The argument in one line.
The AI agent framework you build is a permanent asset; the model powering it is a rentable, swappable brain — and matching the cheapest capable model to each specific task is the highest-leverage cost optimization most builders skip.
Read if. Skip if.
READ IF YOU ARE…
- You already use Hermes agent and want to reduce your API bill without sacrificing capability.
- You default to Claude or ChatGPT for every task and have never benchmarked alternatives by task type.
- You want to add multimodal image reading and live web browsing to your agent without paying frontier prices.
- You are building or running AI automations at scale where token costs compound quickly.
SKIP IF…
- You have never set up an AI agent framework — the setup steps assume Hermes is already installed.
- You are not running tasks that benefit from cheap, high-volume token usage.
The full version, fast.
Most people lock their agent to one model and overpay. The Hermes agent is model-agnostic — you can swap any LLM in as the engine while keeping your tools, memory, and skills intact. MiniMax M3 matches GPT-5.5 on coding at 4% of the price and edges Sonnet 4.6 on reasoning benchmarks at 8% of the price, with a 1M context window, multimodal image and video input, and live web browsing built in. The setup takes minutes: grab an API key from platform.minimax.io, select MiniMax Global in Hermes config, paste the key. The practical workflow: use a frontier model for initial strategy and architecture; delegate execution to MiniMax M3 as the cheap, fast worker. Always verify outputs — treat any model as a worker, not a boss.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:37
01 · Stop Overpaying For The Wrong Brain
Cold open flagging the hidden mistake of defaulting to Claude/ChatGPT for every task. Intro of host.
00:37 – 01:38
02 · Why Your Model Choice Matters
The five things an agent does (code/build, tool use, sight/web, reasoning, memory/context) and why different models score differently on each.
01:38 – 02:12
03 · Own The Car, Swap The Engine
Anchor analogy: Hermes is the car you build once; the model is the engine you swap per task. Any model can drop in.
02:12 – 03:44
04 · MiniMax M3 vs GPT and Sonnet
Benchmark walkthrough: ties GPT-5.5 on coding at 4% of price; edges Sonnet 4.6 at 8%. 1M context, multimodal, live web, open-weight.
03:44 – 04:42
05 · How Sparse Attention Works
Technical explanation of why M3 is cheap: sparse attention routes only relevant tokens, cutting compute to 1/20th. Plain-English meeting analogy.
04:42 – 05:29
06 · Model Agnostic and Cost Per Token
Output-tokens-per-dollar comparison. M3 leads significantly. GLM beats M3 on terminal coding but cannot see images. Always pick by task.
05:29 – 06:46
07 · Grabbing a Plan and API Key
Walkthrough of minimax.io token plans ($20/$50/$120), API pricing tab, and copying the API key.
06:46 – 07:39
08 · Connecting MiniMax to Hermes
Install Hermes via terminal, select MiniMax Global, paste API key. Switch models in Telegram via /model command.
07:39 – 08:23
09 · Testing Multimodal Image Reading
Drops a screenshot into Telegram chat and asks MiniMax to describe it. Model responds with well-structured breakdown of the Hermes OS UI.
08:23 – 10:01
10 · Firecrawl Brand Identity Scrape
Asks Hermes/MiniMax to scrape Glaido.com, extract brand identity, generate HTML overview. Model asks clarifying questions, executes quickly, correctly identifies brand stats.
10:01 – 11:10
11 · Reading Files and Web Lookups
Tests: find latest desktop video and read its first line; look up latest Jack Roberts YouTube video and report the intro. Both succeed nearly instantaneously.
11:10 – 12:56
12 · Voice Mode and Dynamic Routing
Hermes Agentic OS voice terminal demo with MiniMax. Explains skill-model assignment (Orpheus for deep reasoning, M3 for general work, GLM for terminal). Shows voice reminder creation.
12:56 – 14:01
13 · Worker Not Boss, and Verifying
Three caveats before wiring it in: treat it as a worker (always verify); it is metered; open-weight commercial use is free below $20M revenue.
