17:06Ras Mic · Tutorial
Trust Me, You're Sleeping on Cloud Agents
A walkthrough of the four ways cloud-hosted coding agents replace manual bug verification, QA, and code review — plus the setup behind all of it.
July 15thThe argument in one line.
Codex wins on overall subscription value not because its harness is the best of the four — Cursor's is — but because OpenAI subsidizes usage the hardest and its models are the cheapest per unit of coding output.
Read if. Skip if.
READ IF YOU ARE…
- You already pay for one or more $20-200/month AI coding subscriptions and want to know which one is worth keeping.
- You're deciding between Codex, Claude Code, Cursor, and Devin and weighing raw output quality against cost.
- You want a concrete script for convincing your employer to expense an AI subscription.
SKIP IF…
- You're looking for a deep tutorial on any single tool rather than a cross-tool cost comparison.
- You don't use a paid AI coding subscription and have no near-term plan to.
The full version, fast.
After viewers complained he never talks about cost, the creator ranks four AI coding subscriptions. Codex, Claude Code, and Cursor each run their own frontier model (GPT, Claude, and Grok respectively), while Devin mixes in its own SWE models; Claude's Fable 5 is the smarter model, but GPT 5.6 Sol is cheaper and nearly as capable for most coding work. All four subsidize usage, but Codex resets limits most aggressively. Cursor's harness scores highest for raw output (9/10) against Codex's 8/10 and Claude Code's 7/10, yet Codex is named the single best subscription once model cost and subsidies are weighed in. The video closes with Codex power-user tips and a script for getting an employer to cover the bill.
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 – 01:40
01 · Cold open
Comments called him out for never talking about cost; he agrees, tells the Pizza Iola story to prove he knows how to squeeze a deal, then names the four contenders: Codex, Claude Code, Cursor, and Devin.
01:40 – 03:45
02 · Models: Fable 5 vs GPT 5.6 Sol
Breaks down which model powers each harness (GPT for Codex, Claude for Claude Code, Grok/all models for Cursor, SWE/all models for Devin) and calls Fable 5 the smarter model but GPT 5.6 Sol the better all-around pick for cost and general programming.
03:45 – 05:33
03 · Subsidies
All four products subsidize usage, but Codex/OpenAI resets and extends usage the most aggressively — he hits Claude Code's limits far faster in daily use.
05:33 – 07:05
04 · Sponsor: Blacksmith
Ad read for Blacksmith, a CI runner that cut his team's build times from roughly 8-10 minutes to under 2.5 minutes with a one-line config change, plus an agent (CodeSmith) that opened the PR for him.
07:05 – 10:12
05 · Cost benchmark and harness scorecard
Uses CursorBench to show GPT 5.6 Sol beating Fable 5 on High while costing less ($5.69 vs $8.77), then reveals harness scores: Codex 8/10, Claude Code 7/10 (saved by its Workflows sub-agent feature), Cursor 9/10 for best raw output — but declares Codex the overall winner once model, subsidy, and cost are combined.
10:12 – 12:13
06 · Pricing across all four
Walks the pricing ladders: Codex $100/$200 (ChatGPT Pro 5x/20x), Claude $20/$100/$200, Cursor $20/$60/$200 (with new 'generous limits' for Grok and Composer), and Devin $0/$20/$200 — all converging near the same $200 ceiling.
12:13 – 15:50
07 · Power-use: plugins and in-app browser
Recommends exploring Codex's plugin store, singling out Convex as his favorite for spinning up real-time backends by tagging it directly in chat, then demos the in-app browser running and testing a local app, including a standing '/goal' prompt to hit a Lighthouse score of 100.
15:50 – 17:20
08 · Power-use: computer use
Demos Codex's computer-use agent driving a real browser autonomously, shown filling out a RoboForm test form field by field and stopping short of entering a credit card number.
17:20 – 18:58
09 · Thread maxing and model usage
Advises spinning up new threads instead of maxing out context on one long thread, and lays out a model-effort mapping: x-high for coding, light-to-medium for knowledge work, and a warning to avoid Ultra, which he calls token-hungry and low quality.
