38:05Theo - t3․gg · Essay
Why I'm Moving to Linux (For Real)
A Mac loyalist explains why agentic coding broke macOS for him -- and how a fleet of $400 Linux mini-PCs fixed it.
July 3rdThe argument in one line.
Given six weeks of early access and roughly $200,000 in inference spend, GPT-5.6 shifted from a model that needed constant steering into one that can be handed a vague, open-ended goal and left to run unsupervised for hours or days.
Read if. Skip if.
READ IF YOU ARE…
- You already run AI coding agents daily and want a read on what a frontier model can sustain over a multi-day unsupervised run.
- You're curious what a $150,000-200,000 inference budget actually buys in shipped software, not marketing benchmarks.
- You manage a fleet of your own machines or self-hosted infrastructure and want to see agentic computer-use handle a real BIOS-level recovery.
- You want a first-hand account of a new model's behavior before the official reviews and benchmark posts land.
SKIP IF…
- You want a structured review with benchmarks, pricing, or head-to-head model comparisons — the creator explicitly defers all of that to a future video.
- You're looking for a beginner explanation of what an AI coding agent is.
The full version, fast.
Six weeks before GPT-5.6's public release, Theo burned somewhere between $180,000 and $240,000 of inference across 67 projects to stress-test what the model could sustain unsupervised. The core shift: instead of babysitting a thread or resetting context when it got lost, he could hand it a vague goal and let it run for hours, sometimes over twenty, without intervention — rebuilding a mobile app natively twice, porting a Rust agent and a TypeScript compiler, fixing a broken BIOS entirely through remote computer use, and even watching it autonomously register for a third-party service mid-task. The honest caveat throughout: bigger, more finished-looking output is not the same as shippable output, and this level of spend is not a realistic monthly habit.
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 – 02:13
01 · Cold open: what I won't cover
Announces early access to GPT-5.6, gives a rough spend figure ($180K-$240K), and lists everything intentionally out of scope for this video (other model tiers, benchmarks, pricing, Fable comparison).
02:13 – 03:47
02 · Sponsor: Browserbase
Mid-roll sponsor read for Browserbase, a hosted browser/computer-use API for agents.
03:47 – 05:37
03 · Ground rules for this recap
Clarifies OpenAI didn't pay for coverage, this usage level isn't realistic for most people, and this is raw usage notes rather than a review.
05:37 – 07:18
04 · Sixty-seven projects, one fleet
Overview of working across 67 projects and setting up a fleet of machines with a centralized config and SSH-based computer use.
07:18 – 10:42
05 · Lakebed: the flagship rebuild
Deep dive into Lakebed — moving off monolithic JS files into a modular TypeScript project, building CI/CD, artifact storage, CLI login, and a whitelisting/auth system.
10:42 – 13:22
06 · T3 Code: native rewrites
Two full native mobile rewrites of T3 Code (AppKit/Swift and SwiftUI), a computer-use loop for mobile testing, and daily automated PR triage.
13:22 – 16:51
07 · The rewrite era: Rust ports
A from-scratch Rust rewrite of the Hermes agent and a Rust port of the TypeScript-Go compiler that's fast but, per a second model's review, far from feature-complete.
16:51 – 19:23
08 · FS2: the unsupervised goal run
An open-ended goal run that burned an estimated 71.2 billion tokens and autonomously registered a PlanetScale account without being asked.
19:23 – 21:20
09 · Bootable drive and a BIOS rescue
Builds a bootable Codex/Claude recovery drive, then fixes a broken BIOS/GRUB state on a remote machine entirely through a hardware KVM and computer use.
21:20 – 23:52
10 · Skatebench and side quests
Runs Skatebench (with a known cost-reporting bug), automates Prime Day deal links, builds a terminal token tracker, and ships a small 3D game called Fish Slop.
23:52 – 26:09
11 · Takeaways and what's next
Reflects on being pushed to think bigger, sets up upcoming dedicated review and comparison videos, and closes with a subscribe/bell ask.
Lines worth screenshotting.
- Six weeks of early access to GPT-5.6 cost between $180,000 and $240,000 in inference at fast-mode pricing, spread across 67 separate projects.
- The biggest behavioral shift from the prior model wasn't raw intelligence but persistence — it holds onto a task for hours without losing context or needing a reset.
- One unattended goal run alone burned an estimated 71.2 billion tokens, which would have cost around $91,000 at fast-mode pricing.
- During that same unsupervised run, the agent autonomously registered an account with PlanetScale on its own initiative to build the infrastructure it decided it needed.
