# Claude Fable 5 is here!

## Метаданные

- **Канал:** 1littlecoder
- **YouTube:** https://www.youtube.com/watch?v=9Lm7ASzuV6w
- **Дата:** 09.06.2026
- **Длительность:** 8:02
- **Просмотры:** 261
- **Источник:** https://ekstraktznaniy.ru/video/52963

## Описание

https://www.anthropic.com/news/claude-fable-5-mythos-5

Today we’re launching Claude Fable 5: a Mythos-class1 model that we’ve made safe for general use.

Fable 5’s capabilities exceed those of any model we’ve ever made generally available. It is state-of-the-art on nearly all tested benchmarks of AI capability, showing exceptional performance in software engineering, knowledge work, vision, scientific research, and many other areas. The longer and more complex the task, the larger Fable 5’s lead over our other models.

Releasing a model this capable comes with risks. Without safeguards, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage. We’ve therefore launched the model with safeguards that mean queries on some topics will instead receive a response from our next-most-capable model, Claude Opus 4.8. To release the model both safely and quickly, we’ve tuned these safeguards conservatively—they’ll sometimes catch harmless requests, though they 

## Транскрипт

### Segment 1 (00:00 - 05:00) []

Remember the LLM that Anthropic said it's too dangerous to release a couple of months back Mythos? The same family of model, Mythos family of models, latest version Fable 5 has been launched. This is a new class of model along with the existing three items, Sonnet, Haiku, and Opus. Fable probably marks the highest tier of Anthropic Claude models. That also means it's going to be one of the most expensive models available on the planet right now. Right at the start, the pricing is double, almost double the price of Opus. So, input token is $10 per million input token and output token is $50 per million output token. Now, if you think this is expensive, this is still three times cheaper than GPT 5. 5 Pro. So, you can basically make out whether this model is for you or not purely based on your existing usage. If you have been using GPT 5. 5 Pro, probably Fable 5 is a model that you would love, you would need. But, if you are not using that kind of a model or you don't have use cases for that kind of model, then this model is a really, really good to have model, but you may not have a lot of use cases. The model is already available on claude. ai. You can use it with the different IDs like Cursor, but before you go and fire up the first prompt, just remember that this model is extremely slow, very heavy, and the model is context hungry. It is going to finish all your limits and tokens in a very short span. So, very wisely use this model. Don't use this model just to ask how many hours are there in strawberry. This model crushes almost all the benchmarks available for programming, especially for programming. So, if you were to compare agent decoding sweet bench pro, this model has scored 80. 3% Opus 4. 8 has scored 70%. This is like a 10% point increase. This is a frontier code. This is a benchmark. I think this benchmark is from Cognition and Devin's company. This model is again scored 30%, which is a double of what the previous version is. Much, much higher, like six times higher than GPT 5. 5. And you can see across all the benchmarks, whether it is computer use, whether it is tool usage, whether it is knowledge work, whether it is like agent decoding, terminal bench, any benchmark that you pick, this model is extremely good at that particular task. Does it mean benchmarks are exhausted? Let's pick a particular benchmark which is called Cursor Bench, which is an independent benchmark, to understand how the cost benefit ratio works for this particular model. Now, when you look at this chart, you would immediately notice that Fable 5 Max is probably the best. So, it has scored 72. 9%. So, for you giving a comparison, Kimmy K2. 6 scored less than 50% on the same benchmark. Sonnet 4. 6 High So, when you see a benchmark like this, you can see that this model has scored almost 20% point more than any average model that we usually use for a lot of different use cases. But, the cost for that particular task is $18 for that particular task. So, it used 63,000 tokens, it took 76 steps, and it costed $18. Now, when you go down and take a comparison, Kimmy K2. 6 costed $1. 2. So, this model is technically almost like 17 times more expensive than Kimmy K2. 6, but it is 20% point or 20. something percentage point more effective in terms of the score. And it's very important for you to understand this distinction and understand the cost benefit. So, if you have got tasks where Fable 5 can excel, then Fable 5 Max, Fable 5 Extra High, Medium, and all these different modes can take your, let's say, the accuracy of the task that you have got extremely high. So, when you compare it with Opus 4. 7 Max, this is much better. GPT 5. 5 extra high. Fable 5 low is almost equivalent of GPT 5. 5 extra high. And there are like a lot of other versions of this model. You can see that this model is crushing all the other models and it is the leaderboard topper. But at what cost is something that you have to understand. And don't get me wrong because I'm emphasizing on the cost. Because every time there is a new model launch, everybody gets excited. They're like they're pushing the frontier. But it is not for everybody that is honestly what I want to make. I think Open AI launched something called GPT 4. 5. Around the time the model was super heavy. The model was slow. Everybody used the model and became disappointed. Ideally if you're a programmer, this model is not going to disappoint you. The samples that I've seen with this model, the model has got extremely good taste. So the model is not going to disappoint, but probably the cost aspect is like will it benefit using a human being to do this or the trade off of using an AI instead of human being with this particular model. Today like at this current situation of this cost, does it make sense? Is the question a lot of companies going to ask? But if you're a hobbyist, you don't mind about spending a lot of money or if you're a researcher with a great fund

### Segment 2 (05:00 - 08:00) [5:00]

and great grant, you want to you know push the frontier, then this model is exactly for you. The way this model works is for every task that you ask, the model is going to have extremely good context. run much much longer than any other model that you have used. So if you've got a long horizon task that has to run for longer hours, that has to consume lot of knowledge, then obviously this model becomes an extremely good candidate. In fact, Anthropic in its uh system report has specifically said where people have quoted this model that using Methuselah 5, our internal protein design experts accelerated aspects of drug design process by around 10 times. In one example, they found that Methuselah 5 with protein design and bioinformatics tools, but no human assistant matches or beats skilled human operators. See now, this is exactly the kind of problem where you would need this particular model. One more example. During early testing, Stripe reported that Fable 5 compresses months of engineering into days. For example, in a 50 million line Ruby code base, the model performed a code base wide migration in a day that would otherwise have taken a whole team over 2 months by hand. What we do not get to know from these kind of beautiful good-looking quotes is that we don't know how much tokens that they spent. what kind of like rate limits they had. We don't know how much money that they spent. So, once again, if you are going to use this model for, let's say, a frontier task, like a task that requires pushing human intelligence further, whether it is research, whether it is coding, um then this model is extremely good model. Like, in fact, there is a company called Every. They have got an internal senior AI engineer benchmark, and this model has completely crushed every other model that has been available, and this is like the leader of the benchmark at this particular point. But the question is, do you need this kind of a model for an everyday task? Will your employer or can you, if you are like just a normal human being, can you afford? I don't think normal human beings can easily afford this particular model. Nevertheless, this is a great model. For this particular price point, this model is extremely competitive with GPT 5. 5 Pro, and I think OpenAI has a lot to prove because Anthropic has consistently produced great programming models. But if that is not what you want to do, you don't want to do code migration, protein folding, you don't want to do bio testing, you can't use this. Also, another aspect is you can't use this model for cybersecurity or other aspects where you can misuse the model. So, this model is also extremely protected. I think if you have got a really good use case that requires like let's say a few days of human work, like a bunch of team members coming together and then executing something, then this model is exactly for that if you're willing to pay the price of it. See you in another video. Happy prompting.