The argument in one line.
Claude Fable 5 sets a new performance ceiling for agentic coding, but at $18 per CursorBench task and a June 22 plan cutoff, the window to use it at subsidized cost is closing fast.
Read if. Skip if.
READ IF YOU ARE…
- You are building with Claude Code or an agentic product stack and want a ground-level report on Fable 5 real-world performance.
- You have a Pro, Max, Team, or eligible Enterprise Anthropic plan and have not decided whether to burn credits before June 22.
- You are tracking the CursorBench leaderboard and want context on the cost-vs-performance tradeoff at the frontier.
- You are debugging complex async, websocket, or agent bugs and want to know whether Fable 5 extended thinking is worth the price.
SKIP IF…
- You want a technical teardown — this is a first-impressions reaction video, not an engineering deep-dive.
- You are on a free or Starter Anthropic plan; the pricing and cutoff discussion will not apply to you yet.
The full version, fast.
Claude Fable 5 tops CursorBench 3.1 at 72.9% on Max mode but costs $18 per completed task versus $4.37 for GPT-5.5 Extra High at 64.3%. In practice, the model fixed three production bugs in under 30 minutes each across a 1,000+ commit codebase using minimal prompts, and every PR it generated scored 4-5/5 in Greptile automated review. The catch: from June 23 Fable 5 leaves Pro/Max/Team/Seed plans and requires usage credits. The video also covers the Fable/Mythos split (identical weights, different safeguard layers), system card highlights including a claim of 72 working days of output from two people in 16 hours, and a rebuttal to the take that AI making coding frictionless has removed meaning from the craft.
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 – 00:08
01 · Cold open
Personal rule: only covers model drops that feel like paradigm shifts. GPT-5.4 and Opus 4.8 did not qualify. Fable 5 does.
00:08 – 01:35
02 · Fable 5 vs Mythos 5 explained
Two models, identical weights. Fable 5 has safeguard classifiers; Mythos 5 has them lifted for vetted partners via Project Glaemring. Tweet: Anthropic overlords deciding which prompts the peasants are allowed to use.
01:35 – 02:40
03 · CursorBench 3.1 and pricing
Fable 5 Max: 72.9% / $18 per task. GPT-5.5 Extra High: 64.3% / $4.37. Pricing: $10 input / $50 output per million tokens.
02:40 – 03:52
04 · Live demo — Pluto mobile app
Sets up Claude Code on /effort ultracode to build a mobile companion app for Pluto using a pre-approved GPT-5.5 plan.
03:52 – 04:46
05 · Greptile sponsor
AI code reviewer that scores PRs 1-5. Anything below 4/5 not worth merging. Two-week free trial.
04:46 – 06:27
06 · Bug-fix receipts
Three production bugs fixed by Mythos: chat stream (17 min), computer-use agent (16 min), iMessage connection (28 min). Lazy prompts — just the error copy-pasted.
06:27 – 07:49
07 · Social proof — Greptile scores and Lovable clone
Every Mythos PR scores 4-5/5 in Greptile. Riley Brown one-shots a Lovable clone in 2 prompts using Mythos.
07:49 – 08:32
08 · Lovable demo and jailbreak
Side-by-side comparison of actual Lovable vs Riley's Mythos clone. Notes that Pliny has already jailbroken Fable 5.
08:32 – 09:41
09 · The rug pull — June 22 cutoff
Fable 5 free on plans until June 22. After that: usage credits. Anthropic compute-constrained, paying approximately $1B/month for GPUs.
09:41 – 11:05
10 · Response to Moe
Influencer Moe complains Fable 5 solving his problem too easily removed the meaning of the creation process. Pushback: you complained AI sucked, now AI is good and it is not even about the model.
11:05 – 12:08
11 · Pricing critique
Fable 5 is a rich man's model. Uncertainty about whether open source will close the gap. Advice: milk the $200/month plan before June 22.
12:08 – 13:22
12 · Mobile app check-in and system card
Dynamic workflow scaffolding underway in background. Used Fable to summarize its own system card highlights.
13:22 – 14:20
13 · System card highlights
72 working days from 2 people in 16 hours. Generalists+Mythos beat world-leading specialist teams. Hidden anti-distillation safeguard.
14:20 – 14:42
14 · Outro
Glass half full. Use the tools you have, make something happen. Standard subscribe CTA.
Lines worth screenshotting.
