14:42Ras Mic · Talking Head
I spent $1,000 on Claude Fable 5 (Mythos), it's incredible
A 14-minute first-impressions report on the best coding model available — and the 12-day window before it stops being free.
June 11thThe argument in one line.
Claude Fable 5 delivers a genuine capability step-change over Opus 4.8 — finding security flaws Opus missed and producing more critical research — but Anthropic is simultaneously walling it off from agentic systems and pricing it out of daily use, marking the start of a frontier AI premium tier.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude through a paid subscription or API and want to know whether upgrading your model tier changes outcomes, not just speed.
- You run an agentic setup (OpenClaw, Claude Code, Hermes, or similar) and need to understand how Anthropic is changing access rules.
- You do competitive research, security audits, or strategic analysis where model quality materially affects the output.
- You track the cost trajectory of frontier AI and want a practitioner's view, not a press release summary.
SKIP IF…
- You're looking for a hands-on tutorial on how to set up Claude or any agentic system — this is a review, not a guide.
- You only use free-tier AI tools and pricing changes at the -/month level don't affect your decisions.
The full version, fast.
Fable 5 tops every major benchmark by a significant margin and, in practice, finds authorization flaws in a codebase that a well-prompted Opus 4.8 cleared, and produces more critical, less flattering competitive research. The catch: it costs twice the API price of Opus, was available free on Claude subscriptions for only 12 days post-launch, and Anthropic is simultaneously restricting its use inside agentic systems like OpenClaw. The pattern across Anthropic, Google, and OpenAI is consistent — the most capable model tier is moving to /month and above, with no ceiling yet in sight.
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 – 00:34
01 · Cold open — why he waited
Igor explains the delay: wanted real data and community results before publishing.
00:34 – 01:22
02 · Benchmarks — head and shoulders above
SWE-bench Pro: Fable 5 at 80%+, Opus 4.8 at 69%, GPT 5.5 at 58%. Reddit consensus: mature, calm, down-to-earth programmer.
01:22 – 03:00
03 · Cost and access window
Available on Claude subscriptions for only 12 days. API costs twice Opus. Anthropic blocking agentic use on subscriptions. 'AI capitalism' framing.
03:00 – 05:13
04 · Security audit — Alfredo Hub
Ran same audit prompt on Opus 4.8 (B minus) then Fable 5 (C minus, after Opus issues fixed). Fable found authorization inconsistencies Opus missed.
05:13 – 07:08
05 · Community demos — games and websites
Pokemon clone one-shot. 3D space shooter with AI-generated soundtrack. 3D Delhi map (/1.5M tokens). Victor's tutorial website.
07:08 – 08:30
06 · Death Star test + guardrails
Death Star above LA is most aesthetically impressive yet, but guardrails blocked exact Star Wars replication — produced a generic megastructure instead.
08:30 – 11:09
07 · Deep research — Fable vs Opus head-to-head
Same research prompt in both models. Opus evaluated Fable's output as the stronger strategic document. Fable weighted stale community size data more literally.
11:09 – 12:30
08 · Verdict and cost reality
Twice the cost. Claude Code usage doubled as sweetener. Frontier AI trending expensive.
12:30 – 13:17
09 · Apple Intelligence + Gemini
Siri partnering with Google Gemini. Developer beta in July. Igor skeptical based on Apple Intelligence track record.
13:17 – 14:42
10 · NotebookLM agentic upgrade
Ultra plan (/mo) adds VM-backed agent, internet research, extended reasoning loop.
14:42 – 15:10
11 · Wrap
Recommends running a cybersecurity audit on any production app. Sign-off from Igor Pogany.
Lines worth screenshotting.
- Running Fable 5 on a security audit that Opus 4.8 passed found a new set of authorization flaws — deeper reasoning produces qualitatively different results, not just faster ones.
- Fable rated a platform C minus after Opus rated it B minus — and that C minus came after the B minus issues had already been fixed.
- Asking one model to judge another model's output on the same task is a practical benchmark: here, Opus called Fable's research report the stronger strategic document.
- Fable weighted stale context more literally than Opus did — ranking a business fifth instead of first based on an outdated member count — which is more accurate, not less useful.
- Claude Fable 5 scored 80%+ on SWE-bench Pro versus GPT 5.5 at 58% and Opus 4.8 at 69% — the gap is large enough to be practically meaningful, not just a benchmark artifact.
- The 12-day free window on Claude subscriptions was a deliberate taste-test, not generosity — the model moves to API-only billing after that window.
