26:57Riley Brown · Tutorial
The Latest Codex Updates and The Truth about Opus 4.8
A 27-minute breakdown of why Opus 4.8 barely moved the needle and why Codex platform updates mattered far more.
May 31stThe argument in one line.
Claude Mythos will be the most expensive AI model ever shipped, and the only way to justify that cost is to build your benchmarking infrastructure — examples, tool integrations, and ROI metrics — before it arrives.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code or Claude for business tasks and want to evaluate whether Mythos is worth the premium over Opus.
- You work at a company and need to build an ROI case to get budget approved for an expensive new AI model.
- You build AI-powered workflows (email automation, content scripts, coding pipelines) and want to know where Mythos fits versus cheaper alternatives.
- You track the Claude model release cadence and want a clear timeline from Claude v1 through the Mythos preview.
SKIP IF…
- You have no existing Claude workflow — the preparation advice assumes you are already using Claude for real work.
- You are looking for hands-on Mythos demos — the model is not yet public and no live usage is shown.
The full version, fast.
Claude Mythos is a new class of AI model sitting above the Haiku/Sonnet/Opus tier — priced at roughly $150 per million tokens (115x cheaper alternatives like DeepSeek v4 Pro) and designed to run multi-day autonomous agent tasks. The preparation framework has five steps: get organizational permission, define how you will verify ROI against humans and other models, build a library of high-quality examples for your specific tasks, set hard token spending limits, and reframe the cost as R&D rather than operational overhead. The video argues that Mythos earns its premium only on tasks where you can measure output quality against a clear benchmark — and building that benchmark now, before the model ships, is what separates early adopters who win from those who waste budget.
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 – 00:51
01 · Hook
Urgency open: Mythos is days away, most people are not prepared, the video will fix that.
00:51 – 01:42
02 · Roadmap slide
Five-section overview: What is Mythos, Why it matters for business, Where to use it, How to measure it, How much it costs.
01:42 – 05:58
03 · What is Mythos?
New model class above Opus. Claude model timeline from 2023 to 2026. Currently behind closed doors — large enterprise and government only. Project Glasswing: 271 Firefox vulnerabilities found.
05:58 – 10:49
04 · Why does Mythos matter for business?
Best coding and security model, strong general-purpose research, can run for days. All tools being rebuilt for agents. Practical focus: connect your company data to Mythos.
10:49 – 13:34
05 · Where should you use Mythos?
Claude Code for apps and code, Claude Cowork for documents and HTML. Plugin ecosystem comparable to Cursor and Codex. Cowork praised for document design quality.
13:34 – 18:45
06 · How to prepare — high-quality examples
Objective tasks easier to train than subjective ones. Key prep: build a library of high-quality examples. Foreplay API case study: have Mythos scrape 300 competitor ads, filter top 100, generate UGC scripts.
18:45 – 20:34
07 · How to measure Mythos
Email drafting test: give Mythos full email context, generate 50 draft replies, count how many you send without editing. Compare to Opus and Sonnet. That ratio is your ROI proof.
20:34 – 22:33
08 · How much does Mythos cost?
Pricing table: DeepSeek v4 Pro $1.30, GPT 5.5 $35, Opus 4.6 $30, Mythos $150 per 1M tokens. 115x DeepSeek, 5x Opus. Mobile app prompt: $120-$500.
22:33 – 27:33
09 · Final advice — the five-step playbook
Get Permission. Verify ROI vs humans and other models. Experiment, Compare, Optimize. Set Limits. Be Ready to Spend More — treat it as R&D. Subscription tier prediction: $500/$2K/$5K/month plans coming.
Lines worth screenshotting.
- Claude Mythos is priced at $150 per million tokens — 5x Opus 4.6 and 115x DeepSeek v4 Pro.
- A single complex mobile app prompt with Mythos will cost $120 to $500; the same task on Opus 4.6 costs $30-$40.
- The only objective measure of whether Mythos is worth it for a given task is counting how many outputs required zero edits versus cheaper models.