18:58 – 21:24
10 · The free-subscription hack
Closes with the promised trick: pitch your employer to cover the $200/month subscription as a productivity investment and tax write-off, including a scripted pitch aimed straight at the viewer's boss.
Lines worth screenshotting.
- Codex hits its usage limits far less often than Claude Code, according to a creator who pays for and daily-drives all four major subscriptions.
- On CursorBench, GPT 5.6 Sol on Max beat Claude's Fable 5 on High while costing $5.69 per task versus $8.77.
- GPT 5.6 Sol run at extra-high effort costs under $4 per task and performs roughly as well as Fable 5 on Medium.
- Cursor's harness scored highest in hands-on review at 9/10, ahead of Codex's 8/10 and Claude Code's 7/10 — yet Cursor still wasn't the recommended subscription once cost was factored in.
- Claude Code's multi-agent 'Workflows' feature, run via Ultra Code effort at extra-high reasoning, is described as the one thing keeping it competitive against Codex and Cursor.
- Codex's 'Ultra' reasoning tier is explicitly discouraged for regular use — it burns through token usage fast for marginal quality gains.
- Every top-tier plan across Codex, Claude, Cursor, and Devin converges on roughly $200/month regardless of provider.
- Codex's in-app browser can run a local app, test it, and execute a standing prompt like 'reach a Lighthouse score of 100' unsupervised for hours.
- Codex's computer-use agent refused to enter a credit card number when autonomously filling out a test web form.
- Pitching an AI subscription to an employer as a $200/month, tax-deductible productivity investment is offered as a legitimate way to get it comped.
- Grok 4.5 inside Cursor is priced competitively and considered close to — or arguably better than — Claude Opus 4.5, despite getting the least screen time in the comparison.
Cheaper models and bigger subsidies beat a better harness
COST BREAKDOWN
Across every metric that isn't raw output quality, Codex's cheaper models, heavier subsidies, and power-user features make it the best AI coding subscription for anyone watching their budget.
02Models: Fable 5 vs GPT 5.6 Sol
- Claude's Fable 5 is the more powerful model for complex, large-codebase work, but GPT 5.6 Sol handles most day-to-day programming just as well for less money.
- Codex, Claude Code, Cursor, and Devin each wrap a different model line — GPT, Claude, Grok, and SWE respectively — so picking a harness also means picking a model family.
- Cursor is the outlier: it serves every major model plus its own Grok-based option, rather than locking users into one provider's models.
03Subsidies
- Every major AI coding subscription is subsidized to some degree — usage resets, extended free periods, and below-cost compute are the norm, not the exception.
- OpenAI/Codex was observed resetting and extending usage far more aggressively than Anthropic, based on tracking how often a public figure at Codex tweets about resets.
- In daily heavy use, Claude Code's usage limits were hit noticeably faster than Codex's, despite both being subsidized.
05Cost benchmark and harness scorecard
- On a published cost-per-task benchmark, GPT 5.6 Sol beat Claude's Fable 5 on quality while costing roughly a third less per task ($5.69 vs $8.77).
- Running GPT 5.6 Sol at extra-high effort delivered results comparable to Fable 5 on Medium for under $4 per task.
- Hands-on harness scores landed Cursor highest (9/10) for raw output quality, ahead of Codex (8/10) and Claude Code (7/10).
- Even with the lower harness score, Codex was still named the overall winner once model cost and subsidy generosity were weighed against Cursor's better output.
06Pricing across all four
- Every provider's top tier — Codex, Claude, Cursor, and Devin — lands close to $200/month, so price alone won't differentiate them; value-per-dollar does.
- Codex's $200/month tier (20x the base rate) was recommended as the best bang-for-buck option for anyone already spending real money on AI subscriptions.
- Cursor's move to add 'generous limits' for Grok and its own Composer model could shift the value calculation in Cursor's favor going forward.
