20:45Matt Wolfe · Essay
Why Everyone Is Freaking Out About Mythos
A 20-minute reality check on Claude Fable 5 and Mythos 5 — what the hype got right, what it got wrong, and what the safety leash actually costs you.
June 11thThe argument in one line.
Fable 5.2 is the strongest model available and a real leap for very hard problems, but for everyday coding and knowledge work Opus 5.5 remains the better daily driver at half the price.
Read if. Skip if.
READ IF YOU ARE…
- A developer deciding whether Fable 5.2 is worth its higher per-token price versus Opus 5.5 for daily coding work.
- Someone building agent setups who wants to know how Anthropic suggests using Fable as an advisor or orchestrator over smaller models.
- A creator or tinkerer curious what the model can actually produce, from games to 3D-printable hardware designs, before spending usage on it.
SKIP IF…
- You want a rigorous benchmark analysis; the host says he puts less weight on benchmarks and tests informally on a few prompts.
- You need a step-by-step tutorial; this is a launch-day reaction and demo reel, not a how-to.
The full version, fast.
Anthropic released Fable 5.2, a guardrailed version of the cybersecurity-restricted Mythos 5.2, and it tops most benchmarks, with the biggest gains in scientific research and hard problems rather than everyday coding. It costs $8 per million input tokens and $40 per million output, double Opus 5.5, and the extra-high reasoning setting often beats max at lower cost. In Matt Wolfe's tests it built a mechanically flawless Megabonk-style game in 16 minutes and a full 3D-printable desk gadget with parts list, firmware and assembly manual. Early-access users showed games, app clones and code-made videos. His conclusion: use Opus 5.5 daily, reach for Fable on the hardest tasks.
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 – 01:58
01 · Launch and the two models
Fable 5.2 vs Mythos 5.2, and Anthropic's advice to keep using Opus 5.5 for most work.
01:58 – 04:30
02 · Benchmarks
Gains by category: modest in coding, large in scientific research and hard problems.
04:30 – 06:40
03 · Effort, cost and artificial analysis
Diminishing returns at max effort and the cost per task.
06:40 – 08:35
04 · Availability and pricing
Paid plans only, launch-week resets, and API pricing against Opus 5.5 and GPT-6 Astra.
08:35 – 09:08
05 · Anthropic's showcase
The 1,000 sub-agent operating system, protein folding and hardware designs.
09:08 – 12:05
06 · Megabonk-style game test
Refuses to clone the exact game, builds an original in 16 minutes, mechanics perfect, graphics so-so.
12:05 – 14:58
07 · Desk gadget hardware test
Parts list, firmware, STL files, and an assembly manual generated on a follow-up.
14:58 – 18:26
08 · Community demos
Gears of War clone, Zelda and Iron Man games, Photoshop clone, code-made travel ad, mirror-earth game.
18:26 – 20:47
09 · Verdict and sign-off
Huge for very hard problems, less so for knowledge work; use Opus 5.5 daily.
Lines worth screenshotting.
- Fable 5.2 is Mythos 5.2 with extra cybersecurity guardrails, which is why the public gets Fable while Mythos stays with security testers.
- Anthropic itself recommends Opus 5.5 as the daily driver and Fable only for the most ambitious open-ended technical work.
- On Frontier Code 2.0, Fable 5.2 scores better at extra high than at max, so letting it think longer made it both pricier and slightly worse.
- Fable 5.2 on high matched Opus 5.5 on max for coding while costing less, so a higher-tier model can be the cheaper option at a lower effort setting.
- Agentic scientific research jumped from 58.7% on Opus 5.5 to 74.6% on Fable 5.2, one of the few big leaps in the benchmark table.
- Typical coding benchmarks moved only a few points, such as 45.5% to 49.1% on Frontier Code, so most coding users will feel little change.
- Fable 5.2 API pricing is $8 input and $40 output per million tokens, double Opus 5.5 but cheaper than Fable 5.1 at $10 and $50.
- At max reasoning the average artificial-analysis task cost about $26.54, making it the most expensive model per task despite top intelligence.
- Anthropic says one engineer had Fable 5.2 coordinate over 1,000 sub-agents for eight days to build a bootable prototype operating system.
- Claude now refuses to clone a named game outright but will build an original game in the same genre without the protected IP.
- The same guardrail applies to real people: asked for a Gary Busey device it built an original robot character instead of his likeness.