- All the tools you already use — Slack, GitHub, Linear, your CRM — are being rebuilt from the ground up to be controlled by agents.
- Claude Mythos is not yet public; access is restricted to large enterprise contracts and certain government entities under Project Glasswing.
- Subjective tasks like content scripts or thumbnails are harder to train AI on because human evaluators disagree on what is good — leading to higher variance outputs.
- Giving an AI model access to find its own high-quality examples via APIs is more scalable than manually curating them.
- The email drafting test is a clean benchmark: give the model your full email context, let it draft 50 replies, count how many you send unedited.
- Subscription plans for Claude at $500, $2,000, and $5,000 per month are expected to follow the current $20/$100/$200 tiers.
- The framing that unlocks Mythos budget inside a company: position yourself as R&D, not as an operational cost center.
- Claude Cowork outperforms Codex for document and presentation tasks; Claude Code is the right surface for full-stack app development.
- Claude 3.5 Sonnet was the model that triggered the vibe coding movement; Mythos is positioned to trigger the autonomous agent movement.
- DeepSeek v4 Pro performs comparably to GPT 5.5 and Opus on many general agent tasks at 23-27x lower cost — making it the right baseline to benchmark against.
- Mythos is rumored to run tasks autonomously for days or weeks; Mythos 2 may sustain month-long autonomous work sessions.
ROI is the only question that matters with Mythos.
WHAT TO LEARN
The most expensive AI model in history earns its price only if you can measure the delta — and the time to build that measuring infrastructure is before the model ships.
- Benchmark before you buy: define what good looks like for your specific tasks now, so you have a baseline to compare Mythos against when it arrives.
- Objective tasks like working code or sent emails are easier to verify and therefore easier to justify at Mythos pricing than subjective tasks like ad copy or thumbnails.
- Giving an AI agent access to find its own high-quality examples via APIs scales further than manually curating an examples library.
- The email drafting test is a clean ROI proxy: count how many model-generated drafts you send without edits, then compare that number across models at their respective costs.
- Cost per task is the right unit, not cost per token — a model may cost $400 per mobile app prompt, but if it eliminates a week of contractor work, the math changes.
- Framing AI spend as R&D inside a company changes what budget it competes for and what success metrics apply — this is a positioning move, not just a vocabulary choice.
- Setting hard token-spend limits before running long autonomous tasks is not optional; unmonitored overnight runs have burned $10,000 in a single session for other teams.
- Cheaper models like DeepSeek v4 Pro perform comparably on many general agent tasks at 23x lower cost — Mythos is only justified for tasks where that gap measurably closes.
Terms worth knowing.
- Claude Mythos
- An unreleased Anthropic AI model representing a new class above the existing Haiku/Sonnet/Opus hierarchy, currently available only to large enterprise and government partners under restricted access.
- Project Glasswing
- Anthropic's initiative using Claude Mythos Preview to find security vulnerabilities in critical software; it identified 271 vulnerabilities in Firefox 150 during testing.
- Claude Cowork
- A Claude interface optimized for general-purpose document tasks — creating documents, presentations, spreadsheets, and HTML files — as opposed to Claude Code, which targets app development.
- Vibe coding
- A style of AI-assisted programming where the developer describes intent in natural language and the model generates working code, popularized by Claude 3.5 Sonnet in mid-2024.
- Foreplay API
- A third-party API that scrapes and surfaces competitor advertising creatives, used in this video as an example of giving an AI agent access to high-quality real-world examples at scale.
Things they pointed at.
16:00toolForeplay
11:35toolClaude Cowork
05:04linkProject Glasswing
Lines you could clip.
00:00
“Rumor on the street says that Anthropic is just days away from releasing the most powerful AI model in the entire world, Claude Mythos.”
14:00
“The key to getting the most out of Claude Opus or in the future Claude Mythos is coming up with really high quality examples for your company and then turning those into skills.”
22:10
“DeepSeek v4 Pro is 115 times less expensive than Mythos will be. We have never seen an AI model this expensive.”
