16:11Matthew Berman · Tutorial
7 INSANE loops you need to try right now
A 16-minute walkthrough of 7 copy-paste AI coding agent loops — each one runs autonomously until a defined goal is met.
June 19thThe argument in one line.
GPT-5.6 squeezes the last available gains out of the existing GPT-5 training run into a cheaper, more capable model, while a rawer, less-optimized rival model already reasons better and has more headroom to improve.
Read if. Skip if.
READ IF YOU ARE…
- You're choosing between frontier models for an agentic coding or automation workflow and care about real cost-per-finished-task, not sticker price.
- You use tools like Codex or similar agentic coders and want to know what multi-day unsupervised runs actually produce.
- You're weighing a cheaper, highly optimized model against a pricier, newer one with a higher reasoning ceiling.
SKIP IF…
- You don't use agentic coding tools or API-billed models and only care about the consumer chat experience.
- You're not interested in comparing OpenAI and Anthropic pricing or benchmarks.
The full version, fast.
GPT-5.6 is framed as a dot-release that behaves like a generational leap, because OpenAI squeezed the remaining headroom out of the GPT-5 training run. The reviewer let Codex run unsupervised for five days on an eight-word prompt and got a working Excel clone (sorting, formulas, pivot tables), then seven days on a Minecraft clone that reached a playable state in about a day and kept deepening. A sponsored Box benchmark shows GPT-5.6's tiers beating or matching GPT-5.5 on real knowledge work. On price, GPT-5.6 undercuts a rival model (Fable) on both input and output tokens, but Fable is argued to reason further ahead despite being less optimized — an unfinished architecture with more room to grow. The close covers GPT-5.6's three sizes (Luna/Terra/Sol) crossed with five reasoning tiers, and a custom skill for routing tasks across them to save quota.
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 – 00:21
01 · Cold open: the thesis
GPT-5.6 is framed as a dot-release that behaves like a generational leap over GPT-5.5.
00:21 – 02:58
02 · Building an Excel clone with Codex
An eight-word /goal prompt runs for five days and produces a spreadsheet clone with sorting, formulas, data validation, and pivot tables.
02:58 – 03:36
03 · Rough edges and the computer-use bridge
The clone still has rough edges after five days; the agent used computer use on real Excel, and browser use is pitched as a daily-driver capability.
03:36 – 04:51
04 · Building a Minecraft clone ("Blockcraft")
A seven-day run reaches playable Minecraft-like gameplay in about a day, then spends the rest of the time deepening the world.
04:51 – 05:58
05 · Box AI benchmark results (sponsor)
Box's enterprise knowledge-work benchmark shows GPT-5.6's tiers matching or beating GPT-5.5 across the full dataset and industry subsets.
05:58 – 06:45
06 · Pricing: GPT-5.6 vs Claude Fable
GPT-5.6 Sol undercuts rival model Fable on input, output, and cache pricing.
06:45 – 07:34
07 · Honda Civic vs Ferrari: capability comparison
GPT-5.6 is framed as a maximally optimized existing model; Fable is framed as a rawer model with a higher reasoning ceiling.
07:34 – 08:54
08 · Model sizes, reasoning tiers, and delegation
GPT-5.6's three sizes and five reasoning tiers are pitched as a routing grid, with a custom skill to delegate tasks across them.
Lines worth screenshotting.
- An eight-word prompt — 'make an Excel clone, continue until feature parity' — kept an agentic coder running unsupervised for over five days and produced a working spreadsheet with sorting, formulas, data validation, and pivot tables.
- The same /goal workflow ran for seven days building a Minecraft clone, reaching a recognizable playable state in about one day and spending the rest of the time deepening world content rather than adding surface features.
- The coding agent used computer use to open the real desktop version of Excel and cross-reference it live while building its clone, rather than working from training data alone.
- On Box's enterprise knowledge-work benchmark, GPT-5.6 Sol scored 64.4% accuracy versus 63.3% for GPT-5.5, while the smallest tier (Luna) matched GPT-5.5's mid-tier score at a fraction of the cost.
- GPT-5.6 Sol prices at $5 per million input tokens and $30 per million output tokens, versus $10 input and $50 output for the rival model Fable — half the sticker price on both ends.
