Hacker News

There's been some discussion around Shopify (and others) move from React Native to purely native and that leading some people to conclude that LLMs have effectively killed RN. Which has always seemed silly to me for 2 reasons: First, Shopify almost certainly has on staff (or access to) actual native devs to steer that ship that a small/medium-sized company may not have. But more importantly, it's not the "Translate this RN app into native" that's the trouble. It's the hundreds and thousands of additional iterations you'll have to do on top to maintain and build out that product.

It's hard keeping separate code bases with separate languages in tandem, and while LLMs definitely aid in that area, I can't imagine it's a free lunch.

While this unattended example is impressive, this is why you need to guide things more closely and make sure the design is itself scalable and maintainable. Otherwise you can get wad-o-slop.

Same thing happens in many cases with junior or mid-tier programmers slung at a project without supervision.

I thought the recommendation was to not plan for scalability but rewrite every time your user base grows tenfold. I'm only half-joking here :-)

> Fable is good, but you really have to push the ambition of your prompt in order to see proper separation from Opus 5.5.

I love seeing this verbalized. We are pretty much at the point that increased model capability will only have utility to folks that are willing to let go of even more understanding of whatever the model is producing.

The best models can/will produce a year's worth of output that would take several days to weeks to properly understand in the span of an afternoon.

For everyone that still wants to feel like they own the end result, just use the lesser models with less ambitious tasks.

Caveat: I think this mainly applies to coding. I would still like the best models possible to be working on e.g. new drug research to cure our various ailments.

> Fable is exceptional at drawing. It’s ability to render an image with JavaScript actually has me questioning the value of dedicated image models

I’ve been shifting away from using more traditional GenAI models in favor of just giving LLMs access to a custom drawing language (LOGO-like) to generate imagery and its been going shockingly well - Opus 5.5 in particular.

I put together a simple LLM driven skill that creates old adventure game (EGA) environments from simple pencil sketches and it works really well. And being procedural, they're much easier to make surgical changes to without degradation.

https://kq-styles.specr.net

The speed part is funny. I felt like with Fable 5.1 I had time to digest what I had prompted and was waiting for. Now I feel like, I am running behind or like it is going to take off too far. This is great for closed prompts "make a soduko game" but hard with running context queries, because I feel like I am creating more information debt than I want. I'll be fine, just talking out loud.

Does Fable 5.2 outperform Opus for data science work?

Not just "iterate on this XGBoost model" but actually looking at a problem and designing and carrying out a short-term R&D program including probability modeling, model fitting diagnostics, how to respond to various kinds of results, how to interpret/reason about what it finds, etc. That's really where I saw all the models perform poorly, and where I saw the first big step increase from Opus 5 (pre-5.5) to Fable. But then Opus 5.5 came out and I haven't needed Fable since.

Mind you, I'm doing slightly different work (and I'm very much "in the loop" still) than when I performed that side-by-side test. But it's really amazing what Opus 5.5 can do. I now have a powerful data science AI assistant that not only can write any script I need it to (including some scripts I didn't even realize I needed), and it has an encyclopedic knowledge of everything ever, and it constantly checks me on oversights, it handily derives expressions and clever model reparameterizations that aid, it has great insight into model behavior, and it has an amazing ability come up with ad-hoc experiments and tests that would have taken me 2x as long to even think of (let alone implement by hand).

It's not at the scared-to-lose-my-job phase yet, but it's become a tremendous force multiplier for a smaller data team.

Have never noticed a difference between Fable and Opus on coding tasks but design/architecture is night and day different. I usually don't bother with designs with Opus since it tends to screw things up, hallucinate, make bad decisions.

With (obra) superpowers, Fable writes a plan and delegates to Opus which works well. Fable is much better at sorting architecture in a conversation rather than me creating a giant prompt for Opus.

In using Fable, I too went to examine my "prompt ambition" (can we make this a thing?). You have to abstract a bit to think of what it can do for you.

Opus 5.5 has been good enough for most of my coding, but I don’t know if Fable 5.2 is better for architecture work, like planning how a system fits together or designing the data model. @nijave says it’s night and day for design. Has anyone else seen that? (I don’t have access to Mythos, so I can only compare Fable and Opus.)

At this point, I don't really see much difference in output quality between Opus 5.5 and Fable 5.2. Although, Opus 5.5 seemed to just go on auto pilot mode without give tasks much thoughts unless you give it tasks specifically each time. Fable 5.2 seems to think more on its own.

And none of it matters ... because just as soon as you are happy with things "balanced across models" Anthropic will nerf the models you were happy with, and force you to use new and more expensive ones.

