스크립트
00:00:00Fable 5.1 is the latest release from Anthropic, and it's one of the best models you can use right
00:00:05now. But it doesn't behave the same way the previous models did, and there are a few small
00:00:09things you need to change in how you work with it. Because if you use it the way you used the
00:00:13previous models, you're not using it to its full potential, and you might even get worse results
00:00:18out of it. For example, Anthropic says this model is best for long-running tasks, but it only does
00:00:23that well if you prompt it in a specific way. And there's one setting that Anthropic says to keep
00:00:28turned up, but it's actually making the results worse. If this is your first time, we're a software
00:00:32company, and this is our channel AI Labs. And in this video, we're going to go over 7 tips for getting
00:00:37the most out of Fable 5.1. But before we go into those tips, there are some things we need to say
00:00:42about this new model first. With the release of Fable 5.1, Anthropic claimed this model is way
00:00:48better at finishing tasks. But Fable 5.1 isn't doing better work. It's getting through more work
00:00:54for less money. We saw this on our own client projects. We gave Fable 5 and Fable 5.1 the same
00:01:00tasks and ran each one several times. We can't reveal the details of the projects, but Fable 5 did
00:01:05better work on most of the tasks, and Fable 5.1 was only better on one of them. That one was the
00:01:10hardest task of the lot, and on it Fable 5.1 did better work, finished around 40% faster, and cost well
00:01:16under half as much. So Fable 5.1 is stronger when the job gets harder. And even on the tasks where its
00:01:22work wasn't better, it kept going. Fable 5 stopped before the end of some of its attempts, while Fable
00:01:275.1 finished every single one. It also costs less on every task. Now you might already know that Claude
00:01:33Code has a usage command that basically estimates what a session would have cost you if you were paying
00:01:38per use instead of on a subscription. And by that estimate, Fable 5.1 was around 47% cheaper than
00:01:45Fable 5. But it isn't cheaper because the model suddenly uses fewer tokens. It's because the price
00:01:50of Anthropic changed. But before we get to that price change, you need to know one thing. The longer
00:01:55the task runs, the longer the conversation grows. And because of that, the model has to read more
00:02:00before it answers a new prompt. Models don't remember the conversation from one message to the
00:02:05next. So every time a model needs to give you another response, the whole conversation has to
00:02:10be sent to it again. But instead of processing the same conversation from scratch each time,
00:02:16Claude keeps a saved copy of the parts it has already processed and reuses that copy the next time you
00:02:21prompt it. This is basically called a cached read. These cached reads also cost, but they are lower in
00:02:27price as compared to the usual input and output. And with Fable 5.1, Anthropic cut the price of that
00:02:33by 75%. So the conversation that gets sent again before each response now costs way less. That saving
00:02:39might look small on a single prompt because there isn't much conversation to send again. But on a long
00:02:44task, the same growing conversation is sent again and again, so the saving grows every time the model
00:02:50responds. Now that only covers what the model reads. What the model writes is priced separately and that
00:02:55price didn't change. An artificial analysis, who test models and score them independently, found that
00:03:01Fable 5.1 wrote way more than Fable 5 did. So the bill went up even with the cheaper reads. In their
00:03:07tests, it generated around 1.7 times more output than Fable 5 and cost about 20% more per task.
