Transcript
00:00:00GPT 5.6 Soul just came out and right now everyone's asking the same thing.
00:00:04Is it finally the model that beats Claude?
00:00:06As you already know, we're a software company,
00:00:08so we spent the last four days running Soul on our own products,
00:00:11including our custom-built community platform and on other client projects as well.
00:00:16And what we found isn't that it's better or worse,
00:00:18it's that Soul behaves very differently from Fable,
00:00:21so using both models in the same way means you're not getting the best out of either one.
00:00:25It finished some work in one run that normally takes Fable several conversations,
00:00:29but it also kept taking actions in places where Fable would normally stop,
00:00:32and during reviews, it found problems in parts of our app that Fable had refused to inspect.
00:00:37So after four days of testing, we ended up with seven rules for getting the best out of Soul.
00:00:42By the end of this video, you'll know what to use it for,
00:00:44where it beats Fable, where it doesn't, and the one setting you should leave turned off.
00:00:49Before we actually get into the video,
00:00:51let's catch up on OpenAI's latest model release in case you've missed it.
00:00:54So OpenAI just shipped their latest models under GPT 5.6
00:00:58and this time, they released three of them.
00:01:01There's Luna and Terra, which are the smaller and cheaper models,
00:01:04and then there's Sol, which is their most capable one.
00:01:06So what's actually new here?
00:01:07OpenAI's whole pitch this time is efficiency,
00:01:10basically getting more work out of every token and more performance out of every dollar you spend.
00:01:15OpenAI says,
00:01:17Fable on coding and long-running tasks while using fewer tokens and finishing in about half the time.
00:01:22On overall intelligence, their own numbers put it at about the same level as Fable, just at a lower cost,
00:01:28but we'll be taking a look at that later on.
00:01:30They've also folded the separate Codex desktop app into the ChatGPT app,
00:01:34although you can still use Sol through the Codex CLI, which is the terminal version of Codex.
00:01:38OpenAI also added a new mode called Ultra, which runs multiple agents at once instead of one.
00:01:44We'll come back to whether that's worth using later on.
00:01:46So right now, there are two models at the top.
00:01:48There's Fable, the one you already use, and now there's Sol.
00:01:51So the rest of this video is about how to get the most out of Sol, including when to use it over Fable.
00:01:57Let's start with the honest part first.
00:01:58Sol is not smarter than Fable.
00:02:00On hard problems without a clean solution, we've seen that Fable still finds the better answer.
00:02:05The difference shows up once the decision has already been made and the model has to complete the work.
00:02:09Fable is better at figuring out what you should do,
00:02:12while Sol is better at continuing through every step without needing you to keep bringing it back to the task,
00:02:17and that difference changes how you should use it.
00:02:20Starting off with the first rule, you need to set a boundary before you let Sol work on its own.
00:02:24We didn't have to think about this as much with Fable because Fable normally stops or refuses when it thinks an action is risky.
00:02:31Sol behaves differently.
00:02:32It keeps taking actions until it finishes the task.
00:02:35The clearest example happened while Sol was working through one of our tasks.
00:02:39It found files and running processes in its way,
00:02:41so it deleted the files and stopped the processes without asking us first.
00:02:45Fable would normally stop at that point, but Sol removed what was blocking it and continued with the task.
00:02:50Other people testing Sol saw a more extreme version of the same behavior,
00:02:54and we even saw one case where it deleted almost every file on someone's Mac.
00:02:59So before you leave it running, you need to limit what it can access.
00:03:02Before you start a long task, ask Codex to create a separate Git branch and commit the version that already works.
00:03:08The branch doesn't limit what Sol can touch, but it gives you a clean checkpoint so you can undo its changes if something goes wrong.
00:03:14To set this up in the Codex app, open the configuration, set the approval policy to never, and set the sandbox to workspace right.
00:03:21Never means Sol can keep working without waiting for you, while workspace right means it can only touch files inside that project.
00:03:28Just don't select full access because that gives it access to files outside your app as well.
00:03:32Now once you've set those boundaries, you can use that same behavior to your advantage.
00:03:36We saw the other side of it when we gave Sol a full feature in AI Labs Pro.
00:03:40The same kind of work with Fable normally takes us several rounds of back and forth, but Sol stayed on the same goal and worked through the entire feature on its own.
