Claude Fable 5 torna: potència, subagents i el cost del nou filtre

It looks like Fable 5 is back on the menu and this comes one day after Enthropix rather underwhelming release of their Sonic 5 model and it's not that the Sonic 5 model was bad. It's still a great model, but when you look at this chart that shows how Enthropix decided on their pricing of Sonic 5 compared to the previous Sonic 4.6, it doesn't really make sense at all because Opus is supposed to be more expensive compared to Sonic, given the intelligence gap. And now we have this weird overlap where Opus is cheaper in some instances than Sonnets.

What's going on here? Thankfully, this video is about Fable 5 and Fable 5 definitely helps save and Thropache. And I can finally open up Cloud and pick Fable 5 as my model and then ask the model just how many times the letter R appears in the word strawberry. Okay, but seriously, how good is Fable 5 really? Whatever we're dealing with the model that's this good, One of the most practical ways is actually dividing the task hierarchically by the

models' intelligence to optimize on pricing and speed. Here's an example. Let's assume that the Fable 5 model lands somewhere in this upper echelon. So given this hierarchical structure in intelligence, we have a higher intelligent model up here, which is much more expensive, and we have a medium-level intelligence right around here, and it's given pricing, and the lower-level intelligence model right here. What if we just carry this hierarchical structure and bring it straight over to a coding agent like Cloud code?

So in that case, we use the Fable 5 model for deep planning and delegations and we use Opus for deep reasoning and finally we use the SON at model for grunt work. Seems like a solid plan, right? This seems to be a common use case among developers. Here's how Diego implemented this. You can simply start by setting Cloud codes model to be by default on Fable and set the effort on X-Hi. And since the main model is now set to Fable, we can now create two sub-agents, one for deep reasoning using Opus and the other for fast execution using Sonet.

And now that we have this hierarchical structure set up, in Cloud Coach shows two of my sub-agents here, I can outline this in Cloud.md file to tell Fable how to delegate properly. Another this hierarchy set up, I can spin this up by asking Fable5 to build me a website that looks like this, showing the solar system interactively, or even create this game that looks like Starcraft, which is one of my favorite games that I play growing up. Pretty cool. These two examples alone ended up showing through 47% of my session usage and already 7% of my fabel usage that set to

expire on July 8th. Now beyond coding use cases that we've seen here, there's a lot to speak about when it comes to the implication of how andthropic and specifically the mythus class model actually impacted the AI industry as a whole. We know that the month of June was a rather shocking moment in the AI industry. And Thropic initially released mythes and Fable publicly on June 9 and later on June 12, both models were suspended due to export control.

And for the next 17 days, and Thropic started to work on a huge safeguarding overhaul to ensure that Fable can route unsafe problems from users to Opus model instead. And finally, on June 30th, the control was lifted and here we are. Now, even though this story arc definitely makes it an interesting story to tell, it actually spoke volume when it comes to what the future might hold when it comes to innovation in the AI industry. If the premise is that we still have so much room left on the table,

when it comes to true breakthroughs in the AI industry, demonstrated yet again by models like Fable 5 and GPT 5.6, what does the future hold exactly from models being safeguarded like this in the future. And meanwhile, open models at GLM5.2 and DeepSeek are slowly catching up in performance while avoiding regulation and under much less scrutiny that my stall, the pace of innovation that front-to-lab seems to be going through at the moment. Now, there's also a lot to say about the consistency of the model when it comes to the

Fable 5's behavior. During the 17-day period I mentioned earlier, where Anthropic developed this router that classified unsafe prompts to route to Opus Model instead, this makes this entire user experience rather confusing to say the least. For example, I asked a simple question to Fable 5 this rather innocuous question that goes something like this. How many permutations are there for a 4-digit keypad? This simple question that might even appear in a statistics 101 textbook is routed to Opus Model instead as you can see since it flagged their safety mechanism.

But if I as a user requested for Fable 5, wouldn't it make more sense to just be denied rather than always having to check whether the answer actually came from Opus or Fable? What I would expect is something more like this, where I'm clearly trying to be sketchy in my prompt, as you can see here, by encoding my characters and shifting the characters plus one,

so that the answer it generates follows that encoding. Perhaps by encoding my messaging this way and having the model also generate its answer based on that encoding, maybe I can just bypass their safety mechanism, and clearly anthropic has already thought about this simple trick, and it ended up shutting down the chat entirely, which to me makes more sense. And even though my question here once decoded is innocuous, which asked the same question

earlier about the 4-digit keypad permutation, the router or classifier detected my intention and flat-out rejected my request. And all of this points to just how difficult it is to actually classify properly and route the users request that front-to-lapse might need to adhere to going forward. Only though, Fable 5 is undoubtedly an impressive model across the board. But as much as a model is great, Fable 5 is currently included in the subsidized plan with

its own users' bar as you can see here. And after July 7, Fable will no longer be part of the regular subsidized plan, but rather, It is slated to be charged through the user's credit or API costs, which is not ideal for my case. And it will certainly be interesting to see how OpenAI subsidized plan for their GPD 5.6 model might pair up against and thropics subsidized plan or lack their off for the FV5 model.

The rumor on the street is that OpenAI figured out some wizard trickery when it comes to inference that helped them save costs for inference by a huge factor. If that is true, OpenAI could deliver in competing against the Fable 5 model, not only in performance, but in cost and speed in terms of tokens per second.