00:03:14Anthropic's Fable models have safety guardrails, which are rules that stop them from doing certain types
00:03:19of tasks. And with Fable 5.1, Anthropic says those guardrails interrupt Claude code sessions
00:03:25around 60% less often. But when it does hit one of those guardrails now, Claude code switches to a
00:03:31weaker model without telling you and keeps going. And that weaker model handles the rest of the session
00:03:36even when your next request has nothing to do with what triggered it. So instead of a refusal, you get
00:03:41weaker work and think Fable 5.1 produced it. And this has already happened to someone building a game
00:03:46that had a joke file with the word biological in it. The project had nothing to do with biology,
00:03:51but that one word was enough to switch the model handling the work. And if your project is actually
00:03:55about one of the subjects those guardrails cover, like security or biology, there's no setting and no
00:04:00way of wording it that avoids this. It's just something Fable 5.1 costs you. Now on artificial analysis's
00:04:06tests, Fable 5.1 scores higher than Fable 5 on accuracy, but it also hallucinates more, which is when a
00:04:13model doesn't know the answer. So it invents one and presents it as true. When Fable 5.1 didn't know
00:04:18an answer, it hallucinated 72.6% of the time. Fable 5 did that 63.6% of the time. And a model saying it
00:04:26doesn't know is actually better than a model pretending to know everything because then you
00:04:31can go and check things yourself. But when it makes something up and you trust it because the model said
00:04:35so, that made up answer goes straight into your work. And that matters even more when you're using it
00:04:40for research or writing because verifying that becomes a whole task later on. But before we move
00:04:45on to the tips, it would be great if you subscribe to the channel and hit the hype button. This small
00:04:50gesture of support goes a long way for us. Now after hearing these things, you're going to say that the
00:04:56older model is clearly better in some aspects. That is true, but at the same time it's going to become
00:05:01the default model in the apps that you use. This is why you need to follow the tips that we are going to
00:05:06talk about. Otherwise, you're not going to get its full potential. Now the first way to actually use
00:05:11Fable 5.1 better is to turn down the effort. For those who don't know, effort is the setting that
00:05:16decides how much work the model does before it gives you an answer. And that one setting affects
00:05:21both the quality of the work and what it costs. Code Rabbit, who make a tool that reviews code,
00:05:26tested Fable 5.1 on their real review work and found that it caught 61% of the problems on low effort
00:05:33compared with 57.1% on high, and it finished around 3 minutes faster as well. And we saw the same thing
00:05:39when we used Fable 5.1 to add a feature to our community website. On high effort, it spent a long
00:05:44time going back and forth on one part of the feature, while on low effort, it finished the whole thing and
00:05:49gave us a better result. We recommended low effort in our previous Fable video as well, and after using
00:05:54Fable 5.1, that's still what we recommend. So you should start with low effort and only turn it up if the
00:06:00result gives you a reason to. Otherwise, you're spending time and tokens for very little in return.
00:06:05The next thing you need to do is remove the instructions you wrote for older models. Those
00:06:10instructions stay in your setup when you switch to Fable 5.1, even though the behavior they were
00:06:15fixing is gone. So Fable 5.1 still has to follow a growing list of fixes for problems it doesn't have.
00:06:21They're just wasting your limit, and some of them now cause new problems. We mentioned this in our
00:06:25previous video as well. Now you might think, how are you going to know which instruction is important and
00:06:30which is not? You can test it by taking an instruction out of your Claude.md and giving Claude the same
00:06:36kind of work again. If the result is the same without it, that instruction doesn't need to be there. But you
00:06:41don't need to do everything yourself. Claude Code has a command called Claude API Prompt Audit that finds
00:06:46those instructions for you. You need to run that command in the terminal, and it'll check everything you've
00:06:51saved for Claude. Then it shows you which instructions were written for older models and gives you the exact
00:06:56changes it recommends. It doesn't change anything on its own, so you can review the list first and only
00:07:01approve the changes you agree with. But before we move on to the next tip, let's have a word by our
00:07:06sponsor, Porkbun. So you finally ship your app, but it's stuck on some random default URL, and what it
00:07:12really needs is a proper domain of its own. That's where Porkbun comes in, with .app and .dev domains built
00:07:18specifically for tech projects. A .app domain is perfect for your app or your SaaS landing page,
00:07:23and .dev is made for your coding portfolio or your docs, so the second someone sees it,
00:07:28they know it's a real project. And these aren't just nice names. Every .app and .dev domain sits
00:07:33on the HSTS preload list, which means HTTPs is automatic on every connection with Xero setup,
00:07:39and a free SSL certificate is included, so your site is secure right out of the box. Best part,
00:07:45you can grab your first year for just $5.99 with free HUI's privacy and no upsells or hidden fees.
00:07:52And after year one, renewals stay priced at cost, so it stays cheap long-term. Get your domain for $5.99
00:07:58at porkbun.com/ailabs26. The link's in the description below.
00:08:03Now, Anthropic sells Fable 5.1 for long jobs you can leave running without watching it much, but it still
00:08:09pauses to ask for permission before doing a step that was already covered by your original request.