- A full Megabonk-style game took about 16 minutes at medium reasoning and worked on the first try, though the graphics trailed Opus 5.5.
- Given one paragraph, the model produced a parts list, firmware, STL files for 3D printing and wiring steps, then a full assembly manual on a follow-up prompt.
Fable 5.2 is for hard problems, not daily work
WHAT TO LEARN
Fable 5.2 tops the benchmarks and costs double Opus 5.5, but the gains concentrate in very hard problems, so most people should keep Opus 5.5 as the default.
- Fable 5.2 is Mythos 5.2 with more cybersecurity guardrails, and Anthropic itself recommends starting with Opus 5.5 for most coding and knowledge work.
- Benchmark gains are small in everyday coding, a few points, but large in agentic scientific research, so match the model to the difficulty of the task.
- Higher reasoning effort has diminishing returns: extra high scored better than max on one coding benchmark while costing less.
- API pricing is $8 input and $40 output per million tokens, double Opus 5.5 yet cheaper than Fable 5.1 and GPT-6 Astra.
- Used as an advisor or orchestrator over smaller models, Fable can supply high-level direction while cheaper models do the routine work.
- Expect IP guardrails: the model builds original work in the same genre rather than cloning a named game or a real person's likeness.
- A single detailed prompt can now yield a parts list, firmware, printable files and an assembly manual, though a follow-up prompt may be needed for detail.
Terms worth knowing.
- Mythos 5.2
- The most capable model in this release, restricted mostly to cybersecurity testers. Fable 5.2 is essentially the same model with more cybersecurity guardrails.
- Orchestrator
- A model that holds the big picture and delegates simpler subtasks to smaller, cheaper models instead of doing everything itself.
- Reasoning effort
- A setting such as medium, high, extra high or max that controls how long the model thinks before answering, trading cost and time for quality.
- STL file
- A standard 3D model format that slicing software reads so a 3D printer can print the part.
- Sub-agent
- A separate model instance spun up by a main model to handle one piece of a larger task, such as writing or reviewing part of the code.
Things they pointed at.
01:30productClaude Opus 5.5
02:55productGemini 4 Carbon
07:00productGPT-6 Astra
13:30toolBambu Studio
16:40channelAlex's two-hour Fable 5.2 demo (linked in description)
Lines you could clip.
04:18
“Fable 5.2 actually scores better at extra high than it does at max.”
18:54
“This model is designed for really hard tasks. For most things, 5.5 is going to be just as good.”
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.
Well, it's another day, another best new LLM the world has ever seen. Today, Anthropic launched Clod Fable 5 .2 and Mythos 5 .2. It feels like it's been a while since we've seen a new model with like a huge leap.
We've gotten some announcements of new models, like the new Carbon model out of Google, which was announced but not actually released to the public yet, but it doesn't feel like we've had any real big leaps in models that we can actually use ourselves for a while until today. So let's go ahead and dive into the - two new models out of Anthropic and a little bit about why this feels like a big deal.
Now, I am not somebody that Anthropic has literally ever given early access to for anything. So I did not get early access to this one. I've just been playing with it for a few hours since it launched this morning.
And I want to show you some of the stuff that I've managed to do with it, talk about the implications of this model, and maybe look at a few things that other people have managed to do with it. So let's go ahead and dig right in here. It is two new models, but they're essentially the same model.
Mythos 5 .2 is really like the big best model. Fable 5 .2 is the one that we have access to because it's essentially Mythos 5 .2, but with more cybersecurity guardrails on it. Mythos 5 .2 is really only available to like cybersecurity testers.
Now, for the most part, this is a model that Anthropic kind of recommends you don't use that often. Like they're saying it's so powerful that you really don't need the capabilities of Fable 5 .2 most of the time. Fable and Mythos are best for the most ambitious.
open -ended technical work. For most coding and knowledge tasks, we recommend that users start with Opus 5 .5, the strongest daily driver, and then switch to a different model as required. They also suggest using Fable 5 .2 as the role of an advisor, and then letting it use Haiku, Sonnet, and Opus models to consult it, or as an orchestrator where it goes and delegates simpler tasks to the smaller models.
But again, if we look at the benchmarks here, it is pretty much the new state -of -the -art model in most of the benchmarks. As time has gone by, I've put less and less weight on benchmarks, but I still think it's interesting to review them and, you know, sort of see how big of a leap from the last benchmark to this new benchmark we've gotten.