00:03:48It also finished much faster than the same work normally takes us with Fable.
00:03:52So use Sol for the work you'd normally have to split across several sessions.
00:03:56That could be a full feature, a complete review of your app, or any task that touches several parts of the product.
00:04:01And one place we saw this difference clearly was during reviews, which brings us to the second rule.
00:04:06But first, let's have a word by our sponsor.
00:04:08Salad Cloud
00:04:09Running open AI models usually means renting cloud GPUs that cost a fortune.
00:04:13Salad flips that.
00:04:14Since 2018, they've run a distributed GPU cloud built on the world's idle consumer GPUs, over 2 million of them,
00:04:21so their prices are often a fraction of the big providers, sometimes cheaper by an order of magnitude.
00:04:26For developers, the best part is their AI gateway.
00:04:29It's an open AI compatible API, so you just point a tool like Cursor or Klein at one endpoint and run open models like Quen 3.6.
00:04:38No servers, no cold starts.
00:04:39I swapped my endpoint for Salad's Ranquen 3.6 in Klein, and it built the feature end-to-end, the same workflow, at a tiny fraction of the cost.
00:04:48And next to the usual cloud providers, the price difference is genuinely hard to believe.
00:04:53There's no lock-in, no contracts, and you can scale up the moment you need to.
00:04:57Sign up with my link below, then message their support to claim $10 in free credit and try it yourself.
00:05:02Sol is also better at reviewing work that Fable refuses to touch.
00:05:06And to be clear, we're not talking about the occasional routing to Opus 4.8,
00:05:10which mostly happens when you ask it to explain its reasoning or the task involves something harmful.
00:05:15We're talking about the safety restrictions built directly into Fable.
00:05:18We know this because we've already used Fable for these kinds of reviews, so we knew where it would normally stop.
00:05:24When we gave the same kind of work to Sol, it didn't stop there.
00:05:27It kept reviewing the parts of the app Fable had always refused to inspect, and it found problems in those areas.
00:05:32But this is only the first reason Sol reviews better.
00:05:35The other one is that it can open the finished app and test it itself, which brings us to the third rule.
00:05:40Whenever a task involves something you can see or click on the screen, Sol will automatically start using computer use.
00:05:46Inside the new ChatGPT app, Sol's computer use can work across multiple tabs,
00:05:51use websites you're already signed into, and download files.
00:05:54So it can move through the same screens you use instead of only working from the files inside your project.
00:05:59Here's what this looked like in our custom-built community.
00:06:01We were building a new store feature, and the app doesn't work the same way for every account.
00:06:06An admin has one set of options, a regular member has another,
00:06:09and someone on the $20 plan should be able to do different things from someone on the $100 plan.
00:06:15So checking the feature once isn't enough.
00:06:17Sol went through the full journey for every one of those accounts.
00:06:20It logged in as each type of member, checked what they could access, went through the process of purchasing a product,
00:06:26and then repeated those checks across the other parts of the app that change depending on the account.
00:06:31It tested the feature from the point of view of every person who would actually use it.
00:06:35So when a feature behaves differently for different people, run the review in a new Sol conversation after the build is finished.
00:06:41Give it the different types of accounts and explain what each one should be able to do.
00:06:45Then ask it to complete every journey from start to finish and only report the problems it can reproduce.
00:06:51For our store, that meant logging in as every account, purchasing a product, and checking that each person saw the right version of the feature.
00:06:57Now this community website we are talking about contains all the skills, workflows, and other resources that we build and show you in our videos.
00:07:05So if you found value in what we do and want to support the channel, this is the best way to do it.
00:07:09The link's in the description.
00:07:10Computer use changes how Sol checks the finished result, but you also need to change how you give it the task in the first place.
00:07:17Which brings us to the fourth rule.
00:07:18The fourth rule focuses on restructuring the prompts and skills you were already using.
00:07:23And this applies to both Fable and Sol, because the newer models are better at working out the basic steps for themselves.
00:07:29So the goal isn't to make every prompt as short as possible.
00:07:32It's to remove instructions that don't change the result and make the instructions that do matter more precise.
00:07:37OpenAI explains this in the GPT 5.6 prompting guide.
00:07:41You should still give Sol the outcome, the important limits, and what needs to be true before the task is finished.