07Power-use: plugins and in-app browser
- Codex's plugin store extends the base tool with integrations like Notion, Google Calendar, and Convex — worth browsing before assuming the stock app is all you get.
- Tagging a plugin directly in a Codex chat (e.g. 'build me a to-do app' with Convex tagged) lets the agent scaffold a real, working backend in one pass.
- Codex's in-app browser can run and test a local app inside the harness itself, including standing prompts like 'reach a Lighthouse score of 100' that run unattended for hours.
08Power-use: computer use
- The computer-use agent can operate a browser autonomously enough to fill out real web forms, but it stopped short of entering payment card details on its own.
- Using the in-app browser and computer-use loop to have the agent self-test features (e.g. verify login/auth works after a fix) closes the loop without manual re-checking.
09Thread maxing and model usage
- Splitting work into fresh threads instead of continuing one long thread avoids maxed-out context windows and the sloppier answers that come with them.
- Match model effort to task type: extra-high for coding, light-to-medium for general knowledge work, and avoid Ultra — it burns tokens fast for little quality gain.
10The free-subscription hack
- A $200/month AI subscription can be framed to an employer as a tax-deductible productivity investment rather than a personal cost, especially in a role that already touches engineering or knowledge work.
- Citing a specific coworker whose company already subsidizes their AI usage in exchange for measurable productivity gains makes the employer pitch concrete instead of hypothetical.
Terms worth knowing.
- CursorBench
- A benchmark Cursor publishes that plots AI coding model performance against average cost per coding task, used to compare models on a price-to-quality curve.
- Harness
- The application or interface (Codex, Claude Code, Cursor, Devin) that wraps an AI model with tools, memory, and workflow — distinct from the underlying model itself.
- Workflows (Claude Code)
- Claude Code's multi-agent, sub-agent orchestration mode, enabled by setting effort to 'Ultra Code,' which runs extra-high reasoning across coordinated sub-agents.
- Thread maxing
- Splitting a long AI coding session into multiple fresh threads instead of continuing one thread indefinitely, to avoid context-window overload and degraded answers.
- Computer use
- An AI agent capability that lets a model directly control a browser or desktop — clicking, typing, navigating — rather than only generating text or code.
- Subsidy (usage)
- A provider absorbing more compute cost than a subscription price technically covers, e.g. through usage resets or extended free periods, effectively giving users more tokens than they paid for.
Things they pointed at.
Lines you could clip.
00:17
“I used to go to Pizza Iola in Toronto for $7. You got three slices... and a dip and a drink for $7.”
10:09
“It's not surprising that the winner is Codex.”
12:40
“The Claude desktop app is weak. Terrible. I am not a fan at all.”
18:24
“Don't use ultra... It's terrible. It's bad. It guzzles tokens.”
19:26
“Ask your workplace... that AI will make you more productive and that they should pay or subsidize your subscription.”
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.
metaphorstory
00:00AI is incredible. Incredibly expensive that is depending on what models you're using. And a lot of y'all made it clear to me in the last video when I talked about Cloud Agents where you guys were like, you never talk about cost.
00:10And I'm not gonna lie, the way some of you are badgering me in the comments, this is how you sound.
00:17Yeah. Kinda broke my heart. I'm having too much fun.
00:20But in all seriousness, I humbly admit, you guys are right. I never do talk about cost. I am in a privileged position where I do use all the subscriptions and this allows me to make content so you guys know what is the best harness model, etcetera.
00:33But I'm gonna be honest with you. If anyone knows how to get a deal, if anyone knows how to squeeze dollars out of a tool, it's me. Some of you don't know where I came from.
00:42I used to go to Pizza Iola in Toronto for $7. You got three slices. Sorry.
00:47Three slices and a dip and a drink for $7. If anyone knows how to squeeze a deal, it's me. That's why in today's video, we're gonna talk about codex, we're gonna talk about Cloud Code, we're gonna talk about cursor, we're gonna talk about Devon, and I'm going to tell you which model plus harness makes the most sense if you're tied on a budget.