- Anthropic is also prohibiting the use of Claude subscriptions with agentic systems like OpenClaw and Hermes — the free-tier agentic era is ending.
- Two years ago the highest-tier AI plan was /month; now the plan does not include the best available model.
- The Pokemon clone one-shot demo and the 3D space shooter with soundtrack are useful calibration points: the model can code complex interactive systems in a single prompt.
- A 3D map of Delhi built with Fable cost at 1.5 million tokens — expensive for a demo, but a data point for what is now one-prompt achievable.
- NotebookLM's agentic upgrade (VM, internet research, extended reasoning) is locked to the Google Ultra plan — the premium-tier convergence is happening across every major AI provider simultaneously.
- Igor's Alfredo hub runs ~50 cron jobs and multiple agentic loops — his security audit use case is a practical template for anyone with a production AI-integrated system.
What Fable 5 actually changes, and what it costs
WHAT TO LEARN
Fable 5 is not an incremental update — it finds problems that a well-prompted Opus missed, produces more critical research, and reasons at a level that changes real decisions.
- Running Fable on the same security audit prompt that Opus passed found new authorization flaws — deeper reasoning produces qualitatively different results, not just faster ones.
- Fable weighted stale context more literally than Opus did, ranking a business fifth instead of first based on an outdated member count — which is more accurate, even if less flattering.
- Asking one model to evaluate another model's output on the same task is a practical benchmark you can run for free — Opus called Fable's research report the stronger strategic document.
- The 12-day free window on Claude subscriptions is a taste-test, not a gift — the model moves to API-only billing after that window closes.
- At twice the API cost of Opus 4.8, Fable 5 is not a daily-driver for most workflows — the right use case is high-stakes audits and strategic research where quality matters more than cost.
- NotebookLM's VM-backed agentic upgrade, Google's Gemini integration with Siri, and Anthropic's Fable pricing all landed in the same week — the /month frontier tier is converging across every major provider simultaneously.
Terms worth knowing.
- SWE-bench Pro
- A benchmark that tests AI models on real-world software engineering tasks — fixing bugs and implementing features in production codebases — used by model makers as a standard capability signal.
- Agentic system
- An AI setup where the model takes multi-step actions autonomously — browsing the web, writing and running code, operating other tools — rather than just answering a single prompt.
- OpenClaw / Hermes
- Third-party tools that connect Claude models to agentic workflows, allowing them to run as autonomous agents. Anthropic is restricting their use on standard subscriptions.
- Prompt caching
- A feature that stores frequently-used portions of a prompt so they do not need to be re-processed each call, cutting input token costs by up to 90% on cache hits.
- Claude Code
- Anthropic's official CLI coding agent that operates inside a terminal and can read, write, and execute code across a codebase.
Things they pointed at.
03:00toolOpenClaw (agentic Claude runner)
03:00toolHermes (agentic framework)
10:17productAI Advantage community
Lines you could clip.
02:54
“Anthropic is really moving towards a version of AI capitalism, where you just need serious money to get the best models and agentic capabilities.”
04:55
“The first audit rated the platform as a B minus. The new audit rated it as a C minus. And the C minus was received after I fixed the issues from the first report.”
11:31
“Just two years ago, the most expensive plan you could get was a subscription. Now the plans are not even gonna include the best model available now.”
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.
metaphorstory
00:00Oh, boy. Boy. This week in AI was wild.
00:02We got the biggest model release of the year probably. Usually, in these model releases, I jump.
00:09Within twenty four hours, I'm there reporting on what's happening. Well, two things happened this week. First of all, I kinda looked at this and I was like, I want more information.
00:17I wanna see what people think, and I wanna actually run some of these test prompts that are gonna run for twelve hours and let other people do it because that's what this model does. And then two, I I also also just had a super busy week, I didn't wanna present you with, like, a half baked take. So here we are.
00:31We're gonna cover Fable five, which is Anthropic's biggest and, well, best model, especially on AgenTic and coding related tasks and sciences and actually kind of everything knowledge work. And then we're gonna talk about some other quick stories that came up this week that are relevant.
00:45But, honestly, main topic today, Fable five. It's finally out. It's available on paid plans, and let's dive into it.
00:52So first up, the high level. You probably heard it by now, but if you haven't, this is the best model in terms of benchmarks and user reviews across all of them now.
01:02And it's not by a little bit. Usually, these model releases, they inch up a little bit. This one, head and shoulders above.
01:08If you look at the benchmarks, it tells the story really well. Look. Especially agentic coding is one that I always look at and knowledge work.
