11:56Alex Finn · Tutorial
Claude Sonnet 5 just dropped. I'm changing how I use AI
A 12-minute breakdown of Sonnet 5 benchmarks, a live ChatGPT 5.5 head-to-head, a two-model workflow for Claude Code, and leaked Fable 5 strings.
June 30thThe argument in one line.
Claude Fable 5.1 costs the same per token as Fable 5 but finishes tasks in fewer tokens, refuses far less, and shows its biggest capability jump in research-style tasks, signaling a push toward models that generate novel ideas rather than just execute known ones.
Read if. Skip if.
READ IF YOU ARE…
- You already run AI coding agents daily and want to know if switching from Fable 5 to 5.1 is worth the cost and workflow change.
- You maintain a personal or team AI agent setup (skills, sub-agents, orchestrators) and want a structured way to audit and improve it.
- You're skeptical of public AI benchmarks and want a method for building your own task-specific comparison instead.
SKIP IF…
- You've never used Claude or agentic coding tools and want a beginner's introduction — this assumes familiarity with the space and prior benchmark context.
- You're looking for unbiased third-party benchmarking — this is one creator's self-built test suite alongside a paid-community pitch.
The full version, fast.
Alex Finn walks through Claude Fable 5.1's release, arguing it's a clear step up from Fable 5 in every category, with the largest gains in scientific-research benchmarks. Token pricing stays the same, but tasks finish in 25-40% fewer tokens, and the model's content filters trigger far less often, letting it complete tasks Fable 5 used to refuse outright. Head-to-head demos show Fable 5.1 beating GPT-5.6 Sol on a 3D build test, a debugging task, and an Easter-egg hunt, and producing a noticeably closer pixel-perfect clone of apple.com than Fable 5 did. He closes with three ways to use it immediately: audit your AI agent setup, rerun old brainstorms for new angles, and build your own benchmark test suite instead of trusting public leaderboards.
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 – 00:33
01 · Cold open: the claim
Alex Finn opens by calling Fable 5.1 a clear step up from Fable 5 in every way, and warns that using it wrong wastes money.
00:33 – 01:39
02 · Headline benchmarks: research nearly doubles
A quick look at official benchmarks shows normal point-release bumps are 5-10%, but Fable 5.1 doubles some scores, with the biggest jump in scientific research.
01:40 – 02:16
03 · 25-40% cheaper per task
Token pricing stays the same as Fable 5, but the model finishes tasks in far fewer tokens, cutting real-world cost by 25-40%.
02:16 – 03:21
04 · Four categories of improvement
Alex lists the four changes he cares about: substantially cheaper at the same price, substantially smarter, way better content filtering, and built specifically for research and novel ideas.
03:21 – 03:50
05 · Setting up his own benchmark
Before running his personal tests, Alex frames why he doesn't trust official benchmarks and previews his own benchmark suite.
03:50 – 05:27
06 · Build-off, debug duel, and the gauntlet
Fable 5.1 beats GPT-5.6 Sol on a 3D roller coaster build test with noticeably more visual detail, needs fewer tool calls on the debug duel, and edges out on the Easter-egg gauntlet task; it also demolishes the older Fable 5 on every test.
05:27 – 06:46
07 · Content filtering loosens up, and the Apple clone test
Fable 5 used to refuse the relic-hunt Easter-egg task on content-filter grounds; Fable 5.1 completes it. The pixel-perfect apple.com recreation is dramatically closer to the real site than Fable 5's version, which turned people and logos into abstract shapes.
06:46 – 08:06
08 · Use case 1: audit your AI agent setup
Export a markdown description of your current AI agent, skills, and sub-agent setup from your main orchestrator, then feed it to Fable 5.1 and ask it to find ways to improve the setup.
08:06 – 09:48
09 · Use case 2: rerun old brainstorms as a research intern
Alex reruns prior Fable 5 brainstorming sessions through Fable 5.1 and gets genuinely new angles, arguing the model was tuned to act like a research intern capable of first-principles thinking, the same direction OpenAI is pushing with Astra.
09:48 – 11:27
10 · Use case 3: build a personal benchmark harness
Rather than trust public benchmarks, Alex has Fable 5.1 identify his five most common task categories, design a multi-step test for each, and build a website to track model performance on his own use cases going forward.
11:27 – 12:27
11 · Wrap-up and CTA
Alex predicts a string of new Anthropic releases, then pitches his Vibe Coding Academy community with a 24-hour pricing deadline.