- Two separate full native rewrites of a cross-platform mobile app — one in SwiftUI, one in a different native stack — were each completed in two to four hours.
- A from-scratch Rust port of the TypeScript-to-JavaScript transpiler ran up to 18 times faster than the existing Go version, but a second model's independent review put full completion at only about 5%.
- Confidently wrong answers still happen: the model hallucinated nonexistent BIOS menu options before being redirected to actually look at the screen and verify.
- The same BIOS recovery was then completed entirely through remote computer use and a hardware KVM, with no physical access to the machine.
- A benchmark tool built to score the model against others (Skatebench) had a known bug in its own cost-reporting, a reminder that automated scoring still needs a human check.
- The model was praised repeatedly for coding and infrastructure work but was still assessed as not a strong front-end/visual design model, requiring manual UI cleanup in more than one project.
- The creator explicitly treats this video as raw usage notes, not a review — pricing, benchmarks, and comparisons to other models are deferred to separate upcoming videos.
What sustained agent autonomy actually looks like in practice
WHAT TO LEARN
Six weeks and roughly $200,000 of inference spend showed that autonomy scales with cost and specificity, not raw ambition — the most valuable runs were bounded goals on a fleet the model could fully see.
04Sixty-seven projects, one fleet
- A single agent session can be handed a goal and left to run for many hours without a scripted checkpoint, as long as the task is well-scoped up front.
- Centralizing machine configuration in one repo an agent can read means every new device gets set up identically without manual repetition.
- Long-running autonomy only becomes useful once a model stops losing context mid-task; before that point, checkpoints and goal-resets are a tax you pay to keep it on track.
05Lakebed: the flagship rebuild
- Migrating a sprawling monolithic codebase into a properly modularized project is exactly the kind of large, high-verification refactor that benefits from an agent that won't wander off partway through.
- Building auth, CLI login, and access-control systems from scratch is high-stakes work; treat any agent-built version of it as a draft that still needs a full security pass.
- A working CI/CD pipeline and deploy workflow, once built and verified, can run unattended indefinitely — the upfront investment in getting it right pays off in ongoing hands-off reliability.
06T3 Code: native rewrites
- Full native rewrites of the same app in two different stacks are now fast enough to trial as an experiment rather than a multi-month bet.
- Even a two-to-four-hour rewrite that looks fully complete should be treated as an early build, not shippable; how thoroughly it's tested matters more than how finished it looks on screen.
- Automating a recurring review task, like triaging pull requests daily, is a good default use of agent time because the task is bounded, repeatable, and easy to check.
07The rewrite era: Rust ports
- A large, impressive-looking codebase is not the same as a working product; one rewrite hit 100% test-passing on its narrow slice while an independent review put it at only 5% of the way to a real replacement.
- When a rewrite is done out of curiosity rather than necessity, treat the outcome as a research finding about model capability, not as a roadmap commitment.
- Having a second model independently assess the code surfaced a more honest completion estimate than the raw line count or test-pass rate implied.
08FS2: the unsupervised goal run
- An agent left alone on a genuinely open-ended goal can make consequential real-world decisions on its own, including creating accounts on third-party services; know that risk before granting broad autonomy.
- Not checking in on a long-running goal for an extended period is itself a kind of test: it reveals whether the model can be trusted to make reasonable calls without supervision.
- A single unsupervised run burning the equivalent of tens of thousands of dollars in tokens is a reminder that autonomy and cost scale together; budget guardrails matter as much as capability.
09Bootable drive and a BIOS rescue
- Preparing a bootable recovery tool in advance, rather than improvising after a failure, turns a multi-hour outage into a plug-in-and-fix situation.
- An agent that hallucinates confidently about a system it can't fully see, like an unfamiliar BIOS, should be redirected to look and verify rather than trusted on the first answer.
- Combining remote hardware access with agentic computer use can fully repair a broken machine without physical access; worth having both pieces in place before you need them.
10Skatebench and side quests
- Automated benchmarking tools can produce plausible-looking numbers with real bugs in how they're calculated; verify the scoring logic before trusting a leaderboard.
- Agent-driven web automation is well suited to repetitive verification and link-building tasks, but generated marketing copy still needs a human pass.
- A model can be strong at code and weak at visual taste at the same time; pair it with a different tool or model for anything user-facing that needs to look good.
Terms worth knowing.
- Fast mode
- A higher-throughput, higher-cost pricing tier for running a model. The video repeatedly compares actual spend against what the same work would have cost if run in fast mode instead of the slower default tier.
- Computer use
- An AI agent capability where the model views a screen through screenshots and controls the mouse and keyboard itself, letting it operate ordinary software and websites the way a person would.