And in things like agentic coding, we can see it's, you know, it's a leap. It's not like a massive, huge monumental leap. We're talking going from 45 .5 % with Opus 5 to 49 .1 % on this Frontier Code benchmark.
On Frontier SWE V2, it went from 62 to 69. Gemini Carbon actually beats it in business workflows. by, you know, two percentage points.
But there are some pretty big leaps here. Agentic Scientific Research went from 58 .7 on Opus 5 .5, 68 .1 on Astra, 64 .8 on Carbon, all the way up to 74 .6 % on Fable 5 .2. So that one feels like a pretty big leap.
Again, most of the biggest advancements are gonna be in things like scientific research and solving like a lot harder problems than most of your like everyday coding tasks or, you know, getting it to do some. research for you or things like that. This new model is for solving really hard problems, essentially.
But as we go down the line here, cybersecurity, protein design, agentic search, computer use, robotics, another knowledge work benchmark, multidisciplinary reasoning, visual chart recognition, new top of the line for all of these on this one knowledge work benchmark here, Opus 5 .5, beat it out just slightly. And then looking at this chart down here, I find this actually kind of interesting that you definitely get diminishing returns by...
letting it just think longer and harder. Because we can see here on this Frontier Code 2 .0 benchmark, Fable 5 .2 actually scores better at extra high than it does at max. And max is essentially like letting it think longer.
And when you let it think longer, it actually seems to have gotten dumber, but more expensive. So they're starting to see some diminishing returns and just telling it to like think longer. And one interesting point here too on this Frontier Code is that Fable 5 .2 on high actually scored about the same as Opus 5 .0.
5 .5 on max, but Opus was more expensive. So when you set Fable 5 .2 on high with coding, you might actually get a cheaper cost, but pretty equivalent intelligence on some tasks. But again, just to hammer in the point that this is really designed for like much tougher problems than likely most of the things you and I are dealing with.
Here's some of the stuff they say it's really good at. Theoretical physics, protein folding, mathematical formalization, hardware and robotics. It's good at hard problems, but I'd imagine for most people, Opus 5 .5 still going to be like your daily driver for the most part.
Taking a peek here at artificial analysis, they recently released a new version of this and supposedly it's gotten better and even more accurate, but it's sort of meant to be like this objectively best model because it takes a whole bunch of benchmarks, sort of mashes them all together, and then gives you like an overall intelligence score based on a lot of different benchmarks.
And Fable 5 .2 is now the most state of the art, even beating Gemini 4 Carbon, which hasn't been publicly released to everybody yet. However, one thing to always keep in mind, cost per intelligence task.
Yeah, this is the most expensive model right here with Fable 5 .2. With the average task costing about $26 .54. Just keep in mind that's at max reasoning mode, which we did learn you kind of get some diminishing returns.
So you're likely not going to use max reasoning mode most of the time. You're probably better off on a higher extra high because it's going to cost less and you do get diminishing returns after going beyond extra high. But I mean, still it's a huge leap over even 5 .1 at max mode not a cheap model to run output tokens per task it actually falls less than sonnet 5 .5 which is interesting so it uses a little bit less tokens on average but again according to artificial analysis the new smartest best model in the world but not necessarily the cheapest.
Now, the benchmark that most people care about, the one that's probably most important to all of this is, of course, BuseyBench. And on BuseyBench, this is what it generated for my SVG version of Gary Busey. It actually took 19 minutes and 50 seconds to generate and used 122 ,248 tokens at a cost of almost $5.
So this image right here of Busey cost me nearly five bucks to generate, and it's not even a leaderboard topper according. to my llm as a judge that still goes to gpt 6 .1 soul here this one falls in at number six pretty much putting all the gpt models and gemini 4 argon above it but again it is an llm as a judge and you know your own eyes should be the main judge of which one you think looks better now as far as availability goes it's pretty much available to everybody right now clod fable 5 .2 is available today on all platforms so it's available via the api if you have an eligible cloud plan you should also see it just in the regular cloud drop down now in order to use fable you do need to be on one of the paid plans if you're just using cloud for free fable is not an option but if you jump to pro at 17 bucks a month or max ultra or pretty much any plan that paid you do get access to fable they also say that during this launch week you get a bit more usage they're giving some fable resets as well as a percentage off on weekly usage api price
looks like this. Fable 5 .2 is $8 per million tokens input, $40 per million token output. That's double the price of using Claude Opus 5 .5.