25:10
“Be the guy at the company who is treated as R&D.”
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.
analogystory
00:00Rumor on the street says that Anthropic is just days away from releasing the most powerful AI model in the entire world, Claude Mythos. And my question to you is this, if Claude Mythos showed up in Claude, would you be prepared to use it to advance your business or career? If not, the time to prepare is right now.
00:20When this model releases, I wanna be able to use it right away to help my business. In this video, I'm gonna discuss the things that you need to understand about Mythos before it launches, and then how to prepare to use it in Claude desktop app and Claude code for both general agent work and AI coding work.
00:38My name is Riley Brown. Let's dive in. Alright, guys.
00:41We have a lot to cover today. We're discussing how to prepare for this new type of AI model that blows all of the other AI models that came before it out of the water. But again, there's gonna be a lot of trade offs.
00:53This model is very heavyweight, and it's gonna be very expensive. So the first thing I wanna discuss, what is Claude Mythos? And so Claude Mythos is an AI model released by Anthropic.
01:06And over the last few years, you've heard of models released by the names of Haiku, Claude Sonnet, and Claude Opus. And Claude Mythos will actually be a new class of model that is currently behind closed doors.
01:21So Anthropic is basically deemed these models not safe for the public. Some speculate that this is not actually true and that they're using this as an excuse because they just don't have enough compute to serve the model to the public, but we'll see when it releases. Right now, the only people who have access are large companies that have deals with Anthropic and then certain parts of the government, but soon, it'll be opened up to everyone.
01:47And I think it's useful to actually go through the timeline of the different Claude models because it's been a really fun to watch. And so if we go back to 2023, the first model they released was Claude v one. I don't think I've I'd hardly heard of them at this point.
01:59And I think I first heard about them with Claude two, but it wasn't that good. No one really switched over. However, 2024 was the year that everything changed.
02:08Claude released three models, Claude three Opus, Claude three Haiku, and Claude three Sonnet. These were the first Haiku, Sonnet, and Opus models. And so this was kind of in the first few months of 2024, and this didn't create the massive boom like the next model, which was Claude 3.5 SONNET.
02:28And so Claude 3.5 SONNET was the model that changed everything in the world of AI coding forever. In fact, this model was responsible for vibe coding.
02:36I still remember using it. When I first started using this model in the Quad app, I was copying and pasting code from Quad into Replit, and it was the first time that I'd ever felt vibe coding.
02:48And it was truly a spectacular model. It was added into tools like Cursor, which absolutely blew up and everyone was using Cursor in, like, from this '24 2024 to 2025 era.
03:02And then a model came out in early twenty twenty five, which was Claude three point seven Sonnen. So this model was the first, uh, model that started thinking.
03:11It was you know, at at the end of twenty twenty four, early twenty twenty five, that's when models started thinking. That's when DeepSeek released, uh, r one. And then GPT released, I think it was o three, which were these first, like, models that really thought for a long time before it, um, gave a response.
03:29And that's when AI models started getting really good at tool calls, and it was kind of a step up from 3.5.
03:37The models really started thinking for longer which was a really interesting concept at the time. And then we got Sonnet four and Opus four which were pretty good. I remember this was a decent step up, but it wasn't an insane step up.
03:50And Opus 4.1, which was released towards the end, I think it was August 2025, this is when I started realizing it was one shotting, like, really hard mobile apps. You could create a mobile app with a front end and a back end.
04:03It wasn't perfect. It didn't it wasn't automatic by any means, but you could create a mobile app with a front end and a back end, maybe a simple notes app or even like a Flappy Bird game with a leaderboard that was multiplayer.
04:15You could actually one shot these apps, you know, like 30 or 40% of the time. And so OPUS 4.1 was when I started noticing AI getting really, really good at coding.
04:26And OPUS 4.5 was the first model that kind of blew everyone's mind. This is when Andre Karpathy at the end of last year started talking about how AI is moving way faster than he he thought.