- Fleet
- The creator's term for the set of physical and virtual machines he manages from one centralized config repository, which lets an agent set up or repair any of them the same way.
- Skatebench
- The creator's own benchmark tool for scoring coding-model performance on a shared set of tasks, used here to compare GPT-5.6 against other models despite an acknowledged bug in its cost reporting.
- Reasoning tokens
- A separate token category some models generate for internal step-by-step reasoning before producing a final answer, tracked alongside input and output tokens in the usage dashboards shown throughout the video.
- Bootable recovery drive
- A USB drive preloaded with an operating system and AI coding tools so any machine on a network can be booted directly into a working, agent-ready environment for repairs.
Things they pointed at.
21:04toolSkatebench
17:10toolPlanetScale
07:40productLakebed
11:16productT3 Code
13:39toolHermes Agent
23:41productFish Slop
Lines you could clip.
00:45
“I did more inference in the past month than I had done in my life up until that point.”
03:10
“Turns out you can do a lot with $200,000 of tokens.”
16:53
“Turned out it was 71.2 billion tokens, which at the fast mode pricing would have been around $91,000.”
19:38
“It effectively said hold my beer, reboot it a few times, got into a shell in Grub, booted correctly, and then used the computer use to remote control the computer, open the terminal, and then fix the boot partitions, all autonomously.”
25:50
“It's just kind of capable. It's a workhorse and I made it work hard.”
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.
metaphoranalogystory
00:00GBT five six is finally here, and I don't think I've ever been this excited to talk about a model. There's a reason why though. It's not because it's the best thing ever.
00:06It is really great. But the reason is because I've had access for a while now, since before Fable even first dropped. I've been waiting to talk about this model for so goddamn long that I've been putting out less videos because I've been trying to clean up my computer to hide the fact that I've used it so much.
00:21And when I say so much, I mean it. Depending on how you estimate my usage, I did between a 180,000 and $240,000 of inference in my month and a half ish of access.
00:32And I got an absurd amount done in that time. It took a bit for me to digest what made Five Six so special, but once I did, I started pushing it further than ever pushed a model. I did more inference in the past month than I had done in my life up until that point.
00:46And I have so many thoughts to share now that I've done this. I need to be realistic though. If I share everything I have to say about this model, the video is going to be hours upon hours long.
00:56So instead, I'm gonna break this up into multiple different parts. And this one's particularly special. The reason this one's special is because I haven't read any of the coverage on this model yet.
01:06I have no idea what people think about it, how they're using it, or what they expect from a video like this. I'm just going off the top. I'm sharing how I used the model and how I spent such an absurd amount of money on this model to show you guys what it's actually like, or at least what it was like for me.
01:21What I'm trying to say here is don't think of this video as a review. I'll have a proper five six review coming up really soon after this. Right now, I just wanna show how I use the model and the cool things I built with it, because turns out you can do a lot with $200,000 of tokens.
01:34What that means is I won't be talking about Luna and Terra, because I didn't have access during the testing. I won't be comparing with Fable, that's his own dedicated video. I won't talk about Ultra or Max reasoning levels, because one of those is a reasoning level, the other's a new feature, neither of which are things I had access to.
01:48I won't talk about the death of Codex, even though believe me, I wanna talk about this. I have a lot of things to say. I won't talk about benchmarks or sentiment, because again, I haven't seen any of them other than what was posted with that initial announcement, and those benches didn't really make sense to me.
02:02I won't talk about pricing or usage limits, because again, I didn't have any of that when I was testing the model. And I won't talk about the fast version because it didn't exist. My fast version was 1.5 x faster, not the 750 TPS that's being promised by Cerebras.
02:16Don't even know if that's out yet. Again, I haven't checked anything today. What I have checked out though is today's sponsor.
02:21I gotta be real with you guys. I was super super skeptical of computer use and browser use when I first saw it popping up. It just didn't make sense to me.
02:29Why would an agent be faster than a human clicking buttons if it has to screenshot, manually move the mouse to a specific coordinate, click and then wait and hope it gets things right? In my early experiences with computer use stuff showed just how bad it was and I moved on.
02:42But then something important happened. GPT 5.6. There's a lot of things that make this model special but the one that impressed me the most immediately was how much better it was at using a computer, in particular using browsers.
02:54I could give it a task like set this up in a dashboard, make this change in Notion and just navigate the web for me and it would. But then I had to dedicate my whole computer to letting Codex use it. Well, I did before I set up Browserbase.