01:14It's just not even close. GPT 5.5 at 58, Clawd Opus at 69. This is at over 80% on Software Engineering Bench Pro, an important one that usually all these model makers always publish.
01:26It's just the same story across everything else. And then on the other hand, there's the user experiences and the user reviews, and those are excellent too. I really like this Reddit takeover here, which says that Fable feels like a mature, calm, and down to earth programmer.
01:40It's very impressive. And in short, I'm gonna show you a bunch of examples and talk about my experience and what I saw on the Internet, but I agree. I like that it responds concisely and it just gets work done in a more thorough and in-depth way than anything else.
01:53The cost of that, well, it's literally the cost of the model. If you run it on your Claude Max subscription, the $200 subscription, you're gonna find that within like thirty minutes, you can easily fill up your entire usage for the day.
02:06And if you run it through the API, well, it just straight up costs twice the price of Claude Opus four per day today, which is already known for being notoriously expensive. So the big question, what can it do?
02:17What can you do with it would be worth your time? Because this thing is available within your Clot subscription, but only until June 22.
02:25If you go into your Clot subscription, you will find that, hey, they only gave this out for, what is it, like, twelve days from the release. And then you're gonna have to pay for the tokens directly through the API, which just makes me think in combination with the fact that they're prohibiting the use of your Anthropic subscription with AgenTic systems like OpenClaw or Hermes, which is happening also over the next few days.
02:48Anthropic is really moving towards a version of, like, I don't know, I would call it , where you just need serious money to get the best models and AgenTic capabilities. Anyway, as we have this window and everybody can access this on the paid plans, let's look at what you can actually do with it and what it performs better on.
03:07And there's a ton of examples on the Internet. I'm gonna show you some of the cool games and fun, interactive, and entertaining demos that people were creating. Wanna I start with the one that touched me the most and I thought was the most impressive.
03:19So I don't talk about this a lot on the channel. Actually, I've been thinking about how to talk about it more. But just because I don't make content on it doesn't mean I'm not all over of this agentic revolution that is happening.
03:30Right now, over there, I have a Mac mini running two versions of OpenClaw and a MacBook Pro running a Hermes agent, and then I also use Clawd Cowork a lot. Now what I tried with Fable five is running a security audit on the whole system and the platform that I built for our education company, AI Advantage, inside of OpenCloth.
03:50You might not be familiar, and that is fine, and it goes way beyond this video to guide you for this, but I basically build out this entire hub. Alfredo is my agent with all of these different sub applications for different parts of the business, and there's a lot there.
04:04There's apps. There's almost 50 cron jobs. There's a bunch of Vagintic Loops running too.
04:09But what I did is I ran a security audit on the entire Alfredo hub first with Opus 4.8. And I found some security audit prompt on the Internet that was popular in some Reddit forum. And, look, if you look into my Telegram chat from which I operate Alfredo most of the time, it did pretty well.
04:25It found some inconsistencies. It gave me some quick wins, and then I implemented right away. Fantastic.
04:31Now I switched the model over from Opus 4.8 to Fable five, and I ran the same audit. Look at that. Alfredo Hub Audit by Claude Fable five.
04:39Cool. It just found a whole new list. And in one sentence, I would say that the first audit rated the platform as a b minus in terms of security.
04:49The new audit rated it as a c minus. And the c minus was received after I fixed the issues from the first report. So I don't wanna go into the technicalities, but the different authorization levels that I had that were inconsistent.
05:02Inconsistent. I missed, Opus missed, and I'm so glad it found it because I have different access levels within the hub for different people in the company, and and it cleaned all of that up. And then there's all the other technical things it recommended.
05:13I've been implementing all that, and I'm like, this is just really useful to me. So that's my first take. Because people have been creating a bunch of games with this.
05:21Apparently, it's really good at coding and coding complex things like games and it demos well, so people have been doing that. Look. One shot prompt, make a Pokemon clone.
05:29And apparently, it one shots, well, this version of Pokemon where you can walk around and fight others. Impressive. If you follow the channel, you might know we also run this prompt of this three d space shooter game and just see how it looks in different models.
05:44I mean, do you hear the sound? It did a soundtrack. This is by far the most impressive one.
05:49Look at that. As we begin here, approach here, you can collect these. It got that whoop whoop.
05:55Okay. We're getting it. And then if you crash into something, it's over.
05:59Nope. You lose life. This is the most robust game out of all of them.
06:02Okay. Some more amazing stuff. So here, Kumar on Twitter basically built a three d map of Delhi.