04:39He famously said, I've never felt this far behind as a programmer. And the whole world woke up, especially on Twitter. Everyone started building really cool things.
04:49And Opus 4.5, I will add, is like people started using it for general use cases. People started realizing that you could use Opus 4.5 for like, marketing use cases and, uh, more general use case.
05:02Uh, this is the model that was in OpenClaw when it was released, and OPUS 4.5 was the best model in OpenClaw. OPUS 4.5 was absolutely insane. Then we got 4.6, 4.7, and 4.8.
05:16Personally, I think it's somewhat slowed down in terms of how good it is. 4.6 was really good. Um, it was a step up from 4.5, but I do believe that four point seven and four point eight have been small improvements compared to, like, four point five and four point one.
05:31We have reached this point where it slowed down a little bit. But guys, this is a whole new ballgame.
05:38This is a whole new type of model coming soon. Right? You look at this.
05:41All of these models right here are either Opus. Right? They're either Haiku, Sonnet, or Opus.
05:46We're getting a new model. We're getting a new model within the next few weeks, and this model is right here.
05:53So this model is gonna be called Claude Mythos, and this model is bigger. It is stronger, more powerful. It works for longer than any model that came before it.
06:05However, the pricing, which we'll get to a little bit later, matches with that. This model is going to be insanely expensive.
06:12Okay, Riley. So what? We have this new class of model, Claude Mythos.
06:16Well, why is this important for business? Because if you've been following Claude Mythos, you know that it's extremely good at coding.
06:23It's gonna be the best coding model in the world likely according to Anthropic. It's also gonna be world class at locating security vulnerabilities within large repositories.
06:35In fact, they actually found 271 secondurity vulnerabilities in Firefox one fifty.
06:42And so this model is just known for finding secure security vulnerabilities and really, really good at coding. However, if you've been following the whole AI coding movement since about this time right here, you know that as AI models, as they get better at AI coding, they also get better at general purpose tasks.
07:04And this is why Claude has released two sections within their desktop app which is Claude code for AI coding and Claude co work for general purpose tasks. And we're gonna get to this in just And so this model is going to be very, very good at research.
07:22It's incredible at research and that's one of the ways it finds security vulnerabilities. It's not just natural at it. It's really good at traversing the Internet and finding different ideas, connecting different ideas and reasoning more globally and that's what's gonna make it a really general a good general purpose agent as well.
07:40And it'll be able to perform these tasks for days and even weeks at a time. I'm sure mythos one or mythos two will be able to do, uh, like a task for over a month and be effective. And so these models not only are they can go off for longer and longer, they get better at using tools.
07:58And the reason this matters for a business is it's better at using company data and integrations and controlling the different apps that you and your company use all the time. And so this would be my focus to you if you're not gonna use this for like AI coding purposes.
08:11Uh, if you wanna use this for general purpose, I highly recommend focusing on using AI with tools like email, Slack, all of the data and metrics from your company, Linear, GitHub, all of the social media platforms that you can give tracking to.
08:27This model will be able to do research longer, find connections so that you can be more successful in business. And that is how I'm thinking about Claude Mythos. How do I give it access to all of my data as a company?
08:39And I have two companies. Right? I have chorus.com which is a general purpose agent, uh, platform and then I have my own like creator business and I try to I want I wanna get to the point where I'm creating seven high quality videos like this every single week, like every single day, maybe five.
08:55And I will need AI to help me with these presentations, to help me prepare. And so that's how I'm thinking about it. And so what you need to realize is that all the tools you use are being rebuilt to be used by agents.
09:08If you watch my previous video, you saw that I was using Convex. And Convex I was using Convex inside Codex, and I was using it as the database provider, and I was creating these little things called mini apps inside Codex.
09:21Convex was a database provider prior to this AI movement, like before all of this went down. But this tool Convex, which is, again, it's a lot like Supabase or Firebase or something like that, All of these tools are being rebuilt to be used by agents.
