12:24Nate Herk | AI Automation · Talking Head
Is Claude Mythos Coming?
A 12-minute reality check on the Mythos API leak and why one buried Anthropic sentence beats all the hype.
June 6thThe argument in one line.
Claude Fable 5 and Mythos 5 are the same model split by safeguards -- Fable 5 is Mythos capability made safe for general use, available free on subscription only until June 22 before moving to pay-per-token billing.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code or Claude Pro/Max and want to know what changed on June 9, 2026.
- You are deciding whether to use Fable 5 before June 22 when it exits included subscription plans.
- You want to understand how Fable 5 and Mythos 5 relate to each other and to Opus 4.8.
- You follow AI model releases and want a plain-English read of the Anthropic blog without digging through the footnotes yourself.
SKIP IF…
- You need a hands-on benchmark comparison -- this is a blog walkthrough, not a live stress test.
- You are already reading the Anthropic research docs directly and do not need a summary layer.
The full version, fast.
Anthropic released Claude Fable 5 and Claude Mythos 5 simultaneously. Fable 5 is broadly available with cybersecurity safeguards enabled -- the same underlying model as Mythos 5, which has those safeguards lifted and is restricted to Glasswing partners. Both cost $10/million input and $50/million output tokens -- double Opus pricing. Fable 5 is included in Pro/Max/Team subscriptions until June 22; after that it moves to usage-credit billing. The host had predicted months earlier that a public Mythos button would not appear, and the Fable 5 launch confirmed that read.
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 – 00:44
01 · What just dropped
Fable 5 vs Mythos 5 availability split, pricing, and Glasswing restriction announced.
00:44 – 01:57
02 · Subscription cutoff: use it now
Free on Pro/Max/Team until June 22; after that, usage credits only. How to access Fable 5 in Claude Code today.
01:58 – 02:27
03 · Mythos class explained
Footnote definition: Mythos class sits above Opus in the model hierarchy; Preview was released in April via Glasswing.
02:27 – 03:40
04 · Same model, different safeguards
Fable 5 and Mythos 5 are identical weights; the split is regulatory. Host plays his earlier prediction clip -- it aged well.
03:40 – 05:51
05 · Benchmarks and capabilities
State-of-the-art claims on knowledge work, agentic coding, cybersecurity. Glasswing and US government deployment context.
05:51 – 06:58
06 · Agentic loops: hype vs reality
Models can run autonomously longer -- but model makers profit from that framing. Token-cost warning on open-ended agent chains.
06:58 – 07:57
07 · FrontierCode and vision
SWEbench skepticism. FrontierCode shows a real agentic coding jump. Vision improvements matter for verification loops.
07:57 – 08:41
08 · Closing
Reminder to use Fable 5 before June 22. More hands-on content coming once the host has more time with the model.
Lines worth screenshotting.
- Fable 5 and Mythos 5 are the same model under the hood; the only difference is Mythos 5 has cybersecurity safeguards removed for vetted Glasswing partners.
- Mythos class sits above Opus in the Claude hierarchy; Fable 5 is the consumer-safe entry point into that tier.
- At $10/$50 per million tokens, Fable 5 costs exactly twice Opus 4.8 -- yet the earlier Glasswing Mythos Preview was five times Opus pricing.
- Fable 5 is included in Pro/Max/Team/Enterprise subscriptions only until June 22, 2026; after that it requires usage credits.
- The Mythos Preview being five times Opus pricing while Fable 5 is only two times suggests the public model is a constrained version, not the full Mythos capability.
- Agentic loop advocates want you to run agents that prompt agents -- but model makers have a financial incentive to push that narrative since longer loops burn more tokens.
- Most knowledge-work tasks do not benefit from perpetual agent loops; open-ended prompts with no exit condition mostly hit session limits before completing.
- Opus 4.8 on extra-high effort is likely comparable to Fable 5 on low effort -- a useful cost calibration for heavy reasoning tasks.
- The host dismissed SWEbench Pro after deep-SWE analysis debunked its methodology; FrontierCode is his preferred signal for agentic coding performance.
- Anthropic filed to go public while still unprofitable, selling compute at below-cost prices -- the June 22 billing change is a direct consequence.
- Vision improvements in Fable 5 are practically significant for agentic verification loops where the model checks its own rendered output without a human.