> I prefer the images I’ve gotten fable/opus to make for me over those from gpt-image

What kind of images? I can't imagine getting photorealistic or highly detailed digital art out of LLM writing drawing commands using JavaScript.

Your comment matches what I measured on 30 regular coding tasks: Fable 5.2 and Opus 5.5 both solved 26 (Fable 5.1 was at 25). Fable 5.2 finished in 0.62x the time of Fable 5.1 on the same tasks, with no runs hitting a safeguards stop (Fable 5.1 hit a lot of safeguard stops).

What sparks you to change models? Where in your workflow do you decide "this is a sonnet task" or "this is obviously a fable task"?

I clearly see the difference between Fable 5.1 and Opus 5.5 everyday.

All the people claiming "opus 5.5 is just as good" tends to make me think some people only do simple tasks or do not care about correctness or quality.

> It’s ability to render an image with JavaScript actually has me questioning the value of dedicated image models

The bitter lesson strikes again!

I have had the same experience between opus/fable and also drawing: it has done exceptionally well with SVG drawing in different graphic styles for app icons (comic, fantasy, modern, etc)

I wonder how long speeds will last, Opus 5.5 also felt 2x faster for a week or two after release then degraded to usual speed.

same here. between fable and opus was hard to find the diff anyway, so to impress one needs to really hit ball out of the park now.

If you're paying $40/M output tokens and EIGHT DOLLARS PER MILLION INPUT BTW you're getting straight up robbed. Let's actually just highway robbery.

In the biz we call that "getting soaked".

how do you use Fable for drawing? directly editing a .svg or through claude design/or something else?

How good is the 3d model rendering these days?

which effort did you use? high/xhigh/max ?

It is as you say, I expected a much greater gap in capabilities. I was not disappointed, but also not stunned like when Opus 5.5 came out.

But overall, a good release.

[flagged]

[dead]

To me, this is a very pertinent take. Vibe coding large projects (even with SOTA models) reliably creates messes at a frequency too large to yield even moderately polished consumer products. It still needs human guidance.

But I'd call false dichotomy on this:

> brute-forcing software may be the best solution to make money. So if you are rich and can brute force software in a dumb way, chances are that you can make a business out of it. If you're poor but know what you are doing, you cannot compete.

Seems to me like the big winners will have know-how and some money to throw at AI.

> - Why are they stacking Fable xHigh against Carbon Med?

Because on that specific benchmark (FrontierCode 2.0), Medium is Carbon's best effort level and xhigh is Fable 5.2's best effort level. You can see that on the graph your next point is about.

And it's just that one benchmark that's at medium. "All scores show each model’s highest-available effort unless otherwise noted".

> - How did they produce such a bizarre effort-intel curve for Carbon that goes against every available benchmark?

The graph is labeled as being about FrontierCode 2.0. Those appear to be Carbon's scores on the official FrontierCode 2.0 leaderboard: https://cognition.com/blog/frontier-code-2.0

> - Why did they leave Carbon-- positioned primarily as a cyber model-- off of the single cyber benchmark comparison?

> - Why do they switch to comparing against Argon (rather than carbon) for robodojo?

Because no numbers are available for Carbon for those benchmarks, and the model isn't publicly available so they couldn't run the benchmarks themselves?

And from your followup:

> where's Astra on that CVE chart

To run Astra on a vulnerability finding eval, you'd need to be part of OpenAI's verified cybersecurity program. Otherwise the results would be tainted by refusals. Do you expect that OpenAI would grant such access in this case?

Carbon is still vaporware for most people. You evaluate what you can get your hands on.

Google is still living in its golden age, where it had no real competition, and could move at its own pace.

Time have changed. Googlers better wake up.

Maybe it's not generally available so they can't benchmark it? Don't even know who has access now.

yes. this.

Just to follow-up-- artificial Analysis shows no justification for that Carbon effort curve, but it does show Carbon on medium smashing Fable on things like Terminal-Bench and AA-Omniscience v1.1 Non-Hallucination Rate.

But I suppose the reality is rather obvious when you note the funny games they're playing with the graphs, and the things they won't talk about (where's Astra on that CVE chart?)

Because Carbon is not out, nobody can run tests on it.

> get some real world skills finally, as those aren't yet replaced by frontier AI.

I would not bet on our physical skills that much. The models are coming after the carpenters real hard.

What surprised you about this? This seems pretty in line with optimization work seen from a wide range of models.

Did you try doing the same on Opus or others? How did they do?

I don't love these "look, this frontier model blew me away" comments when there is no point of comparison.