14:01 – 14:43
14 · Big Brain For Strategy
Recommended architecture: use most powerful model for initial planning; delegate execution to M3. Closes with tease for the Hermes OS video.
Lines worth screenshotting.
- The agent framework you build is permanent infrastructure; the model inside it is a rentable utility — confusing the two is the most expensive mistake in AI tooling.
- MiniMax M3 matches GPT-5.5 on coding benchmarks at 4% of the price, which means the same output for 25x less cost on coding tasks.
- Sparse attention cuts compute to one-twentieth by only routing tokens to the context that actually matters — that efficiency, not a discount, is why 1M context costs cents.
- GLM-5.2 beats every model on command-line coding but cannot see images; MiniMax M3 is the reverse — always route by what the task actually needs.
- Model-agnostic agents let you swap the winning model every few weeks as the market shifts, without rebuilding your tools or memory system.
- Using a frontier model for strategy and a cheap model for execution is not a compromise — it is the professional architecture.
- Output tokens per dollar is a more honest metric than benchmark scores alone; M3 generates roughly 7x more output tokens per dollar than GPT-5.5.
- Open-weight commercial licensing below $20M revenue means you can ship MiniMax M3 inside a product without attribution requirements.
- Hermes routes queries dynamically — you can pre-assign specific skills to specific models and let the OS dispatch automatically.
- A 12-word test for any model selection: what are the five things this agent does, and which model wins each?
Match the model to the task, not the habit.
WHAT TO LEARN
Defaulting to the same model for every agent task is a habit that silently inflates cost and caps performance — and the fix is a five-dimension scoring rubric applied before you wire anything in.
- Every AI model has a different strength profile across five task types: code and build, tool use, web and sight, reasoning, and memory/context — score any new model on all five before committing it to a workflow.
- The agent framework (persistent memory, tools, skills) is worth building carefully and keeping; the model powering it is a rentable component worth reconsidering every few weeks as the market moves.
- Sparse attention is the engineering reason a 1M-context model can cost cents instead of dollars — understanding this helps you evaluate any future model claiming cheap long-context, because the mechanism is auditable.
- A model that leads on cost-per-output-token but trails on a specific capability is not a worse model — it is the right model for a different subset of tasks.
- The practical two-tier architecture — frontier model for initial strategy and planning, cheaper worker model for execution within that structure — applies to almost any multi-step agentic workflow regardless of which models you choose.
- Treating an AI model as a worker (verify outputs, route hard decisions upward) rather than a boss (trust autonomously) is not pessimism about the technology — it is the professional posture that scales without data loss.
Terms worth knowing.
- Hermes Agent
- An open-source agentic operating system built on top of Claude Code that supports swapping in any LLM as its reasoning engine while retaining persistent memory, skills, and tool integrations.
- MiniMax M3
- An open-weight multimodal LLM from MiniMax with a 1M-token context window that matches frontier models on coding and reasoning benchmarks at roughly 4-8% of their API price.
- Sparse attention
- An architecture technique where a model only computes attention between tokens that are most relevant to each other, reducing compute to roughly 1/20th of dense attention at the same context length.
- Model-agnostic
- An agent or workflow design where the underlying language model can be replaced without rebuilding the surrounding infrastructure, tools, or memory.
- Firecrawl
- A web scraping tool that extracts structured content from pages by pulling only the relevant HTML tags rather than the entire page, used here as a skill wired into Hermes agent.
- Dynamic routing
- A Hermes feature that automatically dispatches a query to the best available model or skill based on the task type, rather than sending everything to one default model.
- Open-weight
- A model whose weights are publicly released, allowing it to be self-hosted, fine-tuned, or embedded in commercial products under the terms of its license.
Things they pointed at.
Lines you could clip.
01:38
“You're gonna own the car and swap the engine.”
03:57
“It's the same answer, a fraction of the noise, and one twentieth of the compute.”
12:56
“Think of this as a worker, not a boss. It can be brilliant on the routine 90% of the time, but always, always, always verify.”