00:07:46But you don't need to map out every folder when the folders are already named properly,
00:07:50repeat the same rule in three places, or describe every step the model should take to get there.
00:07:55We can show you what this looks like through our animation system.
00:07:58The first thing we kept was the actual code the model needs to replicate in the HTML.
00:08:03Instead of describing the animation and leaving the model to interpret it, every rule gives it a working pattern it can copy.
00:08:09We also kept the exact background, surface, and accent colors so the model doesn't choose new ones on every run.
00:08:15But the more useful part was finding what we could remove.
00:08:18One of our rules told the model to always save the finished animation inside the output folder.
00:08:23We removed that line and ran the same kind of task again, and Sol still saved it there.
00:08:27The instruction wasn't changing the result because the model could already work out where the finished file belonged.
00:08:33We're still keeping that line for now because we switch between models,
00:08:36and a weaker model may not pick up the same pattern, but that's the test you should run on your own prompts and skills.
00:08:42Remove one instruction, run the same kind of task again, and check whether the result changes.
00:08:47If the model still does it reliably, that instruction probably doesn't need to be there.
00:08:52OpenAI reported that this kind of cleanup improved its own results by 10-15% while using 41-66% fewer tokens.
00:09:00Now the prompt controls what Sol is trying to do, but the mode controls how many agents it uses to do it, which brings us to the fifth rule.
00:09:07And if you're enjoying the video so far, subscribe to the channel and hit the hype button.
00:09:12This small gesture of support goes a long way for us.
00:09:15Keep Ultra turned off for now.
00:09:17Ultra is Sol's multi-agent mode.
00:09:19So instead of running one agent, it divides the task and runs several of them at the same time.
00:09:24OpenAI positions it as the mode for the hardest work, but every agent uses its own tokens,
00:09:29which means Ultra can burn through your usage much faster than regular Sol.
00:09:33And when we used Ultra on the same kind of work we'd already been running through regular Sol,
00:09:38we didn't see a meaningful improvement in the finished result.
00:09:41The usage increased because several agents were running, but the quality stayed about the same.
00:09:46OpenAI's own results show the same small difference, with Ultra improving the scores by roughly 2 or 3 points depending on the test.
00:09:54And if regular Sol is your default, the next question is whether you should save more usage by sending the smaller tasks to Terra or Luna.
00:10:01If you're using Codex to build your app, keep Sol as the default instead of switching smaller tasks to Terra or Luna.
00:10:08Terra and Luna are cheaper through the API, which matters when you're paying for every request your own product sends to the model.
00:10:15But when you're using Codex through a subscription, the decision is different because you're working within the usage included in your plan.
00:10:21We found this after routing the lighter parts of our work to the smaller models.
00:10:25They completed the task, but the drop in quality meant we spent more time correcting the result than we normally did with Sol.
00:10:31So we went back to Sol for the full task instead of changing models halfway through it.
00:10:36So that leaves the last question. When should you use Sol and when should you go back to Fable?
00:10:40After using both models in our daily workflow, the clearest way we'd position them is this.
00:10:45Fable is a preview of where coding models are heading, while Sol is the most complete version of how they work today.
00:10:51And the speed difference affects the workflow more than the benchmark scores do.
00:10:55In our daily use, Fable, Opus and Sonnet all take longer to complete the same type of work.
00:11:00On the other hand, Codex models are way faster and we're saying that without even turning on Codex's fast mode, which uses 1.5 times your usage.
00:11:08This in our opinion is one of the greatest deciding factors and it's the reason that we've seen ourselves reaching more for Codex models rather than Claude.
00:11:16So use Fable before the build when you need help deciding how the app should work, how a feature should be structured or why a difficult problem keeps happening.
00:11:24Once those decisions have been made, have Sol build it, that's where you use it for the full feature, the long running work, the review and anything that needs to be checked on the screen.
00:11:32So after four days of using Sol, it's become our default once we already know what the finished result should do, while Fable still handles the decisions that need more judgement.
00:11:41The important part is setting the boundary, giving each model the right job and reviewing the result before you ship it.
00:11:47That brings us to the end of this video. If you'd like to support the channel and help us keep making videos like this, you can do so by using the super thanks button below.
00:11:56As always, thank you for watching and I'll see you in the next one.