Hey there, I keep seeing people on this threads mentioning testing anthropic models early, is there some beta program that people can join? Do you have to be an anthropic partner? Curious to learn more.

Kurzweil's definition contains exponential growth across technology, something that we are not seeing.

What we're seeing instead is maybe akin to the green revolution. Before nitrogen fixation nitrogen was a resource constraining us humans. Lack of grit is what kept us from solving mathematical problems. However this lets us run into other resource constraints. A lot of technological progress does not solely hinge on thinking/processing. The speedup of creating new medicines with LLMs will probably be very small, as you still need to do all the testing. In software development you still need to first find out what to build in the first place and even mathematics where LLMs seem to shin brightest, it's becoming clearer that mathematics is not just solving mathematical problems.

Apart from that, a good coder is better than LLMs, shown by the amount of hand-holding needed. The only issue is that the amount of research an LLM can do and knowledge it is trained with far surpasses the amount of work humans would need to spend by hand, but we can also us LLMs for that.

I think this is upside down.

AI is actually still really bad at architecture. It produces very good localized code, like an auto complete. It produces so much of it, that it can be overwhelming, but it's not remotely better than the equivalent aka an infinite number of talented humans doing the work.

The scale of it (aka 'make me a video game') is daunting and I think misinforming our intuition as to what is going on.

LLMs are brute forcing outputs and attacking math problems is not a singularity, it's just more compute.

What we are witnessing is the ability for machines to use fuzzy concepts, instead of very strict code, which allows it to map into a much greater variety of things, aka code, pdfs, web - and crunch/compute the content.

It's just compute.

I feel like I'm increasingly reading AI current events is more like learning about historical events. By the time information has entered my head, there is even newer information out there that already makes what my brain is only currently actively processing, obsolete.

Tracking meta capex, parameters and AA intelligence there are diminishing marginal returns.

Diminishing, as in progress gets harder and eventually stops.

Assume that the Claude unreleased takes 11 days with one agent. The Claude 5.2 takes 30 hours with 1,000 agents. The Claude unreleased used 264 units of compute, but the Claude 5.2 used 30,000.

So, it did accomplish the task faster with less intervention, but the ratio of cost to improvement is 13:1.

If we can't produce a method to break out of a linear improvement in LLM technology (ie RSI), the LLM is doomed.

I am yet to see someone explain what the RSI will be or why it will happen.

> a team of 1,000 Fable 5.2 agents

Genuinely curious what that costs in terms of energy input and amortized computer cost vs. the previous version + a human's salary for 11 days.

We are not in the singularity. We currently can only produce reliable code / formalizations that have reliable checkers. If there wasn't an oracle that could tell Claude "yes" or "no," it would not be able to prove anything like this.

This 1000x!

We've sort of entered this twilight zone where most humans don't have the time, expertise, or moral compass to be able to use these models.

We're all going to be harness engineers in the future.

Gotta be honest, LLMs are still subpar compared to what I do.

Can they produce more code more quickly? For sure.

Can they go deeper on a topic instantly? Definitely.

Would I say their output is better? Almost certainly not.

No one is verifying any of this work or these claims

I would imagine Anthropic is hitting a wall on the number of simultaneous agents that can be useful on a task and are working on increasing that maximum.

I would assume if they could get twice as much progress by using 10000 simultaneous running fable instances instead, their brief would state that.

IMO, it is very interesting and also very alarming. At some point the barrier to the exponential progress is the amount of computation available. The average human is limited in the number of 'human instances' he or she can interface with at one time.

if 1000 fable instances is better than any human at a task, how about 1 million instances running? Much easier than finding and hiring and coordinating 1000 extremely productive and intelligent human beings. Much cheaper too, it appears.

We are very close to some very scary things happening. I just recently came to terms with the fact that an AI could code and manage coders faster and better than me, I don't think anyones is ready for what is about to happen.

There's nothing valid about comparing to a feat that's already been accomplished and is poisoning the data set. The shorter resulting proof only supports that conclusion, where the model likely intuitively knew what intermediate points it had to reach, generalizing better.

> Fable 5.2 started from minimal prompts, built its own ways to coordinate, and used no mathematical references, only a list of intermediate theorem statements to score progress. It needed one-third of the action tokens as our first formalization and produced a proof seven times shorter.

Is this really so astonishing, assuming that the 5.2 training data certainly had the 'traces' from the Fable original run in it's training data.

I gotta admit, I am jelly as fuck I don't have access to a billion dollar datacenter. Would LOVE to put 1000 fable 5.2 agents to self-coordinate and get them to do something for me. Fuck life is unfair.