- Fable 5 Max costs $18 per CursorBench task to score 72.9% — four times the cost of GPT-5.5 Extra High which scores 64.3%.
- Three production bugs in a 1,000+ commit codebase were each fixed in under 30 minutes with a lazy, copy-paste prompt.
- Every PR Mythos generated scored 4 or 5 out of 5 in automated Greptile review — a bar prior models rarely hit.
- Fable 5 and Mythos 5 share identical weights; the difference is a safeguard classifier wrapped around Fable, not a different model.
- Anthropic removed Fable 5 from Pro/Max/Team/Seed plans on June 23 — usage credits required after that date.
- Anthropic estimated two people working 16 hours with Mythos produced the equivalent of 72 working days of work.
- Generalist biologists using Mythos 5 outperformed world-leading specialist teams in a tabletop biosecurity exercise.
- Fable 5 has an invisible anti-distillation safeguard to prevent competitors from training on its outputs.
- When a model finally does the hard thing well, some builders immediately shift to complaining it removed the journey — the goalposts move.
- The $200/month Anthropic plan is effectively subsidized compute for twelve more days; after June 22 every output token costs full rate.
The real cost of the best model is not the benchmark score.
WHAT TO LEARN
Fable 5 is demonstrably better at agentic coding than anything available, but better now comes with a price tag that makes the tool inaccessible outside a subsidized window.
- A model that scores highest on benchmarks can still be wrong for most builders if its per-task cost is four times the nearest competitor.
- Automated code review scores are a more reliable quality signal than self-reported impressions — objective proxies like Greptile confidence scores cut through the noise.
- Lazy prompts still work with capable models: three complex production bugs were fixed with minimal prompt engineering, just error messages copy-pasted verbatim.
- The 72-working-days-in-16-hours system card claim is the most actionable number in this video — it reframes what is reasonable to charge for work done with frontier tools.
- When a model is temporarily free on your plan, front-load highest-complexity tasks before the cutoff, not exploratory ones.
- Fable 5 invisible anti-distillation safeguard means open-source models cannot easily train on its outputs — the capability gap will not close the same way it did after GPT-4.
Terms worth knowing.
- CursorBench 3.1
- A benchmark from the Cursor team that evaluates AI agents on ambiguous multi-step tasks from real Cursor sessions. Reports both score and average cost per task, making it one of the few benchmarks that weights capability against spend.
- Mythos 5
- The unrestricted variant of Claude Fable 5, sharing identical model weights but with safety classifiers lifted in specific areas. Initially available only to vetted cybersecurity and infrastructure partners via Project Glaemring.
- Greptile
- An AI code review tool that runs on pull requests and returns a confidence score from 1 to 5. Used in this video as an objective quality proxy for code generated by different models.
- Usage credits
- Anthropic's pay-as-you-go billing layer that activates after subscription plan limits are exhausted or after a promotional free period ends. Fable 5 moves to this billing model on June 23 for Pro/Max/Team plans.
- Project Glaemring
- Anthropic's internal deployment program for Mythos 5, initially restricting access to vetted cyber defenders and infrastructure providers before broader rollout.
- Ultracode
- A Claude Code session mode (/effort ultra) that applies maximum effort and dynamic workflow orchestration, consuming more tokens per task in exchange for higher thoroughness.
Things they pointed at.
06:59channelRiley Brown on X and YouTube
09:41channelMoe (@atmoio on X)
12:08linkClaude Fable 5 and Mythos 5 System Card
Lines you could clip.
00:08
“After using it for twenty-four hours, almost blowing a thousand dollars on credits, I can comfortably tell you this is the next paradigm shift.”
08:59
“This almost feels like getting us high and addicted to the drug and taking it away from us.”
13:29
“Two people working sixteen hours produced the equivalent of seventy-two working days of work.”
11:31
“This genuinely feels like a rich man's model.”
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.
metaphoranalogy
I told myself I wouldn't do a model drop video unless it felt like the Opus four five moment we had in December to January. That's why the last video I did was GPT 5.4. Didn't do one on 5.5.
Didn't do one on Opus four eight. But Claude Fable five, mythos with the seat belt is different. After using it for twenty four hours, almost blowing a thousand dollars on credits, I can comfortably tell you this is the next paradigm shift, and that's what I wanna talk about.
We're gonna talk about the model, the pricing, the benchmarks, how I've been using it, why I think it's the best, the rug pull that's going to be happening very soon, and so much more. Let's get directly into it. Now you notice it's two model drops.