14:01
“You own the agent and rent the brain.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00Hermes is the best AI agent in the world, but most people ignore one system that costs you more money and makes your results worse. So in this video, I'm gonna show you exactly how to increase the performance of Hermes' agent by using the best model for each task, and it's something you can do so easily even if you've never used Hermes before, so you can save more time, get better results, and get light years ahead of everybody else.
00:25And if you're new, I'm Jack. I built and sold my last tech startup with like a gazillion customers, and now I build my own AI startups, and I share here the things that actually work. So if you haven't already, grab that beautiful coffee, and let's drive straight in.
00:38And so with our beautiful Hermes agent, the first thing that we have to do basically is to stop overpaying for the wrong brain. The beautiful thing about Homi's agent is the fact that we can use any model we want to on the planet, which is absolutely fantastic. And the core concept to that I want to think about is you have to understand the things that the agents actually do.
00:58They can code and build, tool use, so do I wanna go and scrape things from the internet? Do I want to build things, connect to different softwares? The ability to go onto the web and see things and visualize things.
01:09Its reasoning ability, its ability to store memory and context and how long can it think? Think of it as basically just , has a different skill set for a specific task. And some drivers are more expensive than others, and some are way more capable.
01:26And the core thing we have to do is break the duopoly of only Claude and only ChatGPT because you're missing out too much if you're only restricted to those two models. So think of it like this, you're gonna own the car and swap the engine. There's a really good analogy.
01:41This is our beautiful Hermes and we can drop in anything from Claude, Mini Max, GLM, GPT, whatever we physically want to. And in this analogy, Hermes itself is the car.
01:50It's the wheels. It's the wiring of the dashboard that you build once and then you keep. And all we're ever gonna do is basically just decide which engine do you wanna drop in for each specific task that we're doing.
02:00And if you follow this process, you will get more performance per dollar of input that you get from Hermes agent, meaning that the agent that you're using is just way more capable. And this could apply to any model, but the model I'm gonna be talking about in this video is MiniMax v three.
02:15Just to understand what this looks like, if you compare it to GPT 5.5, on the benchmarks, it is tied on coding for four percent of the price.
02:25Imagine a model that's as capable as GPT 5.5, but it's only like one twenty fifth of the price.
02:32For Sonnet 4.6, it basically edges it on some things. Obviously, you can check out the benchmarks, and that's 8% of the price.
02:38The model itself, and it's so easy to set up. 1,000,000 contacts window, it can seize images and videos which some of the other upcoming models don't at the moment. It can browse the live it can browse the live web and open white.
02:50Now this model is super cool. And I emailed MiniMax and they even agreed to sponsor this video. So thank you MiniMax for doing that.
02:56And I really want to sort of lay home that whatever the model is that we're switching out, to break down this duopoly of basically either Claude or ChatGPT will open up your horizons and your performance per dollar you spend will go further and higher.
03:11And I'll show you exactly what I mean about that. So Mini Monkey three has done something really interesting, which is why I was really excited to talk about it today. And you can see its performance here about how it basically matches up to the best models in the world right now for things like reasoning, command line code, and web.
03:25And you can see the different trade offs in terms of where it's great and where it's up and coming and those different things. But what it did that was really interesting is you have this general line here, right, which is basically how good is the model versus how expensive is the model.
03:38And one of the really cool thing is an outline in terms of its power against the actual price, which is really freaking interesting. So you may be thinking awesome, but how is it actually achieving this result? How is it able to be this good at the certain particular price point?
03:51Well, apart from the fact that models are just evolving and better, is using something called sparse attention. So think about it from this point of view, there's an expression that says that a horse designed by committee is actually a goat. In one scenario, everybody's talking to everybody.
04:04It's chaotic. It's slow. But over here, we only basically talk to the people that actually matter.
04:10Therefore, it's 20 times less work. So if you think about it from this perspective, you picture a meeting where every person talks over each other is pure chaos. M three's minimax sparse attention only lets the people who matter speak.
04:21It's the same answer, a fraction of the noise, and one twentieth of the compute. And you can see a graph just to kinda explain some of the technical stuff that's happening under the hood with the model. But the tilde r is it is this efficiency is the whole reason why 1,000,000 contacts cost cents instead of dollars in other places.