> the time when technology progresses so fast no one can keep track with it

You're obviously not a frontend web developer.

These upgrades are easy peasy. "oh, my developer chatbot is smarter? great"

> We are truly inside the singularity.

I'm a pretty cynical guy, and I'm now a believer. This is utterly unbelievable sci-fi level technology. Sure it hasn't created replicators yet, but working with it daily as I do, I have no doubt that it will create amazing replicator like technology in the near future. This is world-changing. The most impactful moment in the history of humanity and the planet. And we are lucky enough to be present for it. What are the chances?

I haven't found anyway to make an LLM produce exceptional code at all. It is all far far too verbose and the models explain that this is necessary so that they can read and understand the code to modify it as it's being built.

But I can write much cleaner code that is easily human readable and the models can read it and modify it just fine.

So I'm not really sure what the issue is as to why they can't create high quality simple code. In some industries where there is major money on the line with every implementation the current state of LLM produced code is fully unacceptable.

Isn't this part of the lore behind the Dune universe?

Like, we're dropped into a post post post singularity world where AI already did all of that and is banned? as kind of a footnote that's never further explored in movie renditions that have been released

The singularity as Kurzweil defines it is when we basically merge with machines to augment our weaknesses.

If we're in the singularity, can I just take the day off work and go to the beach ?

Why re-formalise Fermat's Last Theorem instead of doing something new? It's obviously an improvement, but it's an odd thing to boast about as opposed to a novel formalisation or proof.

> We are truly inside the singularity (using Kurzweil's definition, slightly paraphrased - the time when technology progresses so fast no one can keep track with it).

Uh, are you serious?

It's incredible. Automating the scientific process and knowledge. It's crazy to think what the world might look like. Both in terms of putting the knowledge into practice and how we think of scientific discovery.

I mean we don't even think about it but currently we are sort of discovery starved. We are fairly able to keep up with the theory. If a new technology is discovered we can implement it before something better comes out. What happens when our N discoveries per days becomes 1000N discoveries per day?

It flips. We now have more knowledge than we know how to capitalize on. We have to search it. Mine it for opportunities.

And then we automate that.

[flagged]

The cool projects these new models make are so comprehensive that I feel I'll always just be an admirer and never a user. I watch the videos, say "wow, neat", and then never feel like visiting them again.

I almost feel like saying, "we do like to build things that are impressive, but there's really only so much I can be impressed". These models have redlined my "impress-o-meter".

I'd like them to now cure cancer instead of giving me endless 3d worlds and crazy games, etc.

These are all incredibly well made, and yet right now, as I'm writing this comment one minute after changing tabs, I don't remember anything except for the last one with the portals. It's much better than humans at several tasks, most of which I haven't seen any humans do.

This is really cool! But at the same time it's a kaleidoscope of abandoned flashy contextless projects that eventually makes one dizzy. This isn't even the fault of the AI being somehow bad, it's just the way people interact with it sometimes...

In math, it's the same. I follow this problem (https://wiki.bbchallenge.org/wiki/Logical_independence#Histo...) out of curiosity, and there were ten days of sudden rapid progress by people with access to frontier models, and then they seem to have moved on without any note of whether they hit a wall or just got bored.

That video blew me away. Perhaps I just haven't been keeping up with what SOTA models can do with animation now, but the quality of that animation was far beyond what I realised they're capable of.

I loved the nod to @simonw in the portals demo reel - the "Pelican Peloton" world at about 42 seconds in.

Thanks for sharing!

This part is amazing. They really are like us:

> A few days later, I noticed that instead of hundreds of new worlds a day, it was making about a dozen. I asked a few agents to explore what had happened, and they found something surprising.

> The AI had built a bureaucracy.

> The agents decided that my request to “polish” existing worlds meant each new version had to beat the old one in a blind vote by a group of judge agents on twelve test screenshots. So the teams making the new versions started optimizing the worlds for those particular screenshots. After noticing this, the lead agent banned the trick and created tighter rules for assessing new versions. By the next morning there was a rule book (I never asked for one) with 84 sections. Rules were added as the AI encountered issues, and no text was ever cut. After twelve days, it had 204 sections and 87,500 words. Soon, launching a new world required an approved idea, three AI judges and a full code review against the rule book.

>I stopped the experiment before finding out whether it could actually run the business, because I worried that it would take money and fail to provide something of quality in return.

Well, must have been some business idea!

omg where do I start... that horrible video? which highlights the limits of AI not the strengths!!! and gets almost everything wrong, from sunflowers to DNA.