09:38Every single huge tool that you use or every single tool in general will be rebuilt from the ground up to be used by agents. So realize that this is where the value is. Connecting the tools you've already used to AI agents, and these AI agents that you're connecting them to are getting a lot smarter.
09:55And, basically, every time we get new models, every iteration of these Claude models, they get better and better at controlling tools. Right?
10:04And these next generation AI models like Mythos and whatever comes after will be able to use these tools for better and for longer and soon fully autonomously. And we can already glimpses with the GPT 5.5 with their goal feature where it can just go off for days. And you could do the same goal feature within Claude, but like Mythos is rumored to be able to just go off for like days at a time.
10:26And so if I were to summarize this section, Claude Mythos will be smarter. It will work for longer. It'll use your own tools that you already use better for coding or for general use cases, and it will be a complete step up from models that came before it.
10:42And as we'll get to a little bit later for a lot more money. That brings me to the next question. Where should I or where should you use Mythos?
10:49So in order to use Claude Mythos, I would definitely use it within their Claude desktop app. Or if you're technical, of course, you can use it in the terminal. Right?
10:58You can open up the terminal and you can just use it here. You'll be able to use it in the terminal within Claude code.
11:07I honestly love using this, but I am starting to use the desktop app a little bit more. I am starting a little bit to like Cowork. I just really like the documents that it creates.
11:17And so, yeah, I've spent the last few days using Cowork because I'm already trying to figure out how I'm going to integrate Mythos into my, like, video prep. I think it's plenty worth it for me instead of, like, hiring someone full time which could end up costing me over a $100,000 per year. I'm definitely gonna be using Mythos and maybe I'll spend 10 to 20 k a year on doing research for these videos.
11:39In my opinion, uh, Claude Cowork is a lot better at Codex at just like spatial understanding within these documents. It can just create way better documents that you can export to PDFs.
11:52It just has better design understanding in general. And their HTML documents, which I have one right here, these are just really good.
12:01I really like the charts that they create. They create really good visuals, and I just find I really, really enjoy using Claude CoWork for these types of tasks.
12:11And then if you wanna actually do full coding tasks where you create an app with like a front end and a back end and one that requires a browser, maybe you wanna use React, you can use Claude Code. The one thing I'll say about Claude Code, if you're using it and if you're coming from Codex, I will say their in app browser is not quite as good yet.
12:29But again, this is where you do general agent tasks and this is where you're gonna do coding based tasks once Mythos is released. You can do it now, of course, and you should prepare for this.
12:39But when Mythos release, I recommend using it within Cowork or within ClaudeCode. And so if you want to do something like a landing page, web app, or create a mobile app, you're gonna use Claude Code. If you wanna create something like a document, a presentation, a spreadsheet, or like I showed you a single HTML file, you can use Claude Cowork.
12:57Both of these can be connected to your own internal tools. Whether you are in Claude Code or if you are in Cowork, you can press this customized tab and you can hit browse plugins.
13:07And just like all of the other tools like Cursor or Codex, you can very easily integrate all of your tools. It has all of them and all of these platforms have all of the integrations which is really cool.
13:20And connecting this to Cloud Code that uses Mythos is gonna be incredibly powerful. So plug ins, connectors, and skills.
13:29And one thing I wanna mention, and I think this is really, really, really important actually, is a lot of these general agent tasks are incredibly subjective. Whereas a lot of the coding tasks are more objective.
13:41Right? Does the app work? Right?
13:43That is very easy to verify. Yes. That works.
13:46Or no. It doesn't work. But the reason why some people like say that it's not as good at general purpose work, you know, something like creating, for example, like a content script or even something like a YouTube thumbnail or, you know, an ad creating all these things.
14:03These are highly subjective. And so these are actually harder to train. When they're training these models, it's harder to verify whether something's good because that's again, that's how basically how these models learn how to do it.
14:15They basically use reinforcement learning and they either give you a thumbs up or a thumbs down and they train the model to be more like the thumbs up. But if something's really subjective, it's it's really hard to do that because some people thinks it's good, some people think it's bad, and it's just this game of picking taste.