03:06Browserbase built the browser for your agents. They allow for programmatic access to the entire web through their puppeteer layer that they are hosting for you. You no longer have to deal with Playwright yourself.
03:17You don't have to set up a browser or a real computer. You just give an API to your agents and now they can do whatever they need to on the web. And this isn't just some small startup guys.
03:26Everyone from Microsoft to Lovable to Amplitude is already relying on Browserbase for running their browsers. Even DeepMind has been using them recently which is just unbelievable. But setting up browsers for your agents isn't easy.
03:37And in a world where 85% of the web's APIs are blocked behind UI, a browser becomes essential for doing real work. Give your agents the entire web at sortev.link/browserbase. Few more things I wanna clarify before we dive too deep.
03:50OpenAI has not paid me for any of this coverage. Normally, I don't do this level of usage when they give me early access. I do maybe a 100 or $200 of inference, which is covered under my paid $200 plan anyways, which still paying full price for.
04:04In fact, I have a bad feeling I might have to open up another one of those accounts soon. I also don't think that my $200,000 of inference is a realistic use case.
04:12I don't think anyone should be doing this every month. That was more me pushing the limits, experimenting, and seeing what I could get away with to an extent because they've been very tolerant of my bullshit. OpenAI has always been too tolerant to the point where others were leaking the fact they had early access without permission from OpenAI, and they didn't do anything about it.
04:30So I could have shared my thoughts earlier, but instead I chose to do what I thought I should do, which is wait until the model's officially out for everyone to access. So again, they're not seeing this. In fact, there's a good chance you're watching this before any OpenAI employee has because I just wanna share my honest thoughts and experience using this model.
04:49I was gonna do my best to not talk about Fable too much in this video. I'm gonna do a dedicated comparison later, so pardon me for ignoring the existence of Fable for a bit. If you do wanna hear all of these other things like my proper five six review, comparison with Fable, day to day usage tips, trying to maximize your utilization, all the things that make you really good at using this new model without having to spend the insane money that I spent on it, make sure you're subscribed and that you hit that bell because I have a lot of videos coming out.
05:15I have so many things I've wanted to talk about with this model that I haven't been able to and I'm finally unleashed. There's gonna be a lot of content in next few days and you're not gonna wanna miss it. I also didn't test the model in ChatGPT or 5.6 Pro at all.
05:27So if you're curious about those, I'll talk about it in the future, as I was saying. So let's go through all of the things I did with this model. I worked across 67 nice projects, and I did a lot.
05:41Some of them I did not too many prompts in, I just set off a goal and hoped for the best. Others I put a lot of back and forth effort into. If you were confused why I had so many different machines on my network recently, this is a big part of why.
05:53I wanted to push this model to its limits and really take advantage of the generous near unlimited usage they were giving me. And I did. I pushed it really, really hard.
06:02It's also worth noting that I set up my fleet using 5.6. Their SSH computer use and a centralized config that I had fan out to all my machines, it was actually really nice to work with. And that's one of my favorite things about this model.
06:14Not only is it much better at understanding intent than Five Five was, it also just crabs onto your task and doesn't let go until it's finished. This is one of my biggest issues of Five Five. It almost felt like it would get lost and then stop, or if anything bad got into its context, you'd have to go reset and make a new thread.
06:32Five Six, I did not have that experience with at all. I just will let one thread rip forever, and it's fine. Even something like slash goal went from wholly disinteresting to me to actually pretty exciting, because the model can manage that and can go for a lot longer without issues.
06:46That said, I didn't find myself reaching for slash goal too often, because if the model needed to run for more than a few hours, it just could. I would tell it to go do a long running task and it will hold on and go for hours upon hours. I've had tasks go for twenty plus hours without setting a slash goal at all.
07:03I just told it to do a big thing, and it went and it did the big thing. I guess it's time to go over all the work I did then, right? I'll start with Lakebed, because I'll be frank, I really wanna get this project out.
07:13Apparently, if I was paying fast mode prices, the work I did would have been about $30,000 in tokens, which sounds insane, but remember, you don't have to use it in fast mode, which cuts the price by almost two x, and you already get around $14 of inference right now on the $200 a month plan outside of resets using Codecs.
07:31So like, my usage here is actually kind of contained within a normal sub. And that's kinda crazy, because I set up five six on some impossible tasks and it went way further than I ever would have expected. One of the biggest things I did was move from giant monolithic JavaScript files that Five Five wrote into a properly broken down and managed Type Script project.
07:49This was a huge change that required a lot of verification and just hands on work. And it did all of it without much issue. It also helped me build a whole new CI pipeline and preview and release workflows that I'm still using to this day.