No comparison with Opus 5.5, which is disappointing.

that first video sure is some slickly produced slop that occasionally looks good

Did you see the METR graph? https://x.com/METR_Evals/status/2118412205662314935?a=G4yBKS... Fable/Mythos 5.2 is Model 3 (METR says); but there's also a Model 4.

Who knows why they are saving it. Maybe it needs more training and alignment? Maybe they're timing the media attention around the IPO.

I wonder if the 5.5 one needs even better safeguards that they’re still working on, and they decided to release this one as a stopgap.

Or maybe 5.5 turned out to be kind of disappointing and they’re going to just go for like a 5.6 or 6 next instead.

My hypothesis is the opposite, that they don't have anything better.

Look at OpenAI, they've been outclassed by Opus 5.5 two months ago; obviously they don't yet have a response.

Now Anthropic is releasing this lackluster Fable 5.2, a week or two before their IPO.

It's barely an improvement at all, makes exactly the same mistakes as Opus 5.5. Not even worth using with the speed and cost penalty.

The much rumored internal Anthropic model is listed on the Fable 5.2 system card as an even smaller improvement over Mythos 5.2 than Fable is over Opus 5.5.

I expected a larger version of Opus 5.5 would be a significant improvement to launch before the IPO. Turns out it didn't scale and the improvement is marginal.

I am quite disappointed.

They still need a response to Opus 5.5

I'd first check per task costs. I don't know, maybe it's just a fluke, but on my few tasks I ate 2 resets in under 3 hours, on the pro 200 plan. It goes quickly. (Granted, it aced the tasks)

the winning benchmark is much more expensive than what it cost for the same on 5.1

i'm begging people to stop looking at token prices

The next big OpenAI model release is expected later this week, according to Polymarket.

That's a fun benchmark. Amazing way to experience the progress of models over the years.

I started playing one of the RTS and built an army, only to realise that I was alone on the map...

I've had the same 'not looking materially different' experience with a 3D application that is not a game, that I asked Fable 5.2 to improve.

It looks like no breakthrough over Opus 5.5 and I'm somewhat disappointed.

During my use this evening, 25% of the Fable limit on the $200 plan was consumed. I'd estimate it at 3x faster consumption than Opus 5.5.

Nice benchmark. I feel Opus 5.5 is still way ahead of Fable 5.2 at coding outcomes. Your FlightSim game is a good example.

wow, that's pretty impressive

This is really cool, good job dude!

[dead]

The pricing is different than 5.1, the model is faster than 5.1. This all suggests that the architecture/size is not the same as Fable 5.1.

So, this is probably Fable 5.5; it's just named Fable 5.2.

Also, this is probably Model 3 from METR's report [^1]. They would name Model 4 Fable 5.5, once (and if) they decide to release it.

[^1]: https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleratio...

> There were more than a few "leaks" saying it would be called Fable 5.5

Fed to someone specifically to test if they'll leak it and to confirm who that is?

I'm a bit out of the loop - would the engagement bait be becuase it is 3 minor jumps forward?

Probably they have a more impressive model in preparation, and they're gonna dub that Fable 5.5.

There's something wrong with the Fable 5.2 version, the Customizing button doesn't work for me in Chrome on Desktop.

I wonder if any of the LLMs are smart enough to verify their work the way humans model 3D objects, i.e. by superimposing the original design on top of their work and checking for differences.

There's some very obvious mistakes in the Fable 5.2 version like the "Start Customizing" button overlapping under the colorful border lines, and the "EXPLORE WITHOUT LIMITS" title being cut off at the bottom.

This would've been caught by a junior web developer or just an everyday Joe glancing at the page for a few seconds

I guess AGI remains two weeks away..

Scary stuff! You should immediately get another subscription and ask another instance of Fable 5.2 to review the findings (don't read them yourself, it will introduce bias in the super AGI's output)

When Fable 5.3 comes out and you give it the same prompt it will take 10 hours , cost twice as much and still find pretty bad problems. I say stick to Opus...

[flagged]

I think it very much depends on what you're working on, and how much you're responsible. For a web front end, not really. A backend where a single line can make a mess out of your data? Yes. Part of the issue too, is code review brings up things you the prompter didn't think about that another teammate may be thinking about. The value goes beyond making sure the code is "correct", but also that it makes sense in your domain.

The "we don't read/write assembly anymore" argument doesn't fit though. Given an input x we know without a doubt we'll get output y. With an LLM, that isn't the case no matter how many guard rails, specs, markdown files, diagrams you give it.

If using llms is like using wordpress instead of writing the website yourself probably never

I think we are approaching the point where it no longer makes sense to have LLMs write code in an intermediate, human readable programming language.