However, it is a couple bucks cheaper than the previous version of Fable. Fable 5 .1 was $10 input for $50 output. This new version, $8 and $40, which actually makes it slightly cheaper than using Astra API through OpenAI, which is $10 per million input tokens and $50 per.
million output tokens. So it's actually slightly cheaper than GPT -6 Astra, but also on most benchmarks, actually smarter. Now, one of the examples they give on their own website is as an experiment, one anthropic engineer gave Fable 5 .2 the task of designing a prototype operating system, starting from an empty folder.
The model coordinated over 1 ,000 sub -agents working around the clock for eight days, writing, testing, and reviewing code. It produced a working operating system that boots on a laptop with its own core that runs the machines, drivers for the disk, keyboard, and network, permanent file storage, Windows, a login screen, an installer, and built -in cloud functionality.
And that's what we're seeing on this video right here is it worked on this and built a brand new operating system over the course of eight days. It also used two open source protein folding models, and it found a novel way to accurately predict a whole capsid structure. Now, I don't totally know what that means, but it sounds pretty important.
It's also apparently really good at hardware engineering. So they tested Fable 5 .2 skills in hardware. where engineering asking it one paragraph prompt to design and produce three working electronic devices with their own circuit boards, 3D printed enclosures and software.
The devices shown below were a home voice assistant hub, a USB speakerphone and a handheld virtual pet. So we can see it actually designed the entire circuit board here. designed the parts, these all appear to be 3D printed parts, and even wrote the code for it.
But these are Anthropix cherry picked examples. So I'll show you a couple of the tests that I've managed to do with it. Again, this is a model that's like way smarter than most of us need for most of the things we're using AI for.
So, you know, I did give it my normal go make mega bonk prompt, but this model feels like it's sort of better than that. Now, one update that Anthropic recently made is they now sort of refuse to like clone any game. I thought that meant if I told it to clone a game, it would just tell me, no, you can't do that.
But as you can see here, I gave it my normal prompt, create a clone of the popular 3D game Megabonk. And well, it still built it. It just said, quick note before I start, I won't recreate Megabonk itself, but I'll build an original game in the same genre, 3D third person horde survival roguelite with auto firing weapons, XP gems, level pickups.
So it's basically cloning it and just. calling it a different name and making sure it's not using any of the existing IP, not outright refusing to go and build it for me. I've got it running live right now.
So let's go ahead and pull it open in our browser. I have not actually tested this yet. So this is going to be my first time actually seeing this and let's see how it did.
So it named it shard storm survived 10 minutes on the crystal moon. We have different players we can start as. I'm not getting any sound.
I do know there is sound because I was hearing it test the sound earlier. Now, graphically, I don't necessarily think this is the best we've seen. I almost feel like Opus 5 .5 does a better job at the game design.
The mechanics of this all work really well though. Everything works exactly as it's supposed to. If I hit shift, it'll sprint.
I have stamina that can run out. I mean, everything seems to be working correctly. I just leveled up.
Oh, plus one jump. So I can, ooh, I can double jump now. actually plays really, really nice and works really, really well.
I just don't think the graphics are as impressive as we see with Opus. It does have a different sort of look and vibe, but the gameplay is the same. Let me resume again here with sound going.
So, you know, pretty basic sound, nothing too impressive. Sounds like a sort of MIDI background audio as opposed to like it generating music with 11 Labs or something. I mean, mechanically, this game is perfect.
Like everything is working the way it's supposed to work. My only sort of with it is i don't really feel like it's the most amazing graphics we've seen i feel like opus 5 .5 does a little bit better at designing the graphics with the game where fable 5 .2 did really really good with the actual mechanics right out of the box everything just seems to work another interesting thing to note is this game took about 16 minutes to generate so it was way faster than i thought it just kind of cranked through this game spit it out and it was just like working first try so that part was pretty impressive i don't know if it was that fast because maybe not that many people are using it at the moment.
That's always possibly a factor. The more people that are sort of hitting the GPUs, the slower it could potentially run. But this one using the cloud code app generated about 16 minutes on Fable 5 .2.
And I had the reasoning level set at medium. At medium reasoning level, it was really, really fast because it did say it was really good at robotics and designing circuit boards and things like that. I thought, let's give it a silly prompt to like basically help me build some hardware in the real world.