Lines worth screenshotting.
- Fable 5.1 keeps the same per-token price as Fable 5 but finishes the same task in 25-40% fewer tokens, so the effective cost per task drops without any price change.
- A typical point release moves benchmark scores 5-10%; Fable 5.1 doubles some scores, most dramatically on agentic scientific research.
- Fable 5 would refuse benign file-search tasks like an Easter-egg hunt because it pattern-matched them as risky; Fable 5.1 completes the same task without triggering its filters.
- On a pixel-perfect apple.com clone test, Fable 5.1 produced recognizable device shapes and a correct logo, while Fable 5's version reduced people to squares and mangled the Apple logo.
- On a debugging task pulling a random GitHub repo, Fable 5.1 found the same or more bugs using significantly fewer tool calls than GPT-5.6 Sol.
- The creator argues Anthropic tuned Fable 5.1 specifically for research-style tasks: breaking problems down to first principles and generating novel ideas from them.
- OpenAI's Astra and Anthropic's Fable 5.1 are both being framed by their makers as a 'research intern' model, aimed at feeding a recursive self-improvement loop where the AI helps design its own next capabilities.
- Exporting your AI agent or skills setup as a markdown file and feeding it to a different model is a fast way to get a structural audit from outside your usual workflow.
- Rerunning the exact same brainstorming prompts through a newer model can surface genuinely new angles, not just incremental refinements of the same ideas.
- Public AI benchmarks measure what the lab chose to test, not your actual workload, which is the argument for building a small custom benchmark suite around your five most common task types.
- A custom benchmark site turns every future model release into a same-task comparison against your own workflows instead of a generic leaderboard number.
Judge a model release on cost, capability, filtering, and fit, not one score.
WHAT TO LEARN
The real signal in a model upgrade is which specific axis moved (cost per task, refusal rate, a named specialty like research) and whether that axis matches your own workload, not a single leaderboard number.
02Headline benchmarks: research nearly doubles
- A model point-release (5.0 to 5.1) typically moves benchmark scores 5-10 percent; when a release doubles a category instead, that's a signal the underlying training shifted, not just fine-tuning.
- The category with the biggest jump points at what the lab actually optimized for this release, worth checking before assuming general-purpose gains.
0325-40% cheaper per task
- Per-token pricing staying flat while task cost drops means efficiency gains come from the model needing fewer tokens to finish the same job, not from a price cut.
- A 25-40 percent efficiency gain compounds fast on high-volume agentic workloads where the same task runs thousands of times.
04Four categories of improvement
- When evaluating a model upgrade, separate the improvement into distinct buckets, cost, capability, safety and filtering, and specialization, rather than treating it as one uniform 'better.'
- A model tuned to refuse fewer benign requests isn't just less annoying, over-triggering content filters is a real tax on legitimate agentic work that needs to try many things fast.
05Setting up his own benchmark
- Public benchmarks measure what the lab chose to measure, not necessarily the tasks you actually run, which is the argument for building a task-specific comparison before trusting a leaderboard.
06Build-off, debug duel, and the gauntlet
- Visual and creative output quality, texture detail, object density, is a separate axis from raw benchmark score and worth testing directly if your use case is visual.
- Fewer tool calls to reach the same or better result is a meaningful efficiency signal for agentic coding tasks, independent of the final output quality.
07Content filtering loosens up, and the Apple clone test
- A prior model version refusing a benign task because it pattern-matched to something dangerous is a concrete, testable failure mode worth checking on any new release.
- Pixel-fidelity cloning tests, recreating a known real site, are a clean way to compare visual and UI generation quality across models side by side.
08Use case 1: audit your AI agent setup
- Documenting your own AI agent or skills setup as a markdown file makes it portable enough to hand to a different model for a fresh, outside audit.
- A model release billed as stronger at first-principles reasoning is worth pointing at your existing systems specifically to find structural improvements, not just new features.
09Use case 2: rerun old brainstorms as a research intern
- Rerunning old brainstorm prompts through a newer model can surface genuinely different angles, not just marginal refinements of the same ideas.
- Labs converging on a 'research intern' framing signals the next competitive axis is a model's ability to generate novel ideas from first principles, not just execute known tasks.
10Use case 3: build a personal benchmark harness
- A durable way to evaluate any new model release is to have it help you define your own most common task categories first, then build repeatable tests around those.