- Mythos 5 is deployed exclusively through Project Glasswing in collaboration with the US government as a cybersecurity tool.
What Fable 5 actually is and what to do before June 22
WHAT TO LEARN
Fable 5 is Mythos capability with the safety rails on -- and you have until June 22 to use it inside your existing subscription before it moves to pay-per-token billing.
01What just dropped
- Fable 5 is broadly available; Mythos 5 goes only to Glasswing partners -- the headline obscures that split.
- Both models price at $10/$50 per million tokens, double Opus 4.8, which sets the baseline for any cost-benefit calculation.
02Subscription cutoff: use it now
- Free Fable 5 access ends June 22; after that, usage credits are the only path -- treat that window as a structured evaluation period, not an infinite free trial.
- Access Fable 5 in Claude Code with /model after updating; the switch is immediate once the client is current.
03Mythos class explained
- Mythos class is a formal tier above Opus, not marketing language -- it has a footnote in the Anthropic blog defining it precisely.
- The tier was introduced quietly via Glasswing Preview in April 2026, months before any general public access.
04Same model, different safeguards
- The Mythos Preview costing five times Opus while Fable 5 costs two times Opus is the clearest signal that the public model is a constrained version, not the ceiling.
- When someone predicted a model outcome months in advance without insider information, the methodology worth studying is the reasoning chain, not the conclusion.
05Benchmarks and capabilities
- Benchmark claims from model makers are worth reading with a calibrated skepticism -- state-of-the-art assertions shift every few months.
- The cybersecurity capability split between Fable and Mythos is not incidental; it reflects a real policy decision about what capabilities can be deployed without government oversight.
06Agentic loops: hype vs reality
- Model makers benefit financially when users believe every task requires an agentic loop -- that conflict of interest is worth naming before designing any multi-agent pipeline.
- Open-ended agent prompts with no exit condition are more likely to hit a session limit than produce a finished result; scope the task before choosing the architecture.
- Fable 5 at low effort likely approximates Opus 4.8 at extra-high -- knowing that calibration prevents overspending on effort level before the task warrants it.
07FrontierCode and vision
- SWEbench Pro has been credibly challenged as a benchmark; discount it when evaluating agentic coding claims and look for FrontierCode results instead.
- Vision improvements in Fable 5 have the most compounding value in agentic workflows where the model reviews its own output -- that removes one of the most common human-in-the-loop bottlenecks.
Terms worth knowing.
- Fable 5
- Anthropic's broadly available Mythos class model, launched June 9, 2026, with cybersecurity safeguards enabled. It shares the same underlying weights as Mythos 5 but is safe for general consumer use.
- Mythos 5
- The same model as Fable 5 but with cybersecurity safeguards lifted; restricted to vetted Project Glasswing partners including US government agencies and critical infrastructure providers.
- Mythos class
- Anthropic's internal model tier sitting above Opus in capability, introduced with Claude Mythos Preview in April 2026 via Project Glasswing.
- Project Glasswing
- Anthropic's restricted-access program that gives select cybersecurity researchers and government partners early access to Mythos-tier models before any public release.
- Usage credits
- Token-based billing on Anthropic subscription plans, charged separately from the flat monthly fee; becomes the only way to access Fable 5 after June 22, 2026.
- FrontierCode
- A benchmark measuring high-quality, maintainable agentic coding ability. The host considers it more credible than SWEbench Pro for evaluating real-world coding performance.
- Agentic loop
- A pipeline where one AI agent's output becomes the input prompt for another agent, enabling multi-step autonomous work without human checkpoints between steps.
- Effort levels
- Fable 5's reasoning-depth settings (low, medium, high, x-high, max), analogous to extended thinking in Opus 4.8 but with more granular cost and quality tradeoffs.
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.
00:00Wow. Claude Mythos is finally here. Or is it really?
00:03We've got Fable five, and we've got Mythos five. And if you go to the very bottom of this page, you see that Claude Fable five is available everywhere today, while Mythos five is restricted to Glasswing partners only. Pricing for both models is $10 per million input tokens and $50 per million output tokens, which is double the cost of OPUS.
00:19And I know you guys already complained that OPUS is super expensive, but, I mean, really, if this is as capable as we think it is, in the grand scheme of things, this is not that expensive. They also expect Fable demand to be very high. Obviously, yeah, because they've been teasing it up and, you know, hyping it up for a long time now.