01:06Which one can you squeeze absolutely as many tokens as you can while not spending as much money as yours truly does? If that's information you're looking for, you're going to enjoy this video. And finally, at the end, I'm gonna share with you a secret strategy no one should know, on how you can get a subscription for free.
01:23I'm saving that for the end. Sit back, relax, and Here comes the money. Here we go.
01:29Money talk. Here comes the money. I'm having a little too much fun.
01:32Let's get into it. Which subscription slash should you pick? We're gonna compare between Codex, Cloud Code, Cursor, and Devon.
01:38First things first, we're going to compare models. When it comes to models, Codex and Cloud Code have their own in house models.
01:45Right? Codex having the GPT models, Cloud Code having the Cloud models. Cursor is in a special position where they serve all models, and they have their own model, the Grok model since they've merged with SpaceX.
01:56It's SpaceX AI. It's kind of a tongue twister. And then you have Devon, which is pretty much the same thing.
02:01It has all the models. Right? It also has its own suite of models named SWE.
02:05That being said, the question really when it comes to picking which model is the best between these harnesses is a simple question between Fable five versus GPT 5.6 Soul. Now let me know if you want a dedicated video on this.
02:18But to simplify this, Fable five is the better model. You just don't need it all the time. GPT five six Soul is great enough for most tasks, even the complex programming tasks, but Fable five, especially for really large code bases or really complicated features and things, it is just a workhorse.
02:38It is amazing. It is smart. It writes really, really good code.
02:43I love Fable five, and to be honest with you, I prefer Fable five over g p t 5.6 sole any day of the week, but I don't need Fable five for everything. Right? GBT 5.6 Soul is more of a general model.
02:54It's great at programming. It's great at a lot of knowledge work stuff. Right?
02:58So if we were comparing sheer intelligence and power, it'd be Fable five. But in this case, for cost perspective and even for usage perspective, I would say g v d 5.6 is the better model.
03:10Now we do have Grok and we also have SWE. I will be honest, I haven't used the SWE models as much, so I can't really speak on them. But Grok 4.5 is an excellent model.
03:20It is great for the price and it is comparable to Opus four eight, dare I say even better, but it's nowhere near GPT 5.6 Soul, and I love cursor as a harness. So when it comes to the model category, I will say Codex takes this one because the GPT models overall are more superior than the rest of the models.
03:39Again, overall, If we're talking strictly programming, I'd go with Fable any day of the week. Number two, subsidies.
03:46And congratulations. All of these companies, all of these products subsidize the AI usage.
03:53Codex, the AI usage is subsidized. Cloud Code, the AI usage is subsidized. Same with Cursor.
03:58Now that they have Grok, now that they have SpaceX AI, now that they have the colossal giant supercomputers, the Grok models, they've been getting resets.
04:06There's extended usage. There have been periods where it's free. And even with Devon, using the SWE models is free.
04:13So there are subsidies all around. The question is who subsidizes more and whose subsidies are better? And there's no question, there's no one who subsidizes more than OpenAI, than Codex.
04:25I searched Thibault, who's one of the leads at Codex, and I just searched reset. I want you to see how many tweets contain the word reset.
04:34This guy is spamming resets left, right, and center. And from a cost perspective, ladies and gents, that's what we need. We need subsidies.
04:42We need resets. Now some of you might say Anthropic is subsidizing Fable and the other models as well. Yes.
04:48But when it comes to using ClaudeCode, I hit my limits a lot faster than I do with Codex. So when it comes to subsidies, no one does it better than Codex.
04:56I mean, you are getting a lot more compute than you are paying for, and it's just . You know what else is a great bang for your buck? Today's sponsor.
05:05Thanks to AI, my team and I have been shipping a lot. I mean a lot. Features are being built left and right like we've never seen before.
05:12We're even migrating code bases from TypeScript all the way to Rust. It really is a beautiful time to be alive. But here's a sad unfortunate truth.
05:21There is a big bottleneck that slows us down and that is our CI pipeline. If only there was something that can help us fix this so that we can continue shipping even faster. Gents and the three ladies that watches, I'm proud to tell you Blacksmith has solved my problems.