08:01I'll probably use indefinitely. It created artifact storage, railway buckets for actually managing files that users can upload, fixed a bunch of CORS stuff, also and handled a bunch of chaos with refreshing deployments with Postgres events. I built CLI login, this was way more complex than it seems to imply here.
08:16Building a relationship between my production server and the CLI that makes sense, that allows you to authenticate, Not trivial, and it did pretty much all of that by itself with just a little bit of guidance on what I was expecting UX wise. It implemented whitelist access, which is a ton of work.
08:31Probably the single biggest thing I had it do in Lakebed, because it had to centralize auth into its own system to do that. I had it fork Shoe, my auth service, and build that all in.
08:42And it works great now. I cannot wait to launch the current staging version of Lakebed as the production version because the vast majority of the changes were written by Five Six. I hardened the Capsule database layer to make the reactivity way more efficient and reliable.
08:55I designed and repeatedly implemented and reviewed the developer files contract, opaque identifiers, mine policy, all all the file object storage stuff, because I wanted to make sure that was all really, really good and reliable. The first party auth layer, which I mentioned before I did for both the white listing and for the CLI.
09:11I had it research different isolate options for the runtime that runs the code that people ship on Lakebed. Because remember, it's a cloud platform to build real apps on. I wanted to make sure if I was to build my own isolate layer that we did it right, and it explored a ton of options and gave me really good HTML plans I could read similar to what we're reading right now.
09:28It helped me model pricing and quotas, which I'll talk a lot more about my thoughts on what it did here because, spoiler, other models did this in a way I much, much preferred. It recovered a ton of PRs for me.
09:40I showcased a few loops I did in the past to show how powerful they were when I had the model spinning up threads to review PRs and whatnot. When I actually did that, all of it was five six. I had to go through and manually swap them all to look like I was using five five.
09:53I'm sorry I did that. I did not realize just how much worse five five was at sub agent stuff than five six. Five Six blasted through that.
10:01I tried the same prompts later on with Five Five, and it just got confused and lost. It also did a bunch of auditing for my open source launch for Lakebed, which I'm very, excited about. It landed a ton of code, dozens of PRs, and really helped to turn Lakebed from a weird small side project into a thing I'm excited to ship.
10:17And I'm so excited to ship it, you guys have no idea. Then of course we have t three code, the thing that I did a lot of my work with. T three code is the project Julius and I have been working on nonstop to make it easier to manage your agents on whatever machine they're on from a unified layer, whether that is the desktop app, the web app, or a phone app, which is actually very, very far along too.
10:36Early access for the mobile app coming sooner than you think. My work on t three code showcased just how thorough and thoughtful five six can be, in particular with mobile.
10:45I've had so many problems with different models on mobile, even modern really high end ones that I'm trying to not just name drop constantly. Five six is so much better at mobile that I realized I needed to push it harder, so I did. I had it rebuild the entire React Native app using AppKit and Swift, as well as SwiftUI separately.
11:02So that's two native rewrites from scratch. And I was blown away with both. They were actually fully complete end to end, functioning with all the same features that existed in the React Native version.
11:15And it did that in like two to four hours for each of them. Unbelievable. I thought it would take a day and make something entirely broken, but when you combine the capability it has with mobile with its computer use capability where I can actually spin up the simulator and check things in it, you can get full end to end insane projects like this functioning.
11:34I had a bunch of other work I did with Five Six in T three Code, but I really wanna lean into this rewrite angle because I was blown away with how much I could rebuild from scratch using the model. I have a lot more to say about these types of bold rewrites, and believe me, we'll be back to that. First, wanna go over the other things I built with t three code in 5.6.
11:52I built a full end to end computer use loop into t three code so that you could do this type of iteration with mobile and computer use with 5.6 in t three code directly. Ended up not even putting up the PR because Julius got it working even better himself, again, the same model. We implemented a collapse worked for history, which was very, very nice because we had too much stuff going on in the history.
12:12Audited the Codex app server primitives in order to try and make sub agents work better. I even got a PR up that did that. I don't know what we'll do with it.
12:19Designed a route back to top bar computer switcher as well because I wanted to try and manage all my different computers in t three code differently. Didn't end up shipping this but it was insane how far it got without anything beyond a vague prompt of what I was thinking of. Code thing is also t three code, it just didn't split correctly when I asked it to combine all these projects.
12:36I rebuilt the marketing site and got it to the stage it's in today using the model. So if you're worried about how good this model is at front end, I'm not gonna say it's perfect, but when it's steered the right way, you can get good front ends out of it. I think this looks pretty good and most of this was done with five six.