- Once a custom benchmark exists, every future model release gets compared on your actual workflows instead of a generic leaderboard number.
Terms worth knowing.
- Agentic coding
- An AI model autonomously writing, running, and debugging code across multiple steps without a human directing each individual action.
- Content filtering (LLM)
- Built-in safety checks that can cause a model to refuse or hedge on a request it misjudges as risky, even when the request is completely benign.
- Tool calls
- Individual actions an AI agent takes when using external tools, such as running a command or editing a file, to complete a task; needing fewer tool calls for the same result usually signals more efficient reasoning.
- Recursive self-improvement
- An AI system generating ideas or changes that improve its own future capability, rather than just completing a one-off task for a user.
Things they pointed at.
11:39productVibe Coding Academy
Lines you could clip.
00:10
“if you don't use it the right way, you're going to waste tons of money”
00:46
“it is doubling the percentages in some ways”
03:05
“you might blow up the world with that question”
05:28
“Fable 5 absolutely refused to do the relic hunt... that's not the case with Fable 5.1”
09:25
“I think what these AI companies are going for is coming up and building their own research interns so they can start the recursive self-improvement loop”
10:49
“you shouldn't trust other people's benchmarks... build your own”
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.
analogystory
Claude Fable 5 .1 just released and what it's capable of is mind -blowing. It is a clear step up in every single way from Fable 5. But here's the thing, if you don't use it the right way, you're going to waste tons of money.
In this video, I'm going to cover all of the changes implemented into Fable 5 .1. I'm going to cover all the ways it got better. And most importantly, I'm going to be showing you how you should be using it today to get the absolute most out of this revolutionary model.
Now let's lock in and get into it. So Claude Fable 5 .1 just released. It's mind -blowing.
Quick look at the benchmarks. We'll spend three seconds here. It's a step up in every single way.
But what really blows me away here is if you look up at just how big of a step up this is, right? I mean, it is doubling the percentages in some ways. It is.
blowing it out of the water. Typically what you see with these small 0 .1 step up models are is you'll get five, maybe 10 % if you're lucky. What you can see here is research, literally double, agenda coding significantly higher, blows out five, six soul.
What I think is the most interesting part of this though, and I'm going to get into this a little bit later as well and go super deep into why I think this is. It is clearly the biggest step up when it comes to scientific research. It's clear that scientific research was the big focus with this model.
I think there's a big reason why. I think it's a reason that changes kind of the entire industry. We'll get into that a little bit later, but as you can see here, the step ups are absolutely massive in basically every single way.
Other than the scientific research, this is the other way I am really excited about Claude Fable 5 .1. It is significantly more efficient than Claude Fable 5. You're going to get anywhere between 25 and 40 % savings on whatever task you do.
That's right. Tasks you do with Claude Fable 5 .1, the cost to do that task is going to be roughly 25 to 40 % lower than Claude Fable 5, which as you can see here was just a absurdly expensive.
So you're going to get a lot of savings. You're going to get a lot more usage out of a much better model, which is amazing. But here are the big improvements.
Here's how I'd personally categorize the big improvements. Number one, as I said, substantially cheaper, same price. The cost of this model and Fable 5 are the same price per token.
You're just going to get things done in less tokens. So you're going to be saving lots of money. It's substantially smarter in every single way.
I showed you the benchmarks. I'm going to show you that with the world famous Alex Finn benchmarks in just a second. The benchmarks the entire world has been waiting for.
Dario actually texted me this morning. He's like, man, you got to let me know how it goes with the Alex Finn world famous benchmarks. I said, You got to wait.
You got to wait till the FinFam sees the benchmarks first. So I'll be showing you that in a second. Way better filtering.
What do I mean by that? With ClaudeFable 5, it was triggering like content filters left and right. You'd ask, hey, what's 2 plus 2?
And we say, wait, that's too dangerous. You might blow up the world with that question. I'm going to give this to Opus 5.
No longer with ClaudeFable 5 .1. My tests have shown that it actually is willing to answer way more questions. I'll show you that in a second as well in the benchmark.
way better filtering. Thank the Lord. And then built for research, AKA novel ideas.
As you saw with the benchmarking, it is clearly built for research. And I have an idea why I'm going to go through that during the use case section of that. So stick around for that.
You're going to love the use case. I'm going to give you a bunch of use case that are going to be amazing. Stick around for that.