05:36Now you might think setting this up might take forever. I kid you not. It is literally one line change.
05:42That one line change will lead to two x faster hardware, four times faster cache downloads, and 40 times faster Docker builds. And I'm not just saying to say that. Our build time, our CI pipeline, it went from, like, ten minutes, eight minutes, seven minutes on some things to literally everything is under two minutes and a half.
06:01I love Blacksmith because it has saved me and my team time. We can ship and view our changes faster in production. We also get some funky analytics.
06:10So far, it's only been a couple of days, and we already have a thousand plus jobs ran through Blacksmith. And remember earlier, I said all you have to do is change one line to use Blacksmith. Well, I didn't even make that change.
06:20They have an agent called CodeSmith that created a PR that did the one line change that I obviously myself merged that made our CI pipeline faster. You can also use CodeSmith to auto fix any failing CI issues or to even review feedback on PRs.
06:37It doesn't get any better than this. It doesn't get any faster than this. Blacksmith is the hero for all our CI problems.
06:43Make sure to check him out. The link is in the description down below. Let's get back to the video.
06:47So model wise, we can agree that the GPT models are a better bank for your buck. Again, I'm not talking about sheer power, but if you take cost into consideration, here's what I mean. You could see that Fable, especially on Macs, is the most intelligent, the most powerful model.
07:01Looking at cursor bench, you could look at Fable five max. It's super powerful, almost 70% on the cursor bench.
07:09And then you look at g p t five six sole on max, it is better than fable five on high, but it is much cheaper. $5.69 compared to $8.77.
07:19You're getting extra high and high for literally under $4, but just as powerful as Fable five medium.
07:27So when I think about what I use these models for, programming and a lot of knowledge work, then aggregating all of that information with the cost perspective, it is no doubt that the GBT models are better for Moola.
07:39We're also getting heavily subsidized. Now finally, harness.
07:43Which harness is the best? And I have mixed feelings and even these ratings take it with a grain of salt. First things first, I'll talk about Devon.
07:50Devon is more of a cloud product. Right? I use cloud agents a lot, especially for, like, big refactors or big feature changes.
07:59I find cloud agents to be a great environment to develop stuff like this, to work on stuff like this, especially for the the ability that the agent can not only has access to a computer, but it can record it finding a bug, it reproducing a bug, it suggesting a fix, working on the fix, deploying the fix, and testing the fix.
08:17All this stuff is automated for me for certain apps I'm working on, and Cloud Agents make that really, really fun. That out the way, if you're going to be using a daily driver, it's probably gonna be Codex CloudCoder Cursor. Now in terms of harness, no doubt I find the Cursor harness to be the best.
08:32The output is better. I like the user experience. I like that I can split between local and cloud, and I can also SSH to my own machine.
08:41The reason this is a seven and not a five is because Claude codes workflow. And if you're not familiar with workflow, I'll just show you real quick.
08:50Workflow is Claude codes, like multi agent, sub agent sort of architecture.
08:58And, basically, the way you set it up is you change effort to ultra code, and it uses x high, and then it uses workflows. This is what's kept ClawCode at a seven, maybe even a eight for me.
09:11Right? But in terms of user experience, I'm not a fan of the app. The TUI sometimes works, sometimes it poops.
09:18I'm really not a fan of the harness, and I find that Fable five actually I get better results on cursor than I do with Cloud Code. That being said, workflows are elite, but we don't get subsidized that often. When we think about harnesses, Codex is up there.
09:35Codex's harness is amazing. And the reason why it's amazing, you get a couple of things. Right?
09:42The threading system is great. The plug ins, there's tons of them. You have a built in browser.
09:47There's a lot of cool things harness wise that you can do with codecs that I will get into in just a second, but it's not no cursor. I still find the cursor harness to be better in terms of results.
10:00But when we look at costs, when we look at models, when we look at subsidies, when we look at the overall picture, ladies and gents, it's not surprising that the winner is Codex.
10:12If there is one subscription you should get, you can't really afford having all of them or two of them. Codex is the one.