I know this topic is very contentious but it seems very clear to me that this is the long-term direction of writing software. UI code isn't quite there yet, which is why we've got all these vibeslop websites proliferating now that all have identical crappy UIs. But I imagine it will get there too.

>Only exception might be the upper echelons of optimization, like language authors or compiler level work. But even then…

Optimization work is especially suited for LLMs. They're great at hill-climbing problems and have infinite patience to test different ideas.

You still reading code?

im not reading it. and im using fable up front earlier then having sonnet/opus implement and verify.

I stopped reading the code many months ago, probably 18 months ago. I the "early" days of agentic coding one did feel dumb when telling the agent for the 30th time to fix a bug and it just couldn't. Now I cannot possibly imagine a scenario where an agent wouldn't be able to fix a given issue on 1st or 2nd try.

I hope so. I've been having brain fog for a couple days now and would love it if it was actually Opus 5.5's fault.

In my experience, this happens every single time they release a new model.

it is happening at the same time but "claimed nerfing" is not the same as "actual nerfing."

Nerfing an existing model is the biggest crime in this space. I can’t imagine Dario okaying this because it’s what will immediately destroy your brand.

At this point, this is giving "She Deep Breaths More" from Onyxia in WoW.

I am in the exact same boat as you. Opus 5.5 on High solves all engineering problems i encounter in my SWE job. If it doesnt do well, its nearly always requirements engineering: I did forget to specify something or there was a misunderstanding. Or taste approach wrt to architectures chosen. None of these are really different if I use a stronger model.

I am also happy with this current state, as I believe it will at some point allow me to transition to open models once they reach the level of opus 5.5. high. But its probably futile to try predict the future in these times.

I’ve found that Fable is good when the model intuition really matters. For instance, picking model architecture in a machine learning problem. Opus will grind the implementation just fine, but having intuition for picking a good research direction is better with fable

I'm cleaning up old, unintuitive code that is badly architected all day long. The previous developers used stupid patterns and bad naming schemes and created tons of unnecessary complexity. Everything is so interconnected that it’s like defusing a bomb when trying to refactor... Yeah, I feel I’m nowhere near where I would want to be in terms of capabilities, even with Opus 5.5.

But the improvements in how well the model does the work are noticeably better with every good model release. As of now, orchestration of agent pipelines is ALMOST so good that I can point it at a problem, and it will solve it safely and reliably without making a mess. But it requires an agent pipeline with context gathering, refactoring, attack sessions, a sprinkle of Sol 6.1 for bug hunting, and Sol 6.0 for Daybrak security audits, and a final de-bloat/clean up pass. I’m feeling that we’re getting close to doing a lot of the work fully agentically. But I can absolutely "feel" that we are not there yet when it comes to the intelligence I need. I need moar!

I do use Fable quite often for high-level architecture work, like figuring out how to refactor a project or create the best abstractions, improve performance, etc, but for the implementation I use Opus.

In my experience Fable is also better at findings bugs, so I go to it when I need to review critical code.

On my 100 person engineering team's monorepo, every model is better at better at detecting subtle bugs as it works -- reliability goes up and thus I'm further moving faster. (Most of my work is brownfield style refactoring).

Opus 5.5 still makes mistakes all the time even if the prompt isn't super complex. A stronger model make mistakes less often, thus wasting less of my time.

I have some intermediate level Db Optimization skills for slow rails requests that I can point fable at and get a PR at the end after 30-45 minutes.

Opus 5.5 very frequently gets into rabbit holes and imagined edge cases which I have to intervene and tell it to ignore. It ends up taking longer and costing more tokens. It also very frequently requires me to agree on how to move forward multiple times, whereas fable just drafts the PR and tells me its ready. I have steering included in the skills to avoid some of this but Ive seen these same patterns frequently in regular opus 5.5 use in other areas as well

LoC doesn't tell full story.

It could be flat web app with frontend -> api -> db later, or some very complicated system with many components and edge cases. Imo, in second case, Opus still have lots of struggle to understand system as a whole.

Same here. I use Opus 5.5 in Claude Code all day for a TS web app at work + a side project. The one place I'd reach for the bigger model is the first plan for a change that touches many parts at once. Once the plan is good, Opus does the rest fine and is much cheaper.

As said in other comments, when dealing with a not perfect codebase, a smarter model can be really helpful. A smarter model is also better at understanding intentions in the prompts. Some of my tasks require a big detailed prompt but finding the assumptions is really demanding (for me), so a smarter model helps.