It's Claude Fable five and Claude Mythos five. Now you might be wondering, what is the difference? The difference is as follows.
For a small group of cyber defenders and infrastructure providers, we're also launching Claude Mythos five. It's the same underlying model as Fable five, but with safeguards lifted in some areas. So you and me get the safeguarded version.
So here's a tweet that best describes it. Are anthropic overlords deciding which prompts the peasants are allowed to use? But in all honesty, the way it works is it says here, we've therefore launched the model with safeguards that means queries on some topics will instead receive a response from our next most capable model, Claude Opus four eight.
Claude Opus four eight is not that capable, if I'm a be honest. So the way it's going to work is if you ask a question that triggers their safeguards, you're going to get a response from Opus four eight, which is basically a very much less capable model. Now benchmarks.
Almost all the benchmarks, Mythos is miles ahead, but I'm not really big fan of most benchmarks. I do like CursorBench by the Cursor team, and you can clearly see just off score alone, Mythos is ahead with a 72.9% on max, and then the closest second place is GPT five five on extra high with Opus four eight max being up there almost close.
In my opinion, I don't think Opus 4.8 max is as capable as extra high to each zone, but here is the stark difference, the pricing. For a task to be complete with Fable five max to get a 72.9% score, $18 was spent versus if you go to extra high for 64.3%, $4.37 was spent.
So this model is the best, but it's very, very expensive. When we get into pricing, we see with Fable five, you're paying $10 per million tokens input, and you're gonna be paying $50 per million tokens output.
Now I will say this. I thought it was gonna be a lot more experience. I thought we were gonna see 20 and a 100.
So 10 and 50, although it's expensive, I expected worse. One thing that I wanna do is I wanna have Claude Fable five run-in the background.
I have a plan for a mobile companion app for my app Pluto. Now if you don't know what Pluto is, Pluto is basically an AI agent platform for businesses. I got this little cool day and night toggle right here for light mode and dark mode.
But basically, what I wanna build is I want a mobile companion app, and I have a plan written here by g v t five five that I've approved that I want ready to be built, and I want Mythos to do it. So we're gonna go on Cloud Code. By the way, Cloud Code is still terrible.
Like, model Fable five is the best. Cloud Code is the worst. So I'm gonna go slash effort, and we're gonna go on ultra code.
We're just gonna burn money, baby. And then I'm going to say, please look at the plan folder in the root folder and go for the mobile folder where there's a plan on creating a mobile companion app. I want you to read the entire plan, read the code base, see if the plan is still up to date and not stale, and I want you to start building out the complete mobile app.
Do not make any mistakes.
Yo. Yo. No.
Wake up. You can't be sleeping on the job. Who's gonna review the code?
I hate code review, but it has to get done. And that's why I love GREPTL. See, I have a huge PR here on an open source project I have called Ralphie.
Over 4,341 lines done by this contributor. Now this is a lot of lines of code, and I'm probably going to miss something.
But the way I use Greptile in my code review is as follows. I have a Greptile review that's run on this PR. Mind you, this is a pretty large PR.
And what I love about Greptile is the summary it gives me plus the confidence score. The confidence score tells me whether this is worth my time reviewing or not. Anything below a four out of five is something I'm not going to personally review.
Imagine spending all that time only to find out this PR still requires work. But with the greptile, all I need to do is read the summary, review the confidence score. If it's under a four out of five, it's not worth merging.
Doing this has allowed me and my team to ship at warp speed. So make sure to check out Greptile. The link is in the description down below, or you can go to rossmike.link/greptile.
They have a two week free trial, and it's pretty generous. So take advantage of it. Believe me, once you start using it, you won't wanna stop.
Check them out. The link is down in the description below. Alright.
And then we're going to hit enter, and we're gonna let this run. So as this is working, let me show you some of the bugs that Mythos fixed for me.
So I'm gonna go back to cursor right here, and we're gonna close this tab, and I wanna go to, yes, chat streams tab. So let me zoom in here so you can see. I had a gnarly bug where the chat stream was just not working properly.
And to be honest, I did not know what the issue was. I tried with g p t five five. I tried with Opus for it.
Couldn't fix it at all. I gave it to Mythos, and I'll be honest, I gave a lazy prompt. I basically gave the error.