04:35And then we look at what a dollar is actually buying with a model like this. So you can see the output tokens per one dollar to mini max is really good. Remember, every model is great for one specific thing.
04:46So it's not just about committing and only abusing one model. We wanna use the best model for a specific job. That's the whole thing is that we are model agnostic.
04:56And only by looking at data points like this can we really start to break down these perceptions we have and get way more output out of the models that we're using. And so you can see, if you compare it to GLM 5.2, which is going super viral right now, which is again an awesome model.
05:09These are all great models for different things. You can see some of the differences that MiniMax can natively see, browser web, and four times cheaper output per 1,000,000 contacts. And GLM itself is crushing it for command line coding.
05:20It is even better, I believe, than 4.8, and the tests are crazy. And it's such an exciting time to have all these different models that we can tag in and use for anything that we want to.
05:29So I'm gonna come over now to minimax.io, and I'm on platform.minimax.io. I'm gonna get grab a plan.
05:35You can do twenty, fifty, a $120, and you can see the tokens that you get for instance of usage. So the $20 plan is 1,700,000,000 tokens.
05:42Pretty much just like you would do with a cloud subscription or a chat GPT. It's basically just how it all connects together. Then once you've chosen a plan that you want to, if you wanna give MiniMax a whirl.
05:51So you've got the token plans, which are gonna be more economical, or you've got the API pricing. So you can just come down and check out what the actual pricing is for this if you just wanna kinda pay as you go with the model.
06:00And to grab your API key, we'll just come down and click on this get API key. And then when you come down here to subscription and just as we copy this value here. Now if you don't have Hermes Agent you're gonna come over to the Hermes website, and all you're gonna do is come down and literally copy this code.
06:12And then once you've got the code, you're gonna come down and do terminal. And when that pops up, we're gonna enter it in like so, and this will download the entire Hermes agent on a computer. And if this sounds like I'm speaking Spanish, I'm gonna pull the link down below for my full Clawd Code Masterclass that will take you from zero to hero.
06:27We go through foundations, building a website, power features, memory systems, Hermes agent from from start to finish, all the way through, and you also get the full beautiful agentic operating system, the Hermes OS that has so many cool bells and whistles that will take you to a completely new level. So I'm gonna put a link for this down below so you can come grab that if you want to.
06:46Now here's a cool thing. Terminal is here. Now what we're gonna do to connect MiniMax is we'll literally just come down and we're going to select it down here, which is MiniMax Global.
06:54Just do space bar like so. And we're gonna go for the MiniMax at the top by clicking space bar, and then it's gonna ask for the API key and we'll just drop in that API key. And once that's done, hit enter like so, and the API key is now basically saved.
07:07And now we can begin to have a conversation with it in Telegram. Again, if you've not seen that before, either click the link down below or check out this video on screen for setting up Hermes agent from start to finish. Beautiful.
07:17Now I've opened up Telegram. Let's switch over to MiniMax and take it for a spin. So I'm gonna forward slash, I'm gonna type in model like so, and then the model selector will pop up with inside Telegram, which is excellent, and we can just switch over.
07:29Cool. So So you can see MiniMax is right there. So we're gonna click on this one, and then you can pick the model that you want.
07:34So we're gonna go with MiniMax m three, and then we are gonna be ready to have a full conversation. And so now that's connected, let's ask it a question. Hey there, which model is this?
07:43And just like that, it should come back and say, hey, this is MiniMax. And the first thing I wanna demonstrate here is the fact it's got multimodal capabilities. So what we're gonna do is take a screenshot of this, drop it in, and ask it exactly what it is.
07:55Let's come down, take clipboard, which is awesome. Drop it over here. Hey there.
07:59Describe to me what you see in this image. Drop it off and see its ability to just work multimodally in anything that it's doing. And just like that, it's come back with this and I love the way this broken this down.
08:08It's got like, bear in mind, when you're using these models in Talagon, they don't always format it the best, and it's done an incredible job of this. So let's put it to the test using our fire crawl skill. So what I'm gonna do is come down and get a prop up.