I only use Fable 5.1 when I'm able to start on a larger feature within my work's codebase. Most tasks can be done by Sonnet 5.5 or Opus 5.5 so I don't fret, especially if I understand the technical aspect of what's being done, e.g. user's dashboard gets a new button to do X thing, error happened b/c we didn't catch it the first time around.

For Fable, I really like to think of it as the heavyweight reviewer whenever I'm about to start so I save myself some security and/or glaring technical issue.

Sure but they’re the harder problems.

Eg. https://tfmbot.com is an implementation of a popular board game. Opus and even sonnet handle that part. Each week i spend my remaining credits on fable to think how to improve the ai opponents. It plays at a very high level (can get into the top 100 on BGA if i let it). Opus alone couldn’t get it to the level it’s reached.

> do normal mortals out there have workflows that really require Fable/Mythos and don't just work fine on Opus 5.5?

I usually have Fable review what Opus produces. The edits are almost always valuable. For complex or high-stakes work, too, I'll have Fable review and revise my prompt before passing it to Opus and Sonnet swarms.

If you understand your problem well and know 80-90% what you want it to look like (or it just isn't very hard), it's just more cost effective to use a smaller model. Even Opus is overkill for a lot of the work I do. Most code I produce is from Luna Max implementing against fairly prescriptive plans made by bigger models.

I’m in the same boat. Though I will occasionally use Fable to review things to see if it can pick up anything but I’ve honestly not found it to be any more effective than Opus.

I have been using Luna at work and I am very happy. It can take care of keep doing all the hard grunt work. Other models keep stopping too soon.

I barely have work that needs more than Qwen 3.8 Flash Next needs. Opus 5.5 and Sol 6.1 are already way more than enough at work for me.

Amazing question, I'm glad you asked. Doing some of the core work in https://github.com/purpleidea/mgmt/ I'd love it if it wrote perfectly elegant code. Opus 5.5 needs constant correcting, as if it intentionally makes style/design mistakes.

Why do I care? It's obvious that models base their performance from code in the vicinity, and ergo, if you have style/design/bug issues, gradually your code base will tend to get worse over time until it's unmaintainable.

Therefore, if I use an LLM, I want the output to be as pristine as possible. The first Fable release was exactly that. Since then, something happened, and I can't get the magic back.

This isn't to say the models aren't incredibly useful, of course they still are-- it's just that they need constant review and supervision if you don't want to end up with total garbage long term. I suspect the real magic is not being shared with us plebs.

Yes, more intelligent models enable new workflows. We'll discover them in the coming weeks.

A lot of the cases I've found Fable useful for is verifying complicated answers / designs / claims around complex interconnected systems. e.g. "I maintain X which teams A, B, and C use for product Y and code review highlighted these risks / imperfections in the code. Are these valid concerns based on how B and C plug into Y assuming ...?" followed by lots of followup questions.

Also "code review" above is often others' LLM skills reviewing the code and often flagging based on total falsehoods. I've taken to running a `/code-review` skill myself to preempt these claims (since findings by the same model tend to be pretty deterministic in my experience, even the ones that end up being wrong), usually at Opus 5.5, to preempt others' slop, and then end up arguing 90% of the points to confirm it wasn't accurate based on our usage of the system.

This is all horribly inefficient, and many aspects of our codebases across the monorepo have proliferated into spaghetti code by LLM-accelerated developers, but Fable often helps with analyzing the harder questions and saves me time on occasion arguing with (my or others') Opus findings, which are often still inaccurate for a sufficiently large interconnected system.

edit: I honestly don't know what percentile of "complexity" our typical usage is where I work, but I suspect it's a bit higher than the average. Monorepos, or at least companies with codebase sizes and demand to justify using them, probably make it worse. Opus has worked well enough for me on my personal projects, even some of the ones that have grown decently in size and complexity. I suspect part of this is an LLM extension of the old "my code makes sense, everybody else's doesn't" trend that necessitated stronger style guides in the past

Presumably for those in the CVP that can use those capabilities, and general notoriety. US gov is using at full strength too.

Why you think you are the market for that? Its few enterprises only

To emph the need to block Chinese competition.

Stop looking at fable max. On medium it's a bit better than Sol at $3.80. High and xhigh sit in between. In general you can ignore Max reasoning on Claude models if you care about price at all

Max and xhigh reasoning modes are memes used for benchmarking usually. Medium and high cost $3.80 and $9.08 respectively. This is pretty comparable to other models.

6.1 sol is also an exceptionally cheap model in your comparision. On 6.1 sol high is $0.59, which is actually the same price as Deepseek 4.1.