10:18Now let's talk about pricing. These are Canadian dollars. Canadian dollars are not a real currency.
10:22So in USD, 250 CAD is $200. So just like do that conversion in your mind. So with ChadGBT, with Codex, you can get the pro plan five x for a $100 a month or the 20 x for $200 a month.
10:36I will say $200 is a lot of money. But if you can cancel the Netflix, the Disney plus, the whatever subscriptions you got, I know there's a lot of nonsense subscriptions you have. The $200 a month subscription, if you're watching this video, you watch these type of videos, is a great bang for your buck.
10:51You get much more compute than you are paying for. With Claude, it's the same thing. Right?
10:55You have $0, $20, a $100, $200. $200 is the best bang for your buck. Again, you're paying more, but you're getting a lot more compute.
11:04If I had to pick one, I'm on a budget, I would go for codex. And Cursor's pricing is basically the same thing. This is actually USD, $20 for pro, $60 for pro plus, and $200 for ultra.
11:16Now the Cursor subscription is going to start getting interesting even for people who are on a budget because now, guess what it says over here, generous limits for grok and composer. So if you want a harness that gives you, you know, a state of the art model that is pretty cheap and affordable and subsidized, you get that with Cursor.
11:35But if you want access to other models, you also get that with Cursor. So this suggestion, this win might change very, very soon.
11:43Pricing with Devon, same thing, $02,200. They're all pretty much the same when it comes to pricing. It's just a matter who gives you the more compute, more tokens, more juice for your subscription.
11:55Codex is the winner. Now that you know that Codex is the best harness for your buck, the best choice of model for your moolah, how to power use?
12:04Now I'm gonna do a dedicated video on this. Let me know in the comments down below if you wanna see it. But there are a couple things I want you to use and check out.
12:11First things first, when it comes to codecs, do not sleep on the plugins. There are some incredible plugins here. First and foremost, the computer use.
12:19I'll talk about that in a second. You know, you can run spreadsheets and presentations. And here's why I also picked codecs from a harness perspective.
12:25You can use Codex like your personal agent, a hope Open Claw and an Ermies agent, and it's low key better than the Claw desktop app. The Claw desktop app is weak.
12:36Terrible. I am not a fan at all. I'll try to use it every now and then.
12:39I am just not a fan. I find myself leaving almost instantly. The Codex app though, even though a lot of people weren't a fan of how they merged with the Chad GBT, all that aside, it's also a great place for you to a lot of knowledge work.
12:53My friend Riley, who I've been on his channel a couple times, Riley Brown, he has a lot of tutorials on how to do marketing and all these type of things on the Codex app itself. So don't sleep on the plugins. There's a lot of interesting plugins here.
13:06I have Notion and Google Calendar installed. My favorite plugin and I'm I'm about to I'm about to I am about to I'm about to crash off for a second.
13:18My favorite plugin is Convex. Now many of you are probably gonna say, oh, it's because you work at CodeX. You're killing Convex because you work at Convex.
13:24First and foremost, go on my GitHub. Go on my GitHub. Every app I build, I use Convex.
13:28And the reason why I use Convex is the agents are good with it. It's really fantastic. It scales well.
13:34It's all code, and I just have a great experience with it. And because I liked the product, believe it or not, they used to sponsor the channel. I joined the company.
13:41So that doesn't change the fact that I can't talk about I can talk about them. And take my word for it. Look at GitHub.
13:47Every project I build with uses Convex. And I can just install the plug in, and I can open a new chat, and I can tag Convex, and I can literally say, build me a to do app. And then instantly, real time, out of the box, you can use Convex's component system.
14:00It's really, really great. I know some of you are gonna say I'm shilling. I'm really just giving you the sauce.
14:05So plug ins. Check out the plug in store. Look at everything.
14:08Install what you think is beneficial to you. Number two is the in app browser. The in app browser is clutch.
14:13First and foremost, they basically took Atlas and they shoved it into the app. So you can see this is a full on browser that I can use normally.