00:34But what you'll notice here is right now, from today until June 22, Fable five is included on pro, max, team, and seat based enterprise plans at no extra cost, so just part of your subscription. But on June 23, they will be removing Fable five from those plans. So if you wanna use it, you're gonna have to do usage credits, which will be token based.
00:53After this point, when sufficient capacity allows us to do so, we aim to restore Fable five as a standard part of subscription plans. We intend to do this as quickly as we can. I mean, they just filed to go public.
01:04They're obviously not profitable. They have so much compute costs, and they're giving away their models for, honestly, much cheaper than it costs for them to actually run them. So I understand that, but that is something to keep in mind.
01:14And you definitely are gonna wanna play around with these models as much as you can, and you're definitely gonna wanna play around with Fable five as much as you can from today until June 22. So we've got, like, two weeks. So, yes, you can do it right now.
01:25You can see it says up here, meet Fable five, our newest model for complex, long running work. Try anytime with slash model. You're gonna have to make sure you update, though, if you don't see it.
01:33So right here, can see I have Fable five. But otherwise, just go ahead and update your Claude code, and then you can actually go ahead and use it. And let's just see how it works.
01:42Yeah. I mean, that is freaking AGI. The first thing you see in this blog is that it says, today, we're launching Claude Fable five, a mythos class model that we've made safe for general use.
01:51Mythos class. If we see the little one here, we go all the way down to see what that little one means. That little one means Mythos class models are a tier of Claude models that sit right above Opus and capability.
02:03The first one, which was Claude Mythos preview, was released in April through project Glasswing. That is today followed by Claude Fable five and Claude Mythos five. So they're two different models.
02:13This one is obviously more powerful than this one. All users who currently have access to the Mythos Preview through Project Glasswing will be able to upgrade to Claude Mythos five, the same model as Claude Fable five, but with cyber safeguards lifted. If you guys remember during the project glasswing announcement, that Claude Mythos preview was actually five times as expensive as Opus, but this one is only two times as expensive as Opus.
02:36So we're clearly not getting the full model capability, which is actually what I pretty much predicted in this video where I said is Claude Mythos coming. I got a little bit clowned for this, but let's just watch this clip real quick. Is Claude Mythos coming?
02:47Alright. Here's where I land on it. And, obviously, I could be completely wrong.
02:50I don't have any insider information. A public model that you can just log in and use called Mythos, I don't think that's happening anytime soon.
02:57And it might not ever happen under that name, at least not the way that we're picturing it right now. I mean, they kind of told us the plan. A leak doesn't change the plan.
03:06So, I mean, everyone waiting for a big Mythos button to show up in the app next week, I think they're going to be waiting a while. So what do I actually think happens? I think the capabilities start to show up, but kind of quieter.
03:17You know, I think the stuff that makes Mythos Mythos, the stuff that makes it so special, those features sort of start to get baked into future versions of Opus.
03:26And once Anthropic decides the safety guardrails are solid enough, then that's when they we start to see some more stuff.
03:33Right? Okay. So, I mean, that could age worse over time, but, I mean, that was pretty spot on.
03:37We're basically getting a extended version of Opus called Fable, and it no one generally is just gonna be able to have access to Mythos right now. But as you can see, Fable five's capabilities do exceed any model they've ever generally released, so I'm very excited to get my hands on it.
03:52And you guys absolutely know I'm gonna be dropping a ton of content once I do. It's state of the art on nearly all tested benchmarks of AI capability, which I always think you should take with a grain of salt, but still important to look at. So exceptional performance in software engineering, knowledge work, vision, scientific research, and many other areas.
04:08Now, obviously, releasing a model like this has risks. And without safeguards in areas like cybersecurity, there could be people that use these models for malicious intent.
04:17I do think at the base, at the core, models are basically just they're they're neutral, and it depends on the human's judgment and the human's intent and use case to say, hey. Do you want to amplify your work for good reasons or for bad reasons?
04:30And so, obviously, they've launched this model with safeguards, and they've kind of toned this way down from Mythos, which is only going out to a small group of cyber defenders and infrastructure providers. Mythos five will initially be deployed through project Glasswing in collaboration with the US government as an upgrade to Claude Mythos Preview.
04:48So it's better than Claude Mythos Preview when this is the one, just to remind you guys, that was five times as expensive as Opus. It has the strongest cybersecurity capabilities of any model in the world, but they are going to be extending through Glasswing to a broader trust access program. Not exactly sure how you apply for that, but probably doesn't apply for the general consumer.
