember, I'm paying $20 per month for this\nwhile I got it for free, but I would be paying\n$20 per month, so it's not like I'm forking out\n$200 per month with the thing\nlike clawed and hardly being able\nto stay within my usage limit.\nAll right, so it looks like the refactor\njust finished up here created this system\ndata context for us.\nAnd actually it looks like this is pretty good.\nIt's pulling everything out of this kind\nof context window that we had exactly as I asked for.\nAnd then it made all the different file changes.\nHere you can see it's mentioned\nall of the things that we can delete now.\nIt just left them there as reference.\nAnd if I go back to the monitor,\nlet's just refresh here\nand we can see that everything is working.\nWe have the memory usage alert,\nwe have the network tab, we have CPU memory.\nAnd it didn't break anything,\nwhich is honestly pretty impressive.\nConsidering again, you know, the relative cost\nof this compared to some of the frontier model.\nSo now let's get a few more last things to do\nand really stress this.\nOkay, so I have a much more complex prompt\nnow where I'm going to say I want to deploy this\nas a lightweight, self-hosted tool that people\ncan install on any Linux server with one command,\nwalk me through the trade offs, should I packages\na Docker container, a single binary, etc.\nthink through this and then give me a recommendation\nand then we'll have it build it out.\nSo let's see how we can reason on something like this\na little bit more complicated with kind of DevOps\nrelated stuff.\nAnd I probably\nshould have put that into the ask mode.\nBut anyways, let's see the result that we get.\nOkay, so that was surprisingly fast.\nIt's giving me a few options here.\nDocker being number one,\nwhich is what I probably would have gone with single\nbinary npm install script.\nAnd then let's see what it's suggesting.\nOkay, cool.\nSo let's tell it go with Docker okay.\nGo with Docker and build it out.\nAnd let's see if we can make that.\nAll right. So that just wrapped up.\nLooks like we now have a full Docker file\nas well as a docker compose dot Yaml.\nAnd then read me instructions\non how to set all of this up.\nOverall, I'm pretty impressed.\nI know it is a relatively small demo, but generally\nspeaking, I mean, it didn't really make any mistakes.\nThere's hardly\nanything that I need to go back and forth with it on.\nI literally just did four prompts.\nSo we built out this entire application.\nAnd again,\nif we go back to the usage here, we can refresh.\nWe've really hardly used anything.\nWhat do we have maybe.\nYeah 4% in like I've been recording for a while now.\nMaybe, I don't know, 30 40 minutes.\nSo if we just spun up a bunch of sub agents kept\nusing this, we're going to be nowhere near our usage.\nAnd again, at just 20 bucks a month,\nI think is a really good deal.\nNow, look,\nI think that this model is really important,\nnot just because of the performance,\nbut because of the way that it was trained.\nIt's super interesting\nthat now AI models are training AI models,\nand we're kind of getting into this\nlike matrix situation where AI is building AI.\nThat builds AI, that builds better\nAI and showed it to minimax\nfor discussing all of this and promising\nto open source these model weights very soon.\nI can't wait to start running this through something\nlike a llama running into my open client instances.\nAnd I just think the cost to performance here\nis really, really good\nand you guys should definitely consider\nchecking it out.\nSo anyway,\nso that's it. I'm going to wrap up the video here.\nIf you guys enjoyed it make sure they like\nlet me know what you think in the comments down below\nand I will see you in another video.",
  "transcript_chars": 18863,
  "ingested_at": "2026-05-21T19:01:18.365009+00:00",
  "source": "retry-no-transcript",
  "yt_meta": {
    "view_count": 64217,
    "like_count": 777,
    "channel_id": "UC4JX40jDee_tINbkjycV4Sg"
  }
}