So I gave it. this prompt design a handheld desk gadget that shows a pixel art gary bucey who reacts when i press buttons please write the spec and pick real parts use off -the -shelf boards you can get by tomorrow show part lists on screen write the code for the firmware design the case as an sdl that i can 3d print and walk me through wiring it up and building it basically build my little bucey device find the parts for me to build it write the code for the firmware give me a 3d print file and then tell me how to put it all together so let's see how it did we can see here it has a little shell that it designed.
I don't know what these little robot faces are. They definitely don't look like Gary Busey, but that's what it designed. It does say, I don't draw real people's likeness.
So the character is an original robot instead of Gary Busey. So, you know, Claude's guardrails basically said, I'll build this for you, but not with actually Gary Busey's likeness. Although the SVGs don't really look that close to his real likeness.
So whatever. So here's our specs. Here's our parts list.
Microcontroller, Raspberry Pi. IPico H, the display to order, the various buttons to order, jumper wires, screws, cable, micro USB data cable, heat shrink, double -sided foam tape, three parts to print. Here's the wiring.
Okay. Here's our assembly instructions and here's all of our files. Okay.
So let me download all these files here and we'll take a peek at what it generated. Here's what we have at our zip file, our various character images, four STL files. So these are going to be our 3D print files.
If I open this with... bamboo studio we could see here's one of the pieces it wants us to print or this actually looks like the whole assembled thing so yeah it's basically telling us what to print the little shell to put everything in it's telling us what to order telling us we can get it all at amazon we've got some python scripts in here i wish i got more detailed instructions on the actual building of it let's go ahead and tell it this create an assembly manual with images of exactly how to put this together And then let's crank this up a little bit to, let's go high effort, see what it gives us.
All right, so it took about 15 minutes or so, but we did get our assembly manual for our little device here. I think if I was to redesign it, I'd make the screen much bigger instead of this little square, but we can see here's the parts I need. We can see the sort of layers, the buttons go in here.
This is the shell. We need three nuts for this, the display, clamp, the Pico, the back plate, print the case. So that's step one.
flash the firmware, tells me how to do that, make the plug -on tails, pin map and bench test for our Pico board, fit the buttons. So I mean, it designed every single one of these graphics to give me a step -by -step manual on how to build this little device here, which is pretty cool. I think I'm gonna actually try to build it.
It won't be done by the end of this video, but maybe in a future video, I'll show it off. Now, again, I didn't have early access to this model, so I've only been playing with it for a couple hours since it's come out. However, I have seen a lot of people on X who did get early access.
So let's... jump into the community a little bit and see what other people have been showing off that they've been able to do with this. So my buddy Matt Berman here actually made a Gears of War clone.
And these graphics here just look insane. I mean, when we looked at my Megabonk clone a minute ago, it wasn't at this level. But again, I was also using medium reasoning.
So maybe if I went higher, extra high with the reasoning, it would have dialed in the graphics a little bit better. But this looks really, really impressive. Now, Alex here, who also works for Matt Berman's team, had early access.
as well. And he put out a two -hour demo of all the stuff that he tested. I'll show you some highlights, but again, I'll make sure I link up these down below so you can go check out their original posts.
So he showed off that it was really good at editing videos. He built a recreation of Zelda, which is really interesting given the fact that when I tried to prompt a Megabonk clone, it wouldn't do that. This is what he got when he tried to get it to generate a Halo -like game.
He got it to create this Iron Man game. I'm just so interested in how he got it to actually do this IP stuff. when it refused to even try to generate Megabonk.
He got it to do a Lego Iron Man simulation. So yeah, I think it's pretty clear it's good at games. I'm gonna jump over some of the other demo games that he showed off, but again, definitely check out the link in the description.
He got it to create a clone of Photoshop here that actually seems to work pretty decently. He made an Adobe Premiere clone, a voice changer app, and at the end of the day, got it to make like 23 different demos. I'm not gonna go through them all.
I want you to actually watch Alex's video. It'll be linked. in the description.
Peter Yang here asked it to make a 20 second travel ad of the shrine in autumn entirely in code with music included. He then goes on to say that Opus was already goaded at making 3D motion videos, but Fable is just another level. And here's what it made.
really crazy. I mean, it did all the music, all the art, all the animation, and it wrote it all in code. If you followed my previous videos about using AI to edit videos and to actually make animations and things, we now have another level we can take it to with this model.