06:09He said it costs around 1,500,000 tokens. So this map cost him $75.
06:14But look at that. It's pretty amazing. It's a three d map.
06:17Use various map libraries to actually create this. And then there's all the websites that people have been creating.
06:22Here's my favorite example of Victor sharing a twelve minute tutorial if you wanna check out exactly how he made it. Now this is not exactly a one shot website, meaning you give one prompt and it gives you an amazing result.
06:34He actually brought in visual elements that he then referenced and had quite a complex prompt with exactly what technology to use and what the outcome is supposed to look like. But look at these websites. I mean, this is next level stuff and all doable within a ten minute tutorial.
06:50If there's a thing to try yourself is making one of these insane websites or any website. It's just really, really good at coding. And what I always love seeing is that the sentiment across the Internet plus my own experiences that they align with also Reddit takes and Twitter comments and it just does.
07:04Like, there's a unified voice across the Internet of this thing being just amazing, especially when you're building things, improving things, creating websites, creating anything. This Reddit user puts it well that he used Fable for stuff he was working on today and it did a great job.
07:17But yesterday, used Opus and it also did a great job. It's hard to judge the relative quality when both models are smashing the tasks I give them, which is true. So it's not like you should be using Fable for everything from here now.
07:29They're both still great, but there is differences. I'll show you two more things. One of the test problems we always also run on these models, you might notice, is creating a death star above Los Angeles.
07:40And, yeah, I just have to say this is the most aesthetic and just objectively best one we have seen so far. The downtown skyline is accurate. You have the four zero five, little seagulls and the Death Star above it.
07:51But here's one note, and people have been reporting this too. Apparently, the guardrails on this are so tight that as the first model here, it had a problem with the idea of replicating the Death Star because that's a Star Wars themed thing and it didn't wanna go there. So it kinda did a version of it, I suppose.
08:07The detail of it is way higher, but you can see that this is not the exact Death Star from Star Wars, whereas some of the other models that ran it exactly recreated, well, the Death Star with, I don't know, this laser making ring that it has. Okay.
08:21One last example to run this out. So one thing that I like to do with these models is just accumulate a bunch of research and then sort through it. You might know the deep research functionality within ChatGPT, Cloud Gemini, it doesn't matter where, where you can basically go into a chat, you know, click plus, and then you say research, and it's just gonna accumulate, I don't know, a 102, 300 links.
08:42Now, I believe in chat, and this was the case a few months ago, but it didn't matter what model you picked. The research always used the same model. In Claude, what I found, it's different.
08:51It actually makes a difference if you select five or Fable five or Opus 4.8. What I did is run two quick researches that actually found helpful on the competitive landscape in the AI education marketplace where we operate. What I did is exactly what I teach.
09:05I ran this inside of my project with all of my context, and then I also enabled research. I didn't even need to fill in my market because all that context is in the project. And voila, I got one result from Fable and one result from Opus.
09:18Now I did the following thing. I copied the result from Fable and put it into the Opus chat and asked if you were to compare this report to a second version I have, what is the difference and which one is better? And it says, verdict for actually making decisions, the version you have, the Fable version, is the stronger strategic document.
09:34Mine is more faithful to the narrow brief you gave me and goes deeper on that specific niche, but that narrowness is also its main weakness. The honest move isn't to pick one. It's to fold a's depth into b's frame.
09:44Even Opus is like, there's different strengths, but I think overall, the Fable one, me reading through it and looking at these differences is superior in multiple ways. I'll just leave it at that without going into all the nuances where you would need context on our business to understand them.
09:58I'll say one thing though. There was a mistake in my context. I need to update one of the files there.
10:03That was still stating that our community is at was it 300 something members? That's information from last year. As we merged with Dean and Tony and built this new community, right now we're at over, and you might not know this, but this is insane, 70,000 users in the new community.
10:19These are paid subscribers to the new community that we run where I run the programs and teach the courses. If you wanna learn how to set up a project, there's a course on that in the community. You can check it out for $1.
10:30We have a fourteen day trial where you get everything in there. See if it's for you. Link in the description.
10:35But my point here is that it really took that number with the 300 something subscribers super seriously and it ranked us at fifth in the competitive landscape, whereas Opus took it less seriously and ranked us as number one. So I think I actually give that to Fable because it gave the critical context where it assumed that we're not able to deliver paid products that monetize and convert people and give deeper value from all our free offerings.
11:02It took that so much more seriously and it baked it into the report and was way more critical and honest, which I always appreciate. So another golden star to Fable. I think, of course, everything that I showed you here is just superior.