12:52So yeah, take that as you will. I also had it use sub agents to go through all of the t three code PRs and help me orchestrate all of them to figure out what to prioritize and merge and not merge. I ended up liking the results here so much that I built an automation with Hermes to do this every single day.
13:06It has helped me stay on top of PRs better than I ever have in my life. Here is the mobile native version. As I mentioned before, I rebuilt the whole thing in SwiftUI and it succeeded.
13:15Here's it talking about that. I also just had it do a bunch of profiling on the web app to find any performance issues that are worth addressing. Okay, I need to talk about the rewrite thing because it is just so cool.
13:26A combination of the model working well for these types of rewrites and the inspiration I had from a certain bun being rewritten in a certain Rust with a certain fable had me inspired to try doing similar rewrites. So I did. I started by trying to rewrite Hermes Agent in Rust with complete parody, and ended up getting pretty damn far.
13:47One of the interesting things I did when I was working on this is I had it archive all of the stuff I had done with my Hermes agent, as well as Ben with his, so I had all of our history. Then I had Five Six go through it and figure out all of the features that we need in Hermes in order to use a port of it. And it came up with the 50 or so percent of features that we actually use, and it went and implemented almost all of them.
14:10I actually got it spun up in its own separate channel and responding, creating threads, using all of skills, calling the model, and doing the things I expect it to do with a goal I ran for about a day. It works.
14:23A full from the ground up rewrite in Rust that uses like 15 megabytes of RAM instead of the giant Python version that is not the most performant thing, but that's not the goal of Hermes. I have it on a dedicated Mac Mini anyways. Who cares?
14:35I was just curious what it would look like to do this and make a very small, memory safe, reliable, minimal Hermes agent style thing that I could put in a small VM and throw wherever. And it went way better than I expected.
14:47If I was to really sit there and work on it further, I bet I could finish this without too much additional time. Not anywhere near feature complete against Hermes Agent, but to a point where I would be able to use it every day without issue. To be fair, that would have been about $13 on fast mode, probably closer to 5 to seven if it wasn't.
15:03Still pretty insane. Not as insane as this port though, and this has been very hard for me to not talk about because I had this one going for multiple days. I started rebuilding the TypeScript Go port in Rust, not because I thought it would work or I thought it would be useful, just because I was curious how it would go.
15:21I was actually able to get a 100% working TypeScript compiler, specifically the transpiler that turns the TypeScript into JavaScript, all written in Rust that is up to 18 times faster than the equivalent Go version. But I didn't get the type checking very far.
15:34I tried, and I had a lot of tests passing, more than I would have expected. The code base that it built is pretty wild.
15:41It ended up being almost 200,000 lines of Rust and a lot of tests passing. Although, admittedly, according to Fable, which I just had analyzed the code, it's a broad prototype with a tiny verified slice and it's far from shippable.
15:55It's 15 to 20% of the way to a usable tool, but only 5% of the way to a proper t s go replacement despite over a 195,000 lines of rust across the 29 crates that it built. It does have a genuinely working, deliberately narrow, no check transpiler from a five file corpus.
16:10Everything else is code that hasn't been proven, and the tests are currently failing. Currently, I have a bunch of work that was stuck in a stash, so I'm gonna clean that up and figure out how far it actually is. Obviously, I haven't had time to actually play with this, so we'll see where that ends up.
16:23I'll show more in a minute. I But wanna talk about the other things I worked on first. If saw my video where I dropped all the ideas I wished someone would build, you might remember the idea of a Dropbox like cloud that lets you sync your dev folder across machines.
16:37You probably understand why I wanted this now because I had all this work across all these computers. But I also decided to just throw out a goal and see what happens. Before this, I hadn't actually analyzed how many tokens that one goal run burned.
16:49Turned out it was 71,200,000,000, which at the fast mode pricing would have been around $91,000. I had not went to go check and see how well this works.
16:58I haven't even opened the project. I just let the goal run nearly forever. And I guess it did.
17:04I actually learned that it was using PlanetScale when one of my friends who worked there mentioned, yo, I heard you were working on FS two again. We saw you registered on PlanetScale. How's it been?
17:12I was like, wait, I did? And then I went and checked the logs and I saw it had autonomously registered for PlanetScale itself in order to build what it was trying to build.
17:20I'm actually much more excited to go try this out now because I did not think it would get as far as it did. So fingers crossed it went well. I'll be sure to check it in the near future.