08:20Hey there. I want you to use the fire crawl skill. I want you to go over to glider.com.
08:23I want you to extract that brand identity, and then I would like to create for me just a little bit of a mini HTML overview that I can open up in browser and just understand about that brand, color palette, logo, that kind of thing. Send that up.
08:36Pretty complicated task. I'm using Firecrawl mainly because it's the best way to actually grab information. And just before I even answer, guys, look how quick this came back and asked me clarification questions.
08:45And it's asking, is it glider.com? I'm be like, yes. It is glider.com.
08:49And yes, you should have a skill to use Firecrawl. Gonna come down, hit enter on that one, and watch it work its magic in the background. This I think it could just be me, but this does seem to be coming back extremely quickly on the mini mags, which is pretty impressive.
09:01Now while that's working in the background, let's fire up Firecrawl over here and just pull this bad boy up. Now in Firecrawl, we can search the web, we can scrape websites, and we can also interact with them as well, making it pretty cool. And one of the reasons why this is so epic, I talk about it all the time, is that you're not pulling back all the HTML from the page.
09:17It's just extracting the correct tags, which saves us a fortune. And one thing I will say is I am super impressed with how quickly it's coming back and giving us updates whilst it's doing everything. Quick update on the website itself, glider.com appears to be a single page app with minimal static HTML.
09:32Glider, of course, is the software that I'm using right now to transcribe everything and has time able to, like, talk and the sort of text just appears. And then just like that, guys, it's come down, it's confirmed what it's done, and this is what it's created for, the Ziglado brand guidelines. Look at this.
09:44It's correctly identified that we save people twenty hours a month. It's got all logos. It's got the color palette, and all these color palette, guys, it actually grabbed from legit website.
09:54I'm really, really impressed with it. I mean, look at this. It's even grabbed this image.
09:57That is actually wild. I can click on this and go over to the actual website, and you can see the similarities. Now, let's test it with something more difficult.
10:05Why don't we ask it to actually check out a video on my desktop and see how it can handle that? So if I come and say, hey, I want you to go into my desktop, find me the latest video that I recorded, and tell me what is the first sentence I say in that video. Now I haven't necessarily given it the tools to do this.
10:20We can see if we can figure it out. And so it's done that and it's found the pizza commercial I did in my ads video, which is crazy. Let's give it one of the challenge here.
10:28Let's go over this and give it this issue here. So why don't I come down and say, hey, now head over to the Jack Roberts YouTube channel, check out my latest YouTube video, and tell me what is the intro for that video. Okay.
10:38Now I'm asking it to go online and grab my information. Okay. So we can see now it's got my channel ID.
10:43Look at my latest video. And I wanted to share how it's working because one of the thing that really jumps out to me with MiniMax is how quick it actually is at processing information. Even the stream and the speed at which you get the information back is crazy.
10:55And look at this. That is actually completely correct as well. It's actually pulled down one of my latest videos, which is pretty incredible.
11:00Cloud Code Agenic systems are the future. Online capabilities at nine times people don't exist. Blah blah blah.
11:05This open source repo will 10 x your Hermes agent. And it got it almost instantaneously, which is crazy. And then we can also head over to the Hermes Agenic OS.
11:13I can come up to intelligence and begin a conversation with Hermes just in the same way using the minimax model. I can come down here, begin the conversation.
11:22So I might say, hey there, who am I talking with today right now? Hey, Jack. You're talking with me, Hermes' voice.
11:28I'm your AI companion. And essentially anything we can do in the chat, can also do in this voice terminal. And the cool thing is you can speak to Minimax in Hermes agent and say, hey, I want you to build me out a routing strategy.
11:41And remember, Hermes will dynamically route to any specific model based on the task at hand. I've actually built a full score for this. I'll put it down below in the description so you can grab it if you wanna check it out.
11:51But just enables Hermes to directly use its intelligence to route to the correct resource. I did a full video on it.
11:57I'll put a link on screen so you can check that out. I remember guys, you can also give reminders like this. Hey, Go ahead and send me a reminder that in two hours time, I need to order some La Croix.