00:08:14It just asks a question and waits, and when nobody's there to answer, waiting is the same as stopping.
00:08:19This happens because its usual behavior assumes you're working alongside it. So Anthropic's
00:08:24prompting guide for Fable 5.1 says you need to tell it in the prompt that nobody's watching,
00:08:29and that it should continue with anything already covered by your request as long as it can be undone.
00:08:34That gives the model permission to finish a long job without stopping for questions that don't need asking.
00:08:39Fable 5.1 also has a habit of doing more work than you asked for.
00:08:43For example, if you ask it to add one feature, it may notice another problem nearby and fix that as well,
00:08:48or extend the feature beyond what you described. We saw this on one of our client projects,
00:08:53where we asked it for a feature and it also edited the test files that nobody had asked it to touch them.
00:08:58The problem is that those extra edits can change parts of your product that you wanted left exactly
00:09:03as they were. So when you give Fable 5.1 a task, you should also tell it what to leave alone,
00:09:08which means it changes only what the request needs. And if it notices another problem,
00:09:13it mentions that at the end. And when we did that on our project and told it specifically not to edit
00:09:18the test files, it only worked on the feature we asked for. Fable 5.1 has another habit and it happens
00:09:24whenever it changes files. When it's working on a task and only needs to edit a small part of a file,
00:09:29it rewrites the whole file instead, even though the rest of it doesn't change at all.
00:09:33This might look fine to you because you're still getting the right result, but every one of those
00:09:38lines costs output tokens and if it's working with a lot of files, it drains your limit much faster.
00:09:43We saw this on our own second brain, where every small change cost more than it should have,
00:09:48because it kept rewriting files it could have just edited. To stop that,
00:09:52you need to tell Fable 5.1 that when a small change will give you the same result,
00:09:56it should edit just that part. The next problem isn't how much Fable 5.1 writes,
00:10:01but how it writes. We already saw this with Opus 5, which used a lot of what Anthropic calls
00:10:06"mannered prose". For those who don't know, "mannered prose" is when the model uses fancy
00:10:11phrases and metaphors instead of just saying what it means. For example, instead of saying a setting is
00:10:16worth changing, the model calls it a dial worth turning. And in a long research task,
00:10:20that kind of language makes it hard to even see the point the model is trying to make.
00:10:24Now, Fable 5.1 uses less of it than Opus 5 and Fable 5, and it also uses fewer stock phrases and less
00:10:31unexplained jargon. But the problem is still there in its writing, and compared with Fable 5,
00:10:36its sentences run longer with fewer paragraph breaks, which makes it even harder to follow.
00:10:41So you just need to tell it to remove all mannered prose, and that fixes both. That's what we did when
00:10:46we ran it on one of our own research tasks, where we were using it for writing rather than building.
00:10:51The last thing is what kind of work you give Fable 5.1. Everything we've covered so far is about getting
00:10:56more out of it, but the kind of job matters just as much, because Fable 5.1 is at its best on the big
00:11:01ones. When we ran the tasks on our projects, Fable 5 actually did better work on the simplest task,
00:11:07but on the hardest one, Fable 5.1 did better work, and it finished around 40% faster and cost 58% less.
00:11:14So the big jobs are where Fable 5.1 wins, because that's where finishing the whole thing without
00:11:19stopping actually matters. That's why you shouldn't break a feature into lots of small prompts and give
00:11:23it one step at a time. You should give it the whole feature and describe the result you want clearly,
00:11:28and then you'll see it finish the whole feature by itself. Now if you want access to all the skills
00:11:33and workflows that we show you in our videos, you can get them in AI Labs Pro, which is our community.
00:11:38So if you found value in what we do and want to support the channel, this is the best way to do it.
00:11:43The link's in description. That brings us to the end of this video. If you'd like to support the
00:11:48channel and help us keep making videos like this, you can do so by using the super thanks button below.
00:11:53As always, thank you for watching and I'll see you in the next one.
커뮤니티 글
아직 글이 없습니다. 이 영상에 대한 첫 번째 글을 작성해 보세요!
이 영상에 대해 글쓰기