17:28I have to get it actually deployed though, which will be a lot easier because the computer use is phenomenal. So I'm really, really excited to see if it can not just set up this project for me, but actually go into all the dashboards and set up all of the things it needs to really deploy it. Oh, the bootable codex drive is actually one of my favorite things I did.
17:47Since I was managing all these machines on my network, I needed to make sure I could actually set them up properly. And I was getting tired of reconfiguring whatever Linux distro over and over again, and also having to recover it when they broke. When I was ripping an SSD out of one of the machines, it didn't properly reconfigure the boot partitions on the existing SSD that I left in the machine.
18:06All of the content it needed was there, but it wasn't actually pointing at the right partitions, so it opened up an empty grub that could only run memory tests. I realized I wanted an easy bootable drive that I could use with Codex and Claud already off, so I could plug in, boot to that, and then use it to fix things. But setting up that image was taking a while.
18:24So while I had one instance of Codex configuring this flash drive for what I wanted here, which it did, and it's awesome, by the way. I highly recommend setting up a drive like this where you can boot any machine to it. Not a Mac, of course, because Macs don't boot to Linux nowadays particularly well.
18:37But all the other machines on my network, could plug in the drive, reboot, boot to the USB, connect the network, and now I have remote access to that machine with Codex and Claude already set up. Super cool. But I didn't even end up needing it for the box I was trying to repair because once I booted it, I was asking Codex for help like getting around weird BIOS things, because HP's BIOS is trash.
18:58It kept hallucinating things I could do in the BIOS. So eventually, said, okay, here, have direct access. I use my remote KVM, which lets me connect to the HDMI and a fake keyboard and mouse, so I can access the computer fully through the web, including the actual BIOS and config and all of those types of things.
19:14It lets me reboot a machine remotely, super helpful. But you also have to time a bunch of presses in order to do things like get to a specific boot menu or get around Grub. So I just figured I would have to do those parts and maybe Codex would be able to help.
19:27When five six started hallucinating those weird BIOS things that didn't actually exist, I told it to use the tab in Helium and look and see and click around to see if it could find what it thought existed. And it failed, and then it rebooted and it got into the broken grub state. So I told it that it was in that broken grub state and asked for help fixing it, And it effectively said hold my beer, reboot it a few times, got into a shell in Grub, booted correctly, and then used the computer used to remote control the computer, open the terminal, and then fix the boot partitions, all autonomously.
19:59These are the types of tasks that used to give me nightmares. I know how to do them mostly, but they require a lot of research and patience and risk tolerance.
20:08Five six just went out and did it with no additional effort. I still struggle to believe that it was actually capable of doing that. It was one of the most like, oh shit, we're really there moments I've ever had, and it made me start pushing computer use way harder.
20:22And it did all of this faster than it was able to set up that recovery boot drive, both of which were unbelievably cool, by the way. I use both a lot. I have computer use through the GL INET KVM, is still one of my favorite devices.
20:33I'll have a link in the description if you wanna grab one because it's so helpful. But I also got the boot drive, and I can use both of these to configure my network and all the machines on it. I also have a repo that I made called Fleet that organizes all of the different computers I have, what their purposes are, what's configured on them, and what my default configs are.
20:50So when I get a new machine, I just open up that repo on this one, I tell it what I got and why I wanna set it up and what I want it to do, and then it uses all of that context to configure that machine exactly how I want without me even having to touch it. It has been so nice. I keep remembering random things I had the model do that impressed me.
21:07Like, I had it overhaul Skatebench again, and it did a great job. It's way more reliable and easy for me to control now. There is still a bug with the reporting for token spend here because I know that those numbers are very, very wrong, especially when you compare it with the costs.
21:21Not a very good SkateBench score, to be fair, but that's not what we're here to talk about. I'm here to talk about how useful these are for coding. I had to build a three d version of Fish Slop that I currently can't find on my machine, so I'll have to figure out where that is later.
21:34I am sorry, I was really thinking I'd be able to grab that, but apparently when I CDN, it just doesn't exist. Oh, I had to help a bunch with Prime Day stuff.
21:42I do a Prime Day thread because I am a nerd for hunting deals, and it was able to verify a lot of the deals I found, but more importantly, create Geniuslinks, which is my link tracker that makes it easier to redirect to the country of choice for whatever country you guys are in.
21:55That's normally a very slow manual process clicking around the web. It automated the whole thing through browser use for me and responded with all of my links ready to go. It was so nice.
22:04It also wrote some copy for me that was absolute garbage, so I threw all of that out, but the rest was impressive. It built my little terminal token tracker, which I'm sure Faze can sneak a picture of somewhere in here. I show it off in the podcast, which by the way, if you wanna hear me yap about this model even more, I have a whole podcast episode where I do that too.