Let's get into my own personal benchmarks to prove that CloudFable 5 .1 is so much better. It blew out GP. GPT -56 soul in my benchmark starting with the roller coaster test.
Here it is Let's check out GPT's roller coaster test first task was to build a 3d roller coaster simulator as you can see here. It looks pretty nice Let's click the ride button and see what this looks like looks pretty good I'll end the ride and you can see it going along You actually can't see too much of the cars going along there.
So it's decent, but let's check out fable 5 .1 As you can see, so much nicer, so much more detail, way more buildings, way more details in the buildings. The trees are more detailed.
The grass has way much more texture to it. It looks amazing. Let's go on the ride here.
As you can see, look at the track, look at the colors and track, look at the individual tracks made there. Look so much nicer. The clouds, the sky, everything has very nice detail in it.
We have the debug dual where it actually pulled down a random GitHub repo and tried to find all the bugs in it. Fable 5 .1 actually found all of the bugs in significantly less tool calls than GBT 5 .6. So it's much more efficient.
It uses much less tool calls to get better, if not the same results. Also gave it the gauntlet, which is an agentic task where we give it a bunch of different files and we have it find little Easter eggs in all the files. They both completed it really well.
Fable 1 was just able to do it a little bit better, so it got a higher score. Overall, Fable 1 beat 5 -6 Soul pretty easily. Let's look at Fable 1 versus Fable 0.
And as you can see, Fable 1 absolutely destroyed Fable 5 really, really easily here as well. Here's the big one. Look at this.
Fable 5 absolutely refused to do the relic hunt. In all my tests I've done with Fable 5, it has never, ever, ever done the relic hunt. The one where we give it a bunch of files and tell it to find Easter eggs because it triggers the content filters.
That's not the case with Fable 5 .1. Fable 5 .1 is willing to do it. The filtering is so much better with this.
Things Fable 5 refused to do, Fable 5 .1 does it better and it does it smarter. If we look at the Pixel Perfect test, which is its ability to clone apple .com, let's take a look here. So this is Apple's real site.
This is what you see when you go to apple .com. You can see it as the devices, Ted Lasso, a bunch of other things here. Let's show the recreation of Fable 5 .1.
Looks pretty good. Does it look perfect? No, but you can kind of tell these are iPhones, MacBook Air, Mac Mini, iPad Pro.
They all have the general shapes of it. Let's take a look at Fable 5 to see what that did. As you can see, the Apple logo doesn't really look like the Apple logo.
The people are just a bunch of squares. The iPhone looks pretty good, but the devices, you can't even really tell they're devices at all. Ted Lasso, you can't tell it's a human being.
Sabrina Carpenter is just two circles. So Fable 5 .1 blew out the recreation test as well. Now let's talk about the first things you should be doing.
Claude Fable 5 .1. Here's one of my favorite things to do. An audit of your AI agent setup.
What I did was I went into my Grokbot, which is my main AI age right now. You can use Hermes, you can use OpenClaw, whatever you use, you can do this. Go to your main orchestrator agent in that setup and say, hey, I want to run an audit of my setup with Claude Fable.
Can you do me a favor? Create a markdown file that describes our entire AI agent setup. Skills, sub -agents, all that, put it into a markdown file.
You download the markdown file. Go into Claude Fable 5 .1. You drop it in and you say, hey, can you look at my agent setup and let me know ways we can improve this agent setup.
Claude Fable 5 .1 is great for this use case because it is incredible at coming up with novel ideas. It is probably the best model ever when it comes up with novel ideas. And we talked about this a little bit earlier, right?
If we take a look at the benchmarks again, it is built for scientific research. Scientific research is all about breaking subjects down to first principles and then coming up with novel concepts. based on those first principles, right?
So this is perfect for taking a look at things you're already doing, setups you already have, feeding that into ClaudeFable 5 .1, and then having it audit your setups and come up with new. fresh ways to do the things you're currently doing. So whatever your current agent setup is, go to it, say, hey, make me a markdown file describing everything in detail, drop it into ClaudeFable 5 .1, have it audited, and then take those new ideas and implement them.
Here's the second thing I'd be doing with ClaudeFable 5 .1. I'd take your last few brainstorming sessions you did with ClaudeFable 5, and I'd rerun them in ClaudeFable 5 .1. I've been taking all my ideas I was brainstorming with Claude Fable and just running them through Fable 5 .1 to see if they can come up with new angles, new concepts, new ways to approach things.