12:07Don't judge me guys. Coke is out. La Croix isn't.
12:10It just happened. Don't blame me. It just is the way that it is.
12:13And just like that, it creates reminders in the exact same way. And bear in mind guys, that Hermes agent can dynamically route your queries. So you can say to Hermes something like, hey, that I want to create for me a system, a skill that will dynamically route my queries based on the model that I'm using.
12:27And bear in mind, Hermes can do this. Obviously, if you're in an agentic operating system, you can actually, and this is a really cool thing, build out specific skills with specific models. For example, you may have the philosopher for some different models.
12:40I have Orpheus, which is deep reasoning on different topics. You can use minimax for certain things and pre allocate certain skills to certain models with certain objectives, meaning that it roots there dynamically and automatically, which is a huge huge benefit.
12:54Now couple things to know before you wire this into your awesome system. So number one is that you need to think about this as a worker, not a boss. It can be brilliant on the routine 90% of the time, but always, always, always verify with any model that you're building.
13:09Okay? Really, really important. And so the key thing to understand is that we want the right model for the correct job.
13:14But whether that's gonna be MiniMax or Opus or GPT, we are model agnostic. We bring in the best guy for that particular job. Basically, me, token efficient, it's gonna increase performance and actually save you cost as you scale at the same time.
13:29And a great way to think about this is that you own the agent and rent the brain, and MiniMax have kindly offered us 12% off any tiers. So if you come down here with a link, I think I'll put it down below, you'll also get 12% off this, which is super generous, and we appreciate that. And remember, you can automate the switching of brains with a really cool, basically, tool routing skill that you can add directly into Hermes.
13:49And the other good thing to know is with MiniMax is if you're using it for commercial purposes, they haven't asked that you put a built with Mini Max m three as a credit on it. And if it's above $20,000,000, then you can basically reach out and have a separate conversation.
14:02Now generally speaking, when we're talking about doing the right job, you want the most powerful brain possible to do the initial planning and conversation. That's where you get the maximum bang for your buck.
14:13Scurping it out, designing the dashboard, whether it's an operating system, whether it's a course curriculum, or whatever it is, right, whatever you're building out, you essentially want the big brain to do big strategy stuff, then we can start to delegate different models, basically to drive that within the structure that we've already created.
14:29Now, knowing the right models to use is one thing, but if we don't have an operating system to combine them all together properly, we're not gonna get the most out of our entire system, which is why the next thing we need to do is leverage all those together, which we cover in this video right
The Hook
The bait, then the rug-pull.
The title makes a bold claim — and then immediately turns it against you. Within ten seconds, the video pivots: the agent is great, but the way you are using it is bleeding money. That bait-and-switch is the actual hook, and it works because the mistake it names is real.
Named ideas worth stealing.
00:53list
Five Things an Agent Does
- Code & Build
- Tool Use
- Sight & Web
- Reasoning
- Memory/Context
A scoring rubric for comparing models: evaluate every candidate on these five dimensions before committing it to a task type.
Steal forModel selection for any agentic workflow
01:38concept
Own the Car, Swap the Engine
The agent (Hermes) is the car — the permanent infrastructure you build once. The model is the engine — a rentable, swappable component you choose per task or per week.
Steal forExplaining why to separate agent framework from model choice in any AI system
12:56concept
Worker Not Boss
Treat any AI model as a worker who handles 90% of routine tasks brilliantly but always requires verification, not as an autonomous decision-maker.
Steal forSetting correct expectations when handing off AI tasks to clients or teams
14:01model
Big Brain for Strategy, Worker for Execution
Use the most capable (expensive) model for initial planning, architecture, and decisions. Once the structure exists, delegate execution steps to cheaper models.
Steal forCost optimization in multi-step agentic pipelines
How they asked for the click.
VERBAL ASK
14:01next-video
“which is why the next thing we need to do is leverage all those together, which we cover in this video right here”
Soft close — cuts to a suggested video card rather than a hard subscribe ask. CTA is embedded in the final sentence of the tutorial content, not a separate outro segment.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