11:16It's twice as expensive and soon you won't be able to access this through the API, which as I pointed towards kind of shifts what's happening in AI. A lot of these best models are gonna become premium and then the competition is gonna catch up and China is gonna come out with an open source version, and then all the prices are gonna go down.
11:33But the frontier of AI has gradually been getting more and more expensive for consumers. Remember, just two years ago, the most expensive plan you could get was a $20 subscription.
11:42Now the $200 plans are not even gonna include the best model available now. So go try it, build a game, run some research reports, and let me know what you think in the comment section below.
11:53That is my take on Fable as of now, but that's on it. There were a few other stories this week that I found interesting. One of them is good news for all Claude users.
12:01Maybe they kind of did this to kind of soften the blow with Fable. They're essentially doubling your Claude CoWork usage for the next month. Simple as that.
12:09If you use CoWork, you can run more scheduled tasks to do more on your existing subscriptions. It goes for everything from $20,100, $200 subscription. And double the usage also applies to you having Fable for the next ten days.
12:22So if you wanna experiment, now there's a little sweet spot to double the usage with the best model ever, but, you know, this makes up for the fact that Fable uses double the tokens. They kind of cancel each other out if using it, but it's good to know about. And, hey, if you're enjoying this coverage, make sure to subscribe to the channel.
12:36It really helps us out. Now, see what's next. Then there were Apple announcements about an improved Siri.
12:42They're partnering with Google Gemini. And in July, we will receive the developer beta of what they actually cooked up.
12:50It's Apple. At this point, they're a few steps behind in AI, and I'm not gonna judge anything based on their announcements because they always sound good, then Apple Intelligence underperforms in practice. So I'm really curious to see this, to test this out for myself.
13:02I mean, everybody on planet Earth, everybody who interacts with tech can see that vision of, you know, Siri actually working just as well as Chad GPT or Claude. But then it has so much unique data and applications that use the the potential there is just unbelievable. And then finally, I wanted to tell you about an interesting update to NotebookLM because they made it agentic.
13:23Now this is unfortunately only on the Ultra plan, which is the $200 plan. Oh, boy. They're making AI expensive.
13:29It's happening, you guys. As you can see it right here. But on the 200 plan, now you can have NotebookLM basically run an agent in the background that has its own virtual machine, and it can code assets that it hasn't been able to do before.
13:41And it can also research the Internet and have a thinking loop that is more extensive than what it had. Remember, NotebookLM basically works in a way where you just add a bunch of sources, and then it just looks at those documents and uses a large language model to produce results from them. It can transform them into various outputs that they have on the right sidebar.
13:58Well, now it can turn it into even more outputs because there's a entire agent in the background that can do things for ultra subscribers, and it can find new sources by itself and kind of reason over everything and go more in-depth. They're really turning it into a research studio that is increasingly autonomous and flexible, but that has its price.
14:20And, yeah, there you go. That's pretty much everything I found really interesting for this week. I think it's really worth getting your hands on Fable and just getting a feeling for it.
14:27It writes well too, by the way, even basic writing. I didn't wanna go there cause that's so hard to judge and there was a lot of other interesting stuff, but even the short stories we wrote with it were like, that's really good. And it has a different flavor than Opus and GPT 5.5.
14:39And if you have any sort of app, just running a cybersecurity audit on it is sort of a no brainer thing that you should immediately do. Alright.
14:46There's your weekly roundup. My name is Igor, and I hope you have a wonderful week.
The Hook
The bait, then the rug-pull.
Igor Pogany waited two days before covering Fable 5 — unusual for a channel that normally ships within 24 hours of a model drop. His reason: he wanted real test results, not a hot take on a model that needs twelve-hour prompts to show what it can do.
Named ideas worth stealing.
09:34concept
Model-judges-model evaluation
Copy one model's output into a second model and ask it to compare and rank both reports. A practical zero-cost evaluation method that doesn't require human expert judgment.
Steal forAny situation where you want to compare two AI outputs without running a formal benchmark
04:16concept
Security audit with model swap
Run the same security audit prompt on a weaker model first, fix its findings, then re-run with a stronger model. The stronger model's second pass catches what the weaker model missed even after fixes.
Steal forAny production codebase, API, or agentic system with auth/access control
How they asked for the click.
VERBAL ASK
10:17product
“You can check it out for . We have a fourteen day trial where you get everything in there. Link in the description.”
Woven naturally into the deep research section where context about the AI Advantage community came up organically. Not a hard pitch — the trial offer follows from a real story point about community size.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