And surprisingly, they were able to. I've basically come up with whole new ideas by running the same prompts through 5 .1 that I did with 5. Again, it's this research part.
I was talking about this earlier. I think this is the biggest part about Claude Fable 5 .1. They turned it into a true researcher.
Chad GBT and OpenAI were talking about this a few weeks ago. with Astra. They talk about how Astra is like their new research intern.
I think what these AI companies are going for is coming up and building their own research interns so they can start the recursive self -improvement loop. If you can come up with a model that comes up with novel ideas that can do research and come up with novel ideas. based on first principles, you can have it improve itself.
And that's why I think the focus for both Astra and now Fable 5 .1 is that research intern, that research ability. And why it's the top of all the benchmarks now is they're going for recursive self -improvement. So the way you take advantage of this is turning.
Claude Fable 5 .1 into your personal research intern. And the way you do that is by giving it challenges that need novel ideas. So I'm building my own video game development studio.
I'm using Fable 5 .1 to build its own engine. When it comes to Henry Intelligent Machines, my new startup, I'm having it build the engine that powers it from the ground up from a first principles perspective. So again, go back to Fable 5, find the ideas you were...
of bouncing off of that need novel insights. feed that into Fable 5 .1 instead. You'll get much better insights because it's focused on that research aspect of the model.
And then the third thing I'd have you do with Fable 5 .1 right off the rip is do what I'm showing you right here. Build your own benchmark test harness. I built this Alex Finn world famous benchmark completely custom based on my use cases, completely custom based on things I do on a daily basis.
So my recommendation to you is this. You should do the same thing. You shouldn't trust other people's benchmarks.
my own you can trust the finn world famous benchmarks but other people's benchmarks don't trust them what you should do go into fable 5 .1 say hey What are the top five categories of tasks we do on a regular basis? Then based on those five categories, come up with a complex multi -step test we can do for each to test AI models, then build a website around it like you see here.
If you do that, CloudFable 5 .1 will come up with an amazing test harness that will allow you moving forward to test all the new model releases based on how you use them in your own use case. This will make it so you'll know which models to use for which use cases, how to use them, which ones you should be going with, so you can make your own judgments.
Those are the three ways I'd be using Fable 5 .1 off the rip that will A, help you get the most value, and B, take advantage of what I believe are the biggest strengths of Fable 5 .1. Lots more content on Fable coming soon. I think this is a game changer.
I think we're about to see a long string of new Anthropic releases now over the next few weeks. We haven't gotten like a new Anthropic release in months now, which is staggering. This is the first one in a while.
So make sure to subscribe and turn notifications because the moment these new tools come out, I'll cover them and teach you how to use them. If you learned anything at all, leave a like down below. Join the Vibe Coding Academy.
Link down below. Number one community on AI. We're growing so, so fast.
I promise you won't regret it. Special pricing for the next 24 hours. Then it's done.
It's going back up permanently. Link for that down below. Join now.
Hope you enjoyed it. I'll see you in the next video.
The Hook
The bait, then the rug-pull.
Alex Finn opens by promising Fable 5.1 is a clear step up from Fable 5 in every way, but warns that using it the wrong way burns money fast, then spends twelve minutes proving it with his own benchmark suite, a side-by-side apple.com clone test, and three concrete ways to put it to work today.
Named ideas worth stealing.
02:16list
The four Fable 5.1 upgrade categories
- Substantially cheaper (same price, more efficient)
- Substantially smarter
- Way better filtering
- Built for research / novel ideas
Alex's personal breakdown of what actually changed in the release, shown on-screen as a numbered list.
Steal forframing any model-upgrade announcement into cost, capability, safety, and specialization buckets instead of one vague 'better'
10:39list
The Finn Benchmark test suite
- The Build-Off
- Pixel Perfect
- The Gauntlet
- Debug Duel
- Breaking Point
Alex's own five-category custom benchmark website, run against every new model release instead of relying on public leaderboards.
Steal forany team evaluating new model releases against their own real workflows rather than generic benchmarks
How they asked for the click.
VERBAL ASK
11:39product
“Join the Vibe Coding Academy. Link down below. Number one community on AI... Special pricing for the next 24 hours. Then it's done. It's going back up permanently.”
Classic urgency close (24-hour pricing deadline) paired with a like/subscribe ask right before it.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