14:30And that's why these models end up kind of behaving different in the general purpose tasks is because the people who train them are just different. They're doing the thumbs up and the thumbs down differently. But there are certain things that are very objective, you know, like writing a good Python code in the back end is just like very straightforward.
14:46And so these models are gonna be very similar in that regard. And so the key to getting the most out of Claude Opus or in the future Claude Mythos is coming up with really high quality examples for your company and then turning those into skills. So if you think about it, when Mythos comes out, you are gonna have a goal and you're gonna try and use Mythos to reach this goal.
15:10Whether your goal is research or having it literally do all of your emails which I'm actually working on right now. I've actually automated a lot of my at least on one side of my business, all my emails. A high level business strategy whether you're using it for ads or content scripts for UGC or even building a full mobile app for your in person store, you absolutely need to make sure that Mythos understands what good looks like.
15:37And you might be thinking, Riley, you you're crazy. You're gonna use Mythos, the model this expensive to create content scripts for UGC? The answer is yes, first of all.
15:46Because Mythos is going to be really, really good at noticing patterns between a lot of data. And with for this example, is content scripts, there's an API that you can use. It's called foreplay.
15:58I don't know why. It's kind of a weird name. It's not like that.
16:01Uh, this is like an API that pulls ads from anyone. So you could tell Right?
16:07You could tell Mythos, hey, Mythos. Please go, uh, find all of my competitors' ads from the last week, pull them, and get the transcripts for all of them.
16:17And it could just do this. Right? And so you could, in theory, pull the transcripts from, like, 40 of your competitors.
16:25Maybe you pull, like, 300 ad 300 ad transcripts.
16:31Right? You can feed that to Mythos. Right?
16:33And you can even filter them for their longest running ads so you could get the 100 highest quality ads from your 10 competitors.
16:44Right? And so these are just high quality examples. And so part of your job with Mythos is you don't need to manually go get high quality examples.
16:55Maybe you just need to give Mythos access to go find high quality examples and give it enough context so that it can find the right ones. If it knows, if Mythos knows a lot about my company, it can go use this API called foreplay and scrape the right competitors.
17:10We can just tell Mythos to go find them, create one exactly like that. And you know what? I love creating content.
17:16I love creating high quality content. I don't wanna spend my time scripting ads manually. What if we just send Mythos off?
17:22It'll come back with really high quality scripts because Mythos is gonna be better than Opus at finding the patterns. Yes. It's gonna be five times as expensive, but more than likely it'll be worth it than me spending my time.
17:35I'd much rather spend my time doing higher leverage tasks than doing ads. Yes. We could hire people, but Mythos might just be cheaper and faster.
17:43And this may be as simple as going to Claude Cowork and saying, hey, what APIs can I give you? Um, or you could say, uh, APIs slash plugins, uh, can I give you so that you can get access to really high quality examples of top performing blank?
17:59In the previous case it was ads but as you'll see in the next video that I make I'm gonna be making a social media agent and I'm gonna give my agent access to scraping any Instagram video or any TikTok video or any YouTube video and it can scrape an entire account and find high quality examples and then it will be able to create content scripts for our company.
18:22And so what you should be focusing on is like how can you give your agent Mythos, access to high quality examples and then that's how you measure how well Mythos is doing at any given task.
18:36And so when you're testing Mythos, whether it's a coding task or it's a general purpose task, For me, I'm gonna use it a lot for doing research for videos just like this. Right? I'm gonna gather a ton of examples of scripts that I've created or research papers that I've created.
18:52I'm going to compare the output of Mythos to those examples that gives you a benchmark. How close is it to a specific example? If you just kind of randomly prompt Mythos, if you randomly prompt any new AI model and you don't have something to compare it to, it's really hard to gauge whether or not it's actually a good model for you or your company.