05:06And, of course, these models do have the capability to do a lot of good for the world because they're so so good at coding, which is obviously why they're good at cyber defense. But anything that is that good is also gonna be that good at breaking the that cyber code. So that's why they have to be careful.
05:20We take a look at some of these benchmarks here of Mythos five and Fable. So remember why we're looking at these two in the same, um, column here is because these are the exact same model under the hood. The only difference is that Fable five has all of those cyber safeguards lifted.
05:33So same reasoning and probably same cost, but the cyber safeguards have been lifted for Mythos five. Either way, you can see that pretty much every benchmark here is much better. The knowledge work took a big jump.
05:43Agenetic coding took a big jump when you're looking at comparison to stuff like OPUS 4.8 and GPT 5.5. Legal, multidisciplinary reasoning, biology, and then, of course, the big one is the cybersecurity.
05:55We've also been seeing a ton of people talking lately about agentic loops and basically saying, hey. You shouldn't prompt your agents anymore. You should have loops of agents that prompt your agents.
06:03And I I think that that's very true, and you can see here these models can work autonomously for much longer than previous cloud models. But think that that's also something you have to be very careful about because if the model makers are saying that, they obviously want you to, you know, kind of, like, use more tokens.
06:18And most things, especially general knowledge work, don't require agent loops that run all the time. And at that point, all you're doing is shooting off an open ended prompt and honestly just begging for your session limit to be hit by the time you come back. And if this thing is really as good as the benchmarks show in comparison to 4.8, which honestly so far, I don't do a ton of software, you know, engineering and, like, deep, deep programming, but I think OPUS four point eight's been really good.
06:39If it's this much better with these different reasoning models or sorry, with these different effort levels, then I'm really excited to play around with it. And this honestly is very accurate to me.
06:48You know, when I play with OPUS 4.8 with max reasoning, I honestly don't see much of a difference between x high and max, so this makes sense. But looks like they've also done a bump here to make the max effort, obviously, most expensive, a little bit of an improvement over x high.
07:01So if you look at right now, the way that you're using Opus 4.8 on x high, that's probably similar to how Fable's gonna be on low, and Fable on low is also gonna be a little bit cheaper. So as far as being able to try to give it a reference, that's what we've got.
07:14Agented coding, you can see here on the SW bench pro, which I don't even honestly look at anymore. I don't know if you guys have seen the deep SWE, but basically just debunked this whole thing.
07:24So I don't even look at this one, like, at all anymore. But the frontier code, that's a pretty noticeable jump. Fable five shows strong performance on complex analytical tasks, which is, you know, knowledge work.
07:34And I'm very excited for that because when I use my, you know, Agentsic OS, my AOS inside of Cloud Code, using Fable five for that and doing all my knowledge work, I'm really excited to see how good it really is. Also, Cloud models in general have been getting so much better with vision, which is insanely important, especially for verification loops.
07:51So now being able to use Fable and, like, know, I'm gonna try editing videos with it, making slide decks, and verifying, building websites, I'm excited to see how much better it is with Vision. So, obviously, on this blog, you can click through a bunch of these different use cases and different experiments that they've run with it.
08:05I'm not just here trying to regurgitate this blog for you guys because you guys can all do that. I thought that I would just share some of the things that I'm excited by and some of the things I'd be thinking about here. But, anyways, like I said, get in there.
08:16Start playing around with stuff. I'll be making more videos once I get my hands on it a little bit more. But remember that this will be no longer available through the subscription after June 22.
08:26They will be pushing to bring it back, but we don't know exactly when. So just keep your eyes peeled on x or my YouTube channel. So, anyways, I know this one was quick, but hope you guys enjoyed the video.
08:36And if you did, please give it a like. Appreciate you guys making it to the end, and I'll see you in the next one. Thanks, guys.
The Hook
The bait, then the rug-pull.
The Anthropic blog dropped Claude Fable 5 and Claude Mythos 5 on the same day -- but the asterisk buried at the bottom of the page tells the real story: one model is available everywhere, the other goes only to Glasswing partners. Nate Herk walks through what each model actually is, plays a months-old clip that called this outcome, and explains why you have about two weeks to use Fable 5 before the billing structure changes.