14:22I can go on x.com. I can like tweets, whatever whatever. But the main way that I use it is I will tell I'll give you guys an example.
14:30Can you run this app locally in the Codex browser? So what I do is I get the agent to run the app and to show me the app on the browser.
14:41And then whenever I've built a feature or there's a bug or whatever there is, I will tell it to use the browser to test, to debug, and I've even done this is actually one of my favorite prompts. Again, this is another video. Let me know if you want me to go in details, but one of my favorite prompts lately with Codex has been slash goal.
14:59So this is a goal. Optimize the site to be blazing fast and reach a lighthouse score of a 100.
15:10Basically, I want my site to be blazing fast, and for it to get a lighthouse score of a 100, it also has to be well optimized, mobile optimized. And I'll tell the agent to do that, and it'll work on it for, like, an hour, two hours. And then I get a message, a notification that it's done.
15:24So you could see I have it running in the browser. Now I could do something like this. You have the app running in the browser.
15:30Can you use computer use and try to log in the app? So I'm a hit enter, and this is going to run once it's finished its initial process. You could see the agent's cursor right here.
15:40It's an action. It's going to basically test out login because I told it to do so. That's what I'm saying.
15:45Not only am I getting subsidized model, but the harness has a built in browser, and it does computer use, and it can do stuff like this. So it's testing out auth for me.
15:56If auth breaks, guess what? I just tell it fix the auth. And the way you're going to confirm that the auth is fixed is by testing, logging in.
16:04And if it works, let me know. And I kinda spoiled the third one, but computer use. Computer use on codecs is incredible.
16:11It is by far the best computer use I've ever seen and so good that I've used it to do random things on my computer. So just to test out the computers, I found out this random form and I'm going to say I typed computer. Can you fill out this form I have on my aside browser?
16:29I'll give you the link. It's already open on the aside browser. Just open the browser and fill out the form for me.
16:35And I'm just gonna paste the link, and I'm gonna hit enter. And, hopefully, we can see some action.
16:42I'm just gonna put this side by side so you could see it. So you can see I don't even need to put it side by side because Codex is showing me right here its browser view. Look.
16:53Look how fast that is. Look how fast that is. Look how fast that is.
16:59How fast that is. No. This is incredible.
17:02This is incredible.
17:06Hey. Hey. Hey.
17:07Hey. Hey. Hey.
17:08Hey. No. No.
17:09No. No credit card. It tried to fill out my credit card details.
17:12You're just going to have to take my word that the computer use agent is fantastic. And finally, I would say to ThreadMax. Right?
17:20Do not make the mistake of staying on one thread and just building on said thread.
17:27The one thing that you could do with codecs, and again, if you want me to do a dedicated video on this, let me know in the comments down below. You can start threads within a thread. You could tell the agent, okay.
17:37Like, you can, let's say, plan with one agent, go back and forth, maybe build out a prototype of what you wanna build, and then you could tell the agent, spin up a new thread with this plan, with this back and forth we've had in context. Right? Or you could tell us spin five threads.
17:50Whatever it is, just use threads. Do not continue to build on one thread. Max out the context window, and now you have sloppy answers.
17:58Thread max to the best of your abilities. Now this wouldn't be great advice if I didn't tell you the model usage. Which model should you use?
18:06I only use sol. Whenever I use codecs, I only use Sol.
18:10And in terms of what level do I use, when I'm coding, it's x high. Don't use Ultra. Ultra is basically the equivalent of Claude Codes workflows Ultra Code, and Ultra is pretty bad.
18:21I'm not gonna lie. It's terrible. It's bad.
18:23. If you wanna burn through your usage, you would use ultra. Don't use ultra.
18:28High is kind of, in my opinion, useless because when I'm doing anything with code, I'm gonna use extra high. And when I'm doing anything with knowledge work, it's between light and medium. Right?
18:37So this is the model usage that I would recommend when using codecs. Anything other than coding, if it's not super, super, super, super difficult, which it probably won't be, I wouldn't use extra high. Medium to light is just good enough.