19:12So the first test that I will be doing with Mythos is I'm just gonna give it full access to my email. When I give it full access to my email, the first thing I'm gonna tell it to do is to analyze all of my emails, figure out what my goals are, figure out everything about me. And then it's gonna use all of that context And then every day, it's basically going to filter out all of the unimportant emails, which is really cool.
19:31And every single day, I'm gonna have it go through the 50 emails that I need to respond to, and it's just going to create a draft. For the first week, I'll have it create a draft with the eventual goal of just having it autonomously run my emails. Right?
19:44As if I could trust it to be an executive assistant. And I'm gonna see like are these emails actually emails that I would send?
19:52And for the first week, I'm just gonna measure how many of these emails are ready to be already to be sent with no edits. Right? Is it getting the email correct?
20:02That's how I verify it. If I allow Mythos to draft 50 emails for me based on all my context, how many of them did I send with no edits?
20:12Right? Maybe it's 31 with Mythos and maybe it's only 16 for Opus.
20:17Maybe I only sent 16 of the emails with Opus and with Sonnet and maybe it's only nine. And you should measure these things and this allows you to make the decision. Is Mythos worth the cost?
20:29Which brings us to the fun question of how much does Mythos cost?
20:35The answer to this question is Mythos is very, very expensive. And I actually tweeted this today because Claude Mythos is gonna be five times as expensive as Opus.
20:47And back when I was spending a lot more time with vibe code dot dev, which was a mobile app builder, we realized that with Opus 4.6, it could it was one shotting really complex apps, but it was still costing, like, sometimes $40 per prompt because it have to generate all the code and it would have to test the app then fix the code.
21:09It would have to build a back end and a front end and then it would need to deploy it fully to the App Store. And all of this would cost like $40, $30 sometimes depending on how difficult the app was to create.
21:22And so if you just think about this, like, Mythos, I guarantee you will be able to just one shot 98% of mobile apps. It'll cost about $120 to $400 when using the Mythos API.
21:35For one single prompt, it could cost up to around $500 if it's a complex app. To dig in a little bit deeper on price, it is five times as expensive as Opus, and it's like four point something times more expensive than GPT 5.5.
21:52And I think it's a fun benchmark, uh, with DeepSeek v four Pro. So DeepSeek is a little bit worse.
21:59Actually, it's reasonably worse than these models, but it's getting significantly better. And for a lot of general agent tasks, DeepSeek v four works as good as GPT 5.5 and Opus and it's a little bit faster.
22:13Except it is literally 23 times cheaper and 27 times cheaper than GPT 5.5.
22:20And if you look at Mythos, DeepSeek v four Pro is a 115 times less expensive than Mythos will be.
22:29And so we've never seen an AI model this expensive. It will be very frequent where you will spend $50 on a prompt that's not even coding related.
22:40It will be very often you hear about people spending 50, maybe up to a $100 on a single prompt, but don't be surprised when you see really high charges for mythos.
22:53So I want to leave you with some final advice as mythos comes to be in our world and the first thing I want to say is just get permission. If you work for a company really try and get permission to use this model.
23:07Really really push to get access and one way that you can get access to this or something that might help you get access to this is make sure your boss understands that you're gonna verify the return on investment and you're gonna measure how well the model does compared to other humans that they could hire and other models that they could use.
23:26You need to prove that you are making money using it. Right? You are gaining dollars per million tokens spent.
23:34That's how you can end up with the unlimited Mythos budget is if you can literally prove to your boss that you are making money or printing money using Mythos. And the next thing I'll say is you just need to constantly be experimenting, comparing, optimizing.
23:49And you can use high quality examples. You wanna make sure you're giving it context. You're giving it access to the best quality APIs and tools out there so that it has high quality examples.
24:00That's gonna be a theme over the next year as I really talk about becoming fully agent native. It's all about context and high quality examples and then set limits. You wanna avoid the nightmare scenario, which is you wake up one morning, you realize Claude was working for a lot longer than you thought, and it just burned through $10,000 worth of tokens, which happens.
24:22I've heard stories about this. You wanna set limits. But also to the other side of this, you also wanna set realistic expectations, and you're gonna you need to be ready to spend more.