I copy pasted it, and I said I picked Pluto inference and half credits, but the chat didn't respond. I even said didn't response, and the WebSocket didn't work.
That was literally it. And it worked for seventeen minutes.
Now, I can't even imagine how much this cost. It worked for seventeen minutes, but guess what?
It fixed it. Another issue right here. I had an issue with the computer use agent not being able to work properly.
Again, a lazy prompt, screenshots. Right?
I have screenshots, and then I basically was like, fix it. It worked for sixteen minutes.
And guess what? It fixed it. Last one, the iMessage connection.
It wasn't working. I would get delivered, but messages were not being sent. It worked for twenty eight minutes.
And guess what? It fixed it. Mind you, Pluto is a pretty large code base, over a thousand commits.
And under thirty minutes in all cases, it fixed the issues. Another thing I noticed with Mythos, my code review agent of choice, GREBTILE, almost loves it.
Every single GREBTILE review I've got is a five out of five, a five out of five, a five out of five, and again, I picked five at random, a five out of five.
I saw maybe one four out of five, but almost every PR that Mythos has made is getting a four out of five or a five out of five with my Greptile review agent. The code quality is a lot higher compared to every other model. A good friend of mine, Riley, who I'm pretty sure you're subscribed to.
If not, you should follow him on Twitter and subscribe to his channel, one shotted a lovable clone. Now we talk about one shotting Airbnb as a page and all that stuff.
This app is using Convex as a back end, Daytona as a sandbox provider, and is basically deploying apps like Lovable.
And he did it all in one prompt. I just want you to see What happens if I wanna His reaction. Doing this And now I'm going to make a Notion like notes app and make it dark mode.
Okay. So this is my version of Lovable that I've hive coded, and here's actual Lovable. I'm gonna enter these prompts at the same time.
Look at this. Mine's faster. This is actually running in a sandbox.
This is
literally lovable. It's lit I wanna correct two prompts. So what?
What if even it's 10 prompts? The fact that you can do this with Mythos is incredible. Literally lovable.
That like thing.
Yes. There's a little extra polish, but, like, what are we doing?
What are we doing here? Okay. Starting the web server.
So this one was about seven seconds before mine, but mine's probably using a better model low key. Alright. So let's open it on the actual Lovable.
And there you go. We have this little notes app. Now I'm gonna open it up in my version of Lovable.
Bang. So create a new page. Look at this.
Deployed on the web. Two prompts, he has a lovable clone. So when I look at why we haven't received mythos, but we got fable, mythos with the seat belt, I understand.
But what's funny is you still have people like Pliny who still jailbreak fable five and has got it to answer some very dark and questionable questions. Right?
So although the guardrails are important and safety is important, there's guys like him who can bypass these things. Now here's the rug pull. It says from today through June 22, Fable five is included on Pro Max team and seed based enterprise plans at no extra cost.
On June 23, we'll remove Fable five from those plans. Using it after that will require usage credits. So we have about thirteen days, twelve days before we get rug pulled.
And this almost feels like . So you have twelve days to use the best in class model with subsidized credits because if you have a 200 a month plan with Anthropic, then you're getting subsidized credits. And for those of you who use Claude CoWork from June 5 to July 5, your five hour usage limit in CoWork is doubled.
So, again, take advantage of these limits. Now to be fair, the reason why they have the cutoff date is because they are constrained for compute so much so that Anthropic is paying SpaceX money. I think almost $1,000,000,000 a month for GPU.
So I don't think they're doing it on purpose. I think they're really constrained. It is what it is.
Now, one last thing I wanna talk about is this influencer Moe. Moe is pretty popular on both on Twitter and YouTube and he's been popping off as sort of this AI doomer like everything is overhyped, things don't make sense and I'll be honest, I agreed with a lot of his takes but this take is just I'm I'm I'm sort of confused.
Listen to what he says. He says, yesterday, I signed up for the Claude Max $200 plan and had it change the whole visual metaphor of the productivity that I've been working on intermittently over the past year. Instead of traditional UI with tables list tools, I told Fable to use a desktop OS metaphor instead for displaying the various built in mini apps.
Right? Look look what he says here. Fable was able to solve the problem, but really I'm beyond the point of being impressed by an LLM doing some upfront task.
Everything worked. Made no mistakes. All tests passed.
It even fixed all tests, But I was like, okay. Whatever. Thanks.
I blew past my $200 limit in two hours. Now the pricing is a real issue. And now I'm sitting here like, okay.