- Cached input tokens are priced far below fresh input on both platforms: as low as $0.50 per million for GPT-5.6 cache hits versus $10 per million for fresh input.
- The reviewer explicitly reframes the cost question as 'cost per finished task, not cost per token' — a cheaper model that needs more retries can end up more expensive in practice.
- GPT-5.6 is described as the fully optimized end-state of an existing architecture, while the pricier rival (Fable) is described as an unoptimized, freshly trained model with a higher ceiling still to climb.
- GPT-5.6 ships in three sizes — Luna (smallest), Terra (medium), Sol (largest) — each selectable across five reasoning tiers from Light to Ultra, turning one model family into a cost/quality dial.
- The reviewer manually routes work across tiers — planning on the largest/highest-reasoning tier, implementation on a mid-tier, and low-stakes tasks like deploys on the smallest/cheapest tier — using a custom-written delegation skill.
Treat one model family as a routing grid, not a single pick.
MODEL ROUTING
The cheapest model per token isn't automatically the cheapest per finished task, so evaluating a model family means checking cache pricing, reasoning tiers, and real task-completion cost together.
02Building an Excel clone with Codex
- A single eight-word instruction can substitute for weeks of feature-by-feature spec work if the agent is willing to run unsupervised for days.
- Letting an agent iterate for five straight days on one goal produced a spreadsheet clone with sorting, formulas, data validation, and pivot tables — multi-day autonomous runs are a real build strategy, not a novelty.
- An agent cross-referencing the real desktop application while building a clone of it is a concrete use of computer use, not just a browser-automation party trick.
03Rough edges and the computer-use bridge
- Long autonomous runs still leave rough edges even after five days, so treat the output as a strong draft, not a finished product.
- Browser and computer use are now solid enough to be used as a daily driver for real tasks like sorting email or making DNS changes with a single prompt.
04Building a Minecraft clone ("Blockcraft")
- A multi-day autonomous run doesn't just add features linearly — once the core loop works, most of the remaining time goes into depth (more variety, more detail) rather than new surface features.
06Pricing: GPT-5.6 vs Claude Fable
- Compare full API pricing, not just the headline rate: check input, output, cache-hit, and cache-write tiers separately, since they can differ by an order of magnitude within the same model.
- Cached-token pricing deserves its own line item in any cost comparison — repeated-context workloads should be evaluated on cache pricing, not sticker input price.
07Honda Civic vs Ferrari: capability comparison
- A cheaper, faster model isn't automatically the better one — a maximally optimized existing architecture can still lose on raw reasoning to a less-polished but newer one.
- When comparing a 'cheaper but polished' model against a 'pricier but smarter' one, weigh headroom for future improvement alongside today's benchmark numbers.
08Model sizes, reasoning tiers, and delegation
- A model family that ships in multiple sizes times multiple reasoning tiers gives you a cost/quality dial instead of one fixed price point.
- Route tasks deliberately: highest tier for planning, a mid-tier for implementation, and the cheapest tier for low-stakes work like deploys.
- A custom routing skill or script that delegates across model tiers is a practical way to cut spend without sacrificing quality on the parts of a task that actually need it.
Terms worth knowing.
- /goal
- A command used to start a long-running autonomous coding loop that keeps working toward a stated objective, potentially across multiple days without supervision.
- Codex
- An agentic coding tool capable of browser use and computer use, used here to run multi-day unsupervised builds.
- Computer use
- An AI capability that lets a model directly operate a desktop's graphical interface, such as opening and interacting with a real application.
- Reasoning tier
- A selectable level of inference-time thinking effort (e.g. Light, Medium, High, Extra High, Ultra) within a single model, trading cost and speed for reasoning depth.
- Model routing
- The practice of sending different subtasks to different-sized models based on how much reasoning or cost each subtask actually needs.
- Cached input tokens
- Previously-seen prompt tokens that are billed at a steep discount versus fresh input tokens when reused within a caching window.
Things they pointed at.
Lines you could clip.
00:00
“Here's the thing about GPT 5.6. It is truly a massive leap from GPT 5.5.”
03:29
“It used computer use. It opened up Excel on my desktop and would just go back and forth between doing something in Excel, the actual Excel, and then recreating it in this new cloned version.”