24:34And so this is for people who are in companies. Right? If you're an individual, maybe you don't have the luxury of just spending a bunch of money on tokens.
24:41But at the end of the day, your company should care mostly about return on investment, and you should frame it like that. Be like, hey.
24:49I think if I use Mythos, I can help the company make a lot more money and be more successful. So you have to frame it like that and be like, hey.
24:57You just just so you know, this is gonna be more expensive, but I truly believe it's gonna be worth it in the end. And so you're you're looking for kind of this like r and d. Be the guy at the company who's treated as r and d, research and development, find ways to use this, and I genuinely believe that any good boss or manager out there will be pretty likely to give you access to Mythos.
25:20And that's kinda how I see it. Get permission. Verify that you get a return on investment compared to humans or other models.
25:26You're gonna experiment and compare and optimize constantly. You're gonna try and increase the ROI. Um, you're gonna make sure you set limits.
25:33You don't do something stupid. And then you're gonna make sure you set expectations with your team. We're gonna spend a lot more money.
25:39It's quite expensive. It's five times more than Opus and Opus and GBD 5.5 are already very expensive AI models.
25:48And the way that you're gonna minimize cost is you're gonna use it within right? You're gonna use it within one of these tools, Claude desktop and Claude co work.
25:56When you use it within the Claude apps, it's way less expensive than if you use it on an external platform. Then one thing that I will say here is be ready for $500, uh, $2,000, and $5,000 plans.
26:10I have heard some rumors that these are coming. Right now there is a $20 plan, I believe. There's also a $100 plan and a $200 per month plan when you sign up for Claude.
26:23This is the next iteration. I made a video in 2024 predicting that we would see a $100 per plan, uh, and 200 per month plan.
26:32This would be very common. I genuinely believe that the next era are the $500, $2,000, $5,000 per month plan.
26:41More and more companies are viewing this as a replacement of an employee, and so this is, you know, this is $60,000 a year. Anyway, thank you guys so much for watching.
26:50This is an incredibly exciting time in the world of AI. We're about to get a whole suite of new models to use in our AI powered super apps like Claude Code, Codex, Cursor, etcetera.
27:03The models are only gonna get better. The open models are gonna get better. Hopefully, they catch them.
27:08Hopefully, they create a model like Mythos for a fraction of the cost. It's gonna be really exciting next six months, and I'll be covering it here in great detail. So make sure to subscribe and like.
27:18I'll be moving into my actual studio next week. I'm really excited to rebuild our studio. I moved from SF to New York.
27:24Super excited to finally get my studio back. I do not like filming in this Airbnb, but I have to.
27:30I owe it to you guys. See you guys later.
The Hook
The bait, then the rug-pull.
The model does not have a release date yet, but the preparation window is already closing. In a single sitting, this breakdown maps the entire Claude lineage from v1 to the still-locked Mythos tier, prices out what it will actually cost per prompt, and hands over a five-step framework for making the ROI case before your competitors do.
Named ideas worth stealing.
22:33list
Five-Step Mythos Adoption Playbook
- Get Permission
- Verify ROI vs humans and other models
- Experiment Compare Optimize
- Set Limits
- Be Ready to Spend More
A sequential checklist for introducing Mythos into a company or solo workflow, designed to build internal credibility and avoid runaway costs.
Steal forany new expensive AI tool rollout pitch
15:03model
You to Mythos to Goal
- Define the goal
- Identify what good looks like (examples)
- Give Mythos access to find and use those examples
- Measure output against examples
A simple mental model for setting up any Mythos task: the quality of the output is bounded by the quality and specificity of the examples you provide.
Steal forany AI task brief or prompt engineering workflow
How they asked for the click.
VERBAL ASK
27:00subscribe
“Make sure to subscribe and like. I will be moving into my actual studio next week.”
Soft close woven into personal news (moving to NYC studio), low friction
MENTIONED ON CAMERA
Visual structure at a glance.