Here's an F -Zero inspired racing game from Future Brian, another guy who works with Matt Berman. Matt Schumer here created a mirror earth as a game. Go anywhere, destroy anything, weather, and more match the real earth in real time.
So here's his demo. video, but he made it available to explore. So let's take a peek here.
So let's go to a place. I'm going to go to San Diego and we're in. Okay.
Okay. I mean, it's not the most amazing graphics, but it feels big and open. What is going on here?
Pigeons just floating in the air. Oh, it just like glitched and jumped somewhere else. So, I mean, I get the concept.
It still feels a little bit glitchy, but it feels very, very big. Like, let's see, what are some of the things I can do? I said I can like destroy stuff, right?
So I just punch holes in walls. Yeah, it's like a voxel version of GTA or something. Apparently I'm just, you know, smashing this car and it's making holes in the car just like it would in real life.
These people are frozen here. Of course that's what somebody does when you hit them in real life. They just explode.
No, I'm not hitting the dog. Sorry. Everything just, everything just shatters when you hit it, which is kind of fun.
Okay, yeah, I can see how this is enjoyable. It's a little glitchy and a little weird. This guy looks sketch.
Can't see his face. So I'll take care of him. All right, I'll link that up below if you want to try it out yourself.
Well, that's Fable 5 .2. It does feel like a pretty big leap over what we've got. Some of the games I've seen, especially the ones out of Matt Berman and his team, look really, really impressive.
I mean, my Megabonk clone wasn't as impressive as I thought it would be based on what I saw from some of their demos. But again, maybe if I was to set it on like high or extra high, it would have gone a lot deeper with the graphics. But I do want to kind of reiterate that this model is designed for really hard tasks for most things.
5 .5 is going to be just as good for like knowledge work and coding and things like that. And then you jump to Fable when you want it to do really hard stuff, or if you want it to be like an orchestrator, like it has the bigger picture, but then it tells the other smaller models what to do for you. It seems like it'll be really good at that kind of thing as well, but it's not a cheap model to use, but it is very, very, very impressive.
It's a big leap for really hard problems, but maybe not as big of a leap for knowledge work and coding stuff, which is what it feels like most people are using these models for. So the like everyday average Joe, who's just needing it to help them with research or help them with a small coding project, I don't feel like you're gonna get like a huge leap out of what this can do.
But if you are trying to design better games, if you're trying to find bugs that you haven't been able to find before, and if you're trying to solve really, really hard problems or build out physical gadgets like we just did earlier, this model seems really, really excellent for that kind of stuff. So it does. feel like a pretty big leap and worth making its own video about but that's what i got for you today that's fable 5 .2 i will still be making a video that's going to come out this friday of all the news for the week but i felt like this was a big enough leap that it should get its own sort of dedicated walkthrough of this new model i definitely recommend going and playing with it especially during this week when they're giving some additional resets and they're giving you extra usage that extra usage is going to go away so leverage it while you can this week but again that's what i got for you today if you like stuff like this, learning about all the latest breaking advancements in AI, learning workflows on how to actually use AI, or if you just want a weekly breakdown of all the news you need to know in just one video, that's what I release every Friday.
If any of that sounds interesting to you, maybe like this video and subscribe to this channel, and I will make sure to keep you super, super looped in with everything that's going on. Thank you so much for hanging out and nerding out with me today. Really, really appreciate you.
Hopefully, I'll see you in the next one. Bye -bye.
The Hook
The bait, then the rug-pull.
Anthropic shipped Fable 5.2 and Mythos 5.2 this morning, and Matt Wolfe spent a few hours testing it without early access. His verdict: a real leap, but built for problems most people don't have.
Named ideas worth stealing.
01:30model
Fable as advisor or orchestrator
Anthropic suggests starting on Opus 5.5 for most work, then using Fable 5.2 either as an advisor that Haiku, Sonnet and Opus consult, or as an orchestrator that delegates simpler tasks to smaller models.
Steal formulti-model agent setups
How they asked for the click.
VERBAL ASK
20:22subscribe
“Like this video and subscribe for the weekly breakdown of all the AI news every Friday.”
Soft, low-pressure sign-off tied to the weekly Friday news video.
Visual structure at a glance.