06:42
“It's kinda like GPT 5.6 is the most souped up Honda Civic you've ever seen. Every single horsepower has been squeezed out of it.”
07:16
“And Fable is like a Ferrari that hasn't been touched yet. Fresh off the manufacturing line, unoptimized, and the potential is just so much higher there.”
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.
metaphor
00:00Here's the thing about GPT 5.6. It is truly a massive leap from GPT 5.5.
00:07It does not seem like they should have just done a dot upgrade. They have effectively squeezed every drop of juice out of the GPT five training run that they possibly could, and the result is one of the most effective models and capable models that I have ever used.
00:27One of the first loops that I gave it was super simple. Look at this. Slash goal, make an Excel clone, continue until feature parity.
00:34Here is the thing about 5.6 plus Codex. It has all the tools in its tool belt to do this really well. Just a simple prompt, which is effectively eight words long, ran for over five days before I manually stopped it.
00:51And here is what it created. This is the Excel clone, and it has a lot of the features that Excel has. And if I would have let Codex keep churning through tokens, it probably would have gotten even further.
01:02Alright. So here, very simple. We have a list of numbers.
01:05You can easily sort it by ascending, descending. If I add another number in there, same thing.
01:14And everything is just super fast, works really well. We have all the formulas. So let's say equals this plus this nine.
01:25And if you double click into it, you get the formula up in this input bar up here like normal. We can add another one minus 43, hit enter, and everything just works.
01:35So we have a bunch of standard Excel features. We have data validation, conditional formatting. We have tables, find and replace, obviously, toggling and sorting, very easy.
01:45We have all of these different formula that you can use right from this single page HTML app. Remember, Excel is a massive piece of software, and a lot of the functionality has been condensed down into this very simple app.
02:00You can go try this out, which I'll drop a link down below for that. Shout out to here dot now for hosting all of the demos you're gonna see today. It also has really deep analysis capabilities, including pivot tables.
02:11Look at how easy this is. Simply select some data with headers. You get all the rows sorted out.
02:18You get filters. You can create the pivot table. And look at that.
02:22We actually have a full pivot table. Done. And, again, this ran for five days.
02:27Imagine how long it took to actually create Excel. Took years and years and years to get it to where it is. And, obviously, this is a subset of the total functionality.
02:34And, obviously, as you can see here, there are some rough edges. But, again, I stopped it after five days, and it was not anywhere close to being done.
02:43And here's the interesting bit. It used computer use. It opened up Excel on my desktop and would just go back and forth between doing something in Excel, the actual Excel, and then recreating it in this new cloned version.
02:58And that is one of the things that GPT 5.6 is so incredible at, browser use and computer use. It is phenomenal at browser use.
03:10I really cannot overstate that. Codex's browser has increasingly become my default, my main driver browser because I can just get stuff done with it.
03:20I can open up Gmail and have it sort through my emails. I've done complex DNS record changes easily with just one prompt with codecs browser use.
03:31And, of course, I had it clone Minecraft. I did slash goal once again, and I just said create a clone of Minecraft feature parity. And it went for something like seven days before I finally stopped it, and it took only about one day to get what looked like actual Minecraft, but what it kept doing was going deep and would build out parts of the world that didn't exist before, would build out mobs that came straight from the actual game of Minecraft, different biomes.
04:01And it would just continue churning and just trying to get to feature parity. And it was just so impressive.
04:08And watch this. This is definitely the best version of Minecraft that I've ever created using AI.
04:13Alright. So here we go. You can see the shadowing is really nice.
04:17Here I can mine the grass blocks pretty easily. Really cool three d animation when you do that. You can pick up all the blocks that you just mined.
04:25Here we have miniature carrots. There's some carrots. We have farmland.
04:29And And it just works really, really well. Here's a tabby cat. Here, I broke the glass.
04:34And so it is just incredibly impressive. It is a very full world, more full, more realistic than any other Minecraft that I've ever created. You can generate different seeds.
04:46Here it is. Here's that new world I just created. Here is my full inventory.
04:51Very easy to use. Very, very cool. And, of course, our friends and partner on this video, Box, has put together their own benchmarks on enterprise grade work.
05:00Let's take a look. We have Box AI complex work eval for GPT 5.6, Solterra, Luna.