18:51Now you might be saying, what about Terra and Luna? And my answer is, I don't have enough time to test all these different configurations. I get the best model, and I use the parameters within the best I haven't really used Terra Luna to say anything about it.
19:04Now finally, if you remember I told you at the beginning of the video, I'm gonna share with you a secret way on how to get a subscription for free. I hope you're ready. Because this one alright.
19:15I'm sorry. I just got a soundboard and I'm really having fun with it. But in all seriousness, the way to do this is to ask your workplace.
19:21If you're, you know, working white collar, you're working in the office, whatever it is, show your workplace that, you know, AI will make you more productive and that they should pay or subsidize your subscription.
19:32A friend of mine, Nick, who you know, me and Nick go to church together. He works in real estate and he uses AI a lot and the company pays for him, but he's built systems and workflows and the company's a lot more productive. He's a lot more productive.
19:44He's getting free subscription and he gets to use the greatest and latest models. It's a win win for everyone. Now you might be like, I don't know what to say to my boss.
19:53I'm going to talk directly to your boss. So from this point on, you can send it to them, then we'll end the video. Hello, boss.
19:59Uh, one of your amazing employees sent this video to you. They've been watching my channel.
20:05If you're not familiar with my content, I talk about AI workflows and how to be super productive in work, whether it be engineering or knowledge work. And they seem to find a lot of the tools I've been sharing very helpful, especially more helpful in your organization was implemented.
20:20The only thing is these tools cost money. And as a business, especially in 2026, I'm sure you're aware that getting into AI, making your company used to these tools and workflow so you guys can be productive and not be left behind is super, super important.
20:35And you have an employee who's ready to take that initiative. I think it would be fair if the company subsidized their subscription.
20:42It's only $200 a month, tax write off, very easy for the business, but I think the return on investment's gonna be a lot higher. I think you should consider getting your employees the subscription that they're asking for.
20:55They're gonna be super productive. And if they're not productive, you know, just cancel it. Don't ever say I don't take care of you guys.
21:01Don't ever say I don't look out for you guys. That's pretty much it for this video. I hope you enjoyed it.
21:05I hope this covers the best way to maximize AI usage while being on a budget. Let me know down in the comments down below what videos you'd like to see.
21:14I have a couple more cooking soon, so make sure you're subscribed. We're about to hit a 100 k, so I'm super excited for that. You've been awesome.
21:20My name is Ross. Thank you so much for watching this video. I'll see you in the next one.
21:23Peace.
The Hook
The bait, then the rug-pull.
Viewers kept telling him he never talks about cost — so the creator who runs every major AI coding subscription at once breaks down exactly what you're paying for across Codex, Claude Code, Cursor, and Devin, and which one actually deserves your $200 a month.
Named ideas worth stealing.
01:27model
Which Subscription/Harness to Pick scorecard
- Models
- Subsidies
- Harness review
A 3-row scorecard comparing Codex, Claude Code, Cursor, and Devin across which models they run, whether usage is subsidized, and a 1-10 hands-on harness score.
Steal forany head-to-head tool comparison video or buyer's guide
18:00concept
Model effort ladder
- light
- medium
- high
- x-high
- ultra
Maps Codex's reasoning-effort settings to task type: knowledge work sits toward light/medium, coding sits toward x-high, and Ultra is flagged as wasteful for almost everything.
Steal forany AI-tool tutorial explaining how to tune model effort/reasoning settings for cost control
19:26concept
Employer AI-subsidy pitch
A scripted pitch framing a $200/month AI subscription as a tax-deductible productivity investment the employer should cover rather than a personal expense.
Steal forany prosumer SaaS tool positioning itself as expensable to an employer
How they asked for the click.
VERBAL ASK
05:33product
“Thanks to AI, my team and I have been shipping a lot... Blacksmith has solved my problems... The link is in the description down below.”
Mid-roll sponsor read woven into the cost-comparison narrative (ties CI cost savings to the video's cost theme), delivered straight to camera with on-screen GitHub PR proof of the one-line fix.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