22:20Filmed with Ben, who also had early access, weeks before the model was officially announced. Very fun episode. I've learned a lot more about it since, but I think that was a good like, oh shit podcast episode for us where you get to see your genuine reactions.
22:33But I set up that terminal with all of the content on it and its ability to get that content from my fleet using this model. It did a very good job, but its UI was incredibly ugly and I had to rebuild what it did there with Opus four eight. Spoiler, this model isn't magically perfect at front end like other models can be.
22:51It's much more steerable and has less bad of taste by default, but it's still not a great front end model by any stretch of the imagination. I was wrong. Turns out I do actually have a version of Fish Slop here that I haven't seen before.
23:02It generated all of the art itself too. It's not great, but it's passable. UI is a bit garbage though, if I'm being honest.
23:11I wanna find the three d version though. Give me a sec to do that. And here is the three d version where itself.
23:17It made the three d environment, it made all of the rocks on the ground, it made its own textures. Some of the textures are better than when I got out of something like Grok, but not all of them. I think the rocks are actually quite a bit worse.
23:29It does get the controls a lot better though. I had to push it hard to make the controls more like a Subnautica style, like adventure where you're steering and driving your ship. It also screwed up the shoot controls a bunch and it made some of the weirdest, most hideous models for the monsters and for the fish.
23:48It's understanding of three d is, the best I can put it is strange. But, yeah, this is one of those areas where, like, the progress is real.
23:58It's still not at the point where I would want to ship this at all, but it is progressing fast enough that I'm definitely keeping my eye on these types of like three d game use cases. I did a bunch of other weird things like trying to create my own Presto alternative that was more minimal, ended up deciding on just using Pure instead directly.
24:15I had to create Claudex, which let me use my Codex off in Claudcode so I could test using Claudcode's workflows with Codex. And it was really cool. Definitely a thing I'll talk about in the near future when I start comparing the models more directly.
24:29Oh, this one's one of my favorites. I created a little app that is hosted on a Mac Mini at my office that checks which Mac addresses are on the network to have a little tracker that shows who's currently home or at the office to see who's in, which makes it really easy to know like what to expect when I walk into the office.
24:44Oh, and I had it overhaul my plan service that I use for hosting all my HTML plans including the one we're looking at now. It ended up building a bunch of really useful stuff for it. I'll be honest though, I had Fable do the majority of the recent cleanup work because I only had access to Fable for a bit.
24:58But, yeah. I know that was a lot. I'm looking at how long I've been recording for and a little scared, but I wanted to show what I've been doing with the model so that in the future when I talk about what the model is and how it works and what impressed and didn't impress me, you have the context of the things I actually did with it.
25:15Spoiler though, those future videos are gonna be even more interesting. This model got me genuinely excited to build more and think bigger. It's without question one of the bigger reasons I gave the talks I gave recently at both Cascadia JS and AI Engineer about thinking bigger, because this model pushed me to my limit.
25:31I had to push it to do bigger, harder stuff because the simple things were too simple for it. I don't have to steer it the way I did with Five Five. I just tell it go, and it goes and goes and goes without needing help managing context or not getting lost.
25:45It's just kind of capable. That said, I'm going into this without any idea how others feel and I'm curious how y'all do.
25:56Let me know in the comments and keep your eye out for the follow-up videos and again, make sure you're subbed and you hit that bell because there's gonna be a lot of content about this model coming very, very soon. Thanks for sticking through to the end, and until next time, peace nerds.
The Hook
The bait, then the rug-pull.
Before any official review, before any benchmarks, Theo spent six weeks and something like $200,000 in inference stress-testing early access to GPT-5.6 across 67 real projects — this is the raw account of what got built, broken, and rebuilt along the way.
Named ideas worth stealing.
05:42concept
The Fleet Setup
A centralized SSH-based config repo that GPT-5.6 could read and fan out to every machine on the network, letting a new box get configured identically without manual setup.
Steal foranyone managing multiple dev machines or self-hosted servers
18:02concept
Bootable agent recovery drive
A USB drive preloaded with Codex and Claude so any machine could be booted directly into a working, agent-ready state for remote repair work.
Steal forhomelab or fleet operators who need remote recovery without physical access
How they asked for the click.
VERBAL ASK
25:47subscribe
“make sure you're subbed and you hit that bell because there's gonna be a lot of content about this model coming very, very soon”
Repeated soft CTA (also at ~5:04) tied to a promise of upcoming dedicated review and comparison videos, rather than a single hard pitch.
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.