05:07And Box's benchmark tests real knowledge work, like reading documents, reconciling numbers, doing due diligence, and reviewing expert output for errors. We have GPT 5.5 on the accuracy, 63.3, and Tera at 59%.
05:23So, obviously, a drop from Soul, even a drop from 5.5. We have 5.6 Luna, basically getting the same score as Tera, and much less expensive.
05:35Here, we have industry subsets. We have the public sector, life sciences, and health care. All three sole dominated GPT 5.5.
05:45Thanks to our partners at Box for putting together this awesome benchmark. I'm gonna drop a link down below where you can read more about their benchmark and specifically how GPT 5.6 did on it. All three of these models are coming soon to Box AI, so go check them out.
06:00I'll drop a link down below. The pricing is also much better for GPT 5.6. Not only is it less expensive, but it uses less tokens to get to the same result.
06:10So $5 per million input tokens versus $10 for Fable, much cheaper on cash hits, and 30 versus $50 per million output tokens.
06:21So, again, you're just paying less, and it just feels like 5.6 has a more direct line of sight to accomplishing the task versus Fable. Now with all of that said, Fable is something else.
06:33It feels much more like a brand new model, something that I hadn't used ever before, and it is very impressive. It sees around corners better than GPT 5.6.
06:45Here's the best way to explain it. 5.6 feels like the absolute pinnacle of an existing model where Fable feels like we're just scratching the surface on what's possible on a brand new training run.
06:57That's the difference. Here's an analogy for you. It's kinda like GPT 5.6 is .
07:07Every single horsepower has been squeezed out of it. The tires are optimized for speed. There's a spoiler, everything.
07:15And Fable is like . Fresh off the manufacturing line, unoptimized, and the potential is just so much higher there.
07:26So I put together a full review. I have a bunch of demos of things that I've built, including an actual operating system. We have a Rube Goldberg lab that you can use.
07:34All of this is available on it. I'm gonna drop a link to the full review down below.
07:39And here's the other thing that is very different about GPT 5.6 versus Fable. It comes in three different model sizes, Luna, the smallest, Terra, the medium, and Soul, the largest. But even within those, you have multiple levels of reasoning.
07:54And if we choose sol, you get all the way up to ultra, which is basically a quota burner.
08:02And so we've been talking a lot about model routing lately, where maybe you're using Fable for planning, and then you can actually call codecs from within Claude code and delegate off to GPT 5.5. And now you can kinda just do everything with the GPT series of models. You plan with Sol.
08:18Maybe you do most of the implementation with Terra on, you know, high reasoning. And then for stuff like deploying or other kind of low requirement work, you offload to Luna. And so now you have all these different sizes, all these different thinking effort settings for these different models, and you can come up with a really nice skill to delegate between them.
08:40And by the way, I wrote a skill just for that. And if you want it, I'm gonna drop a GitHub link down below so you can delegate all within Codecs and save yourself a bunch of your quota and get basically the same quality perform
The Hook
The bait, then the rug-pull.
The video opens on a claim that shouldn't be true of a 'dot' release: a generational leap from a minor version bump. What follows is two multi-day unsupervised coding runs, a sponsored enterprise benchmark, and a live pricing and capability standoff against a rawer, higher-ceiling rival model.
Named ideas worth stealing.
07:59model
Size x reasoning-tier routing grid
- Luna (small)
- Terra (medium)
- Sol (large)
- Light
- Medium
- High
- Extra High
- Ultra
Three model sizes crossed with five reasoning-effort levels give a single model family a cost/quality dial instead of one fixed price point.
Steal forany agentic workflow that needs to route different subtasks to different cost/quality tiers
06:42concept
Honda Civic vs Ferrari analogy
GPT-5.6 is the most-optimized version of an existing architecture (every drop of performance already squeezed out); the rival model Fable is a fresh, unoptimized architecture with a higher ceiling still to climb.
Steal forexplaining the tradeoff between a polished incumbent and a rough but higher-ceiling challenger
How they asked for the click.
VERBAL ASK
08:25link
“I'm gonna drop a GitHub link down below so you can delegate all within Codecs and save yourself a bunch of your quota”
Soft CTA folded into the closing explanation rather than a hard ask; the sponsor read for Box earlier in the video ('link below!') is a harder, more direct ask.
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.