Now what? Do I ship this? Hear me be a whiny b word for a second that it was so easy.
It killed the whole part of the journey of making an app where you become a new person through the creation process. I'm not gonna read the whole thing, but come on, brother. You were complaining this whole time that AI sucked.
Now AI is good, and now you have something else to complain about. And I'm not just trying to hate on Moe. I I completely understand if your whole identity is writing code then this hurts.
My identity is in Christ. Who the sun sets free is free indeed. I have joy regardless of what's going on in this space.
But I just can't help and fathom that some of these people a lot of people have been complaining that AI sucks. We finally have a model that's good and now it's not even about the model. All that being said, I have stuff to complain about as well and that is the price.
This genuinely feels like a . Right?
It feels like we're entering an era where state of the art models are for rich people, are for those who have money because this model is expensive. Even a $200 a month subscription is expensive. So honestly, I I don't know where this goes.
Will, you know, the open source models catch up and will pricing go down with the constraints on hardware and how expensive GPUs and SSDs are? I'm not really too sure. I don't know.
I'm not here to predict the future, but I'm here to tell you Fable five is the best model, but it really is expensive. If you have the $200 a month subscription, milk the heck out of it until June 22.
Ship all those side projects. Use it. Have fun.
Learn. Now checking back at our mobile app deployment, we can see the dynamic workflow is full intact, that the code base is being analyzed, and now it's scaffolding on apps mobile folder, and it's going to continue on working.
Now I don't think I have enough time for this to finish. I wanna get this video out to you guys as soon as possible. So make sure to follow my Twitter.
The link is in the description for the update. And finally, the system card. Now Anthropic and many model providers release these system cards explaining how the model works.
It's a bajillion pages. I'm not about to read all that. I used Fable to analyze the Fable system card and to share with us some interesting facts.
A couple of things I found interesting, it says it's one model wearing two phases. Fable five and Mythos five share identical weights. Mythos five is the unrestricted version handed to a small set of vetted partners.
Fable five is the same model wrapped in safeguard classifier, so that's what you and I have. The bioweapons judgment was generally close. Anthropic classified Mythos five as a c b one but not CB two, calling it a much less clear judgment than for previous models.
This is again, all the nerd stuff aside, watch what it says here. Grades estimated that two people working sixteen hours produced the equivalent of seventy two working days of work when they used Mythos.
And not only that, it says in a tabletop exercise, teams of generalist biologists using Mythos five outperformed world leading specialist teams.
That is insane. Right? That a generalist with the model can outperform a specialist without it.
Another thing that I haven't seen promoted or talked about a lot is they have this hidden anti competitor safeguard. Unlike the visible classifiers, Fable five has an invisible competitive use safeguard meant to stop people using it to build frontier models.
So the deep seeks and all the open source models that distill or that use the new state of the art models by OpenAI Anthropic, they they, you know, they chat with it and sort of distill to train their own models. That's basically not possible.
But, yeah, that's pretty much it. What a time to be alive. We have access to great intelligence, but it's coming at a very heavy cost and we're soon going to be limited.
Is this the future we wanna live in? Does this excite you? I'm gonna look at everything from a glass half full perspective, use the tools that I got to my best, and make something happen.
At the end of the day, if God be for us, who be against us? Thank you so much for watching this video. I'll see you in the next one.
Make sure to like, comment, subscribe. Peace.
The Hook
The bait, then the rug-pull.
Ras Mic set a personal rule: no model-drop videos unless the release actually shifts the paradigm. Claude Fable 5 broke the rule — and nearly broke his API bill.
Named ideas worth stealing.
01:35model
CursorBench cost-vs-performance matrix
Plots benchmark score against average cost per task. Makes visible the tradeoff hidden in capability-only leaderboards.
Steal forAny comparison of AI tools where both cost and quality vary
03:52concept
Greptile confidence score gate
Any PR scoring below 4/5 is not worth merging. Simple threshold turns a score into a binary decision, removing subjective review time.
Steal forAny QA or review workflow where a fast binary gate replaces manual judgment
How they asked for the click.
VERBAL ASK
14:20subscribe
“Thank you so much for watching this video. I'll see you in the next one. Make sure to like, comment, subscribe. Peace.”
Standard verbal CTA at end, no visual card overlay. Low friction, no product pitch beyond Greptile sponsor mid-roll.
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.