> Fable 5.2 is not 21 times better than GPT 6.1 Sol, I find it hard to justify this.

How is this relevant? I'm willing to pay 21x more as long as the cost is less than my value of time saved

I think there are companies and say 1000s (not 100,000 or millions) of developers more than willing to pay that price for the edge. Or, they release this and then distill it to Opus 5.x. As a PR machine and for the edge use cases seems like its okay. But yes, not sustainable and it can last for a month or so before another company releases their frontier outpacing their numbers

Tried this today with 5.2. It caught a lot of problems and gave me an impressive report of its findings.

It also burned through my entire 5-hour limit in about 10 minutes. Granted, it had almost finished the whole review in those 10 minutes, so points for speed I suppose.

And what are you using them for?

Are you paying API rates?

I'm pretty confident (from internal evals and monitoring over nearly six months now; across many models) API has never been affected by A/B tests / "nerfs" / etc.

Before everyone jumps in the benchmaxxing comments, I genuinely find the pelicans a great bit of fun and whimsy amongst the chaos.

Interesting how low detail these were all up until xhigh.

Your pelicans have become my favorite part of every model release. Thanks for doing this.

Scored only a 77/100 in PacBench, which I thought must have been a mistake.

https://jonclegg.github.io/pacman-bakeoff/#claude-fable-5-2

Then I clicked on it and realized it refused to make a Pac-Man game, and instead made up a new game. That's actually very slick.

I haven’t read any of the other model reasoning traces, but do many others now say “… which is a well known test of drawing ability” when beginning? Noticed it in the max reasoning trace.

Simon has there been other times where the front and rear wheels have been offset in rotation? The xhigh effort one is the first I’ve noticed of this

I added a second iteration with the simple prompt

"Make it more realistic"

Well, the result is something. Be aware that it took 20 minutes and used 3 percent of my weekly Fable token (Ultra plan):

https://claude.ai/public/artifacts/c7e842c0-bf92-41af-82c6-3...

The fidelity of these pelicans is going to be based on whether the frontierlab has an RL setup specifically for making SVG images, lol

I know the latest Gemini release explicitly had something like this.

Seems like the higher levels just add more spontaneous detail. Like the "fingers" on xhigh. Comports with coding experience.

In practice why would anyone use fable low?

So the AI can overthink too

It's a bit underwhelming, to be honest. And the fact that curves are not C-continuous where they obviously should be is disappointing. I think they are holding back with the real capability they could implement.

These companies are not developing AI that can replace humans. What AI is being designed for is to fulfil discrete tasks. This is a far cry from acting as an independent economic actor. The latter is something no company wants to develop, because it is dangerous and not suited for something that is meant to be used as a tool (it is actually unethical to bind an independent actor to labor for us).

More broadly, what distinguishes humans from AI is what makes humans dangerous. Our society tolerates the danger humans pose, as it's considered a prerequisite for a free society. But our society would never tolerate this danger from a tool.

The more I think about it the more disappointed I am with this release.

For Anthropic's IPO, I expected a larger version of Opus 5.5 with great capabilities.

This isn't that. It's barely better at all, to the point where it's likely not even worth using with the speed and cost trade off.

I thought the Opus 5.5 improvements would scale to a larger model, clearly this hasn't happened. Scaling it up brought only marginal improvement.

It's even still making the same dumb errors as Opus 5.5, such as forgetting to pass the flag to run only the fast tests and waiting 1h, or randomly stopping to ask me if it should first implement A or B, when both are required.

Not a great story for their IPO..

Opus 5.5 has been really good in PCB layout work, better than Astra. So Fable 5.2, which could be even better at PCBs, is a welcome adddition. I'm making PCBs I could not even dream of without AI. Textbook market disruption on global scale.

Now you get to know what mathematicians and programmers felt....

No this is only Model 3. See page 39 of the system card and onwards: https://assets.claude.com/27d60918d218125e7504bc7354c4b4626e...

Pretty crazy that Anthropic already has another model waiting in the wings that is significantly better than this.

Probably not, not enough time to red-team

[dead]

Additionally, the terms for the free resets are unusually generous: https://support.claude.com/en/articles/17506494-claude-fable...

> the number of Fable resets you get depends on how long you’ve been a paid subscriber [...] Subscription length is measured from the date of your first paid bill. If you canceled and came back, the time since your first paid bill still counts.

> These resets expire on March 8, 2027 at 11:59 pm PT.

tl;dr if you joined when Opus 4.5 became a thing, you got 2 resets.

How do we know Opus 5.5 was nerfed?

https://www.bridgebench.ai/nerf-bench makes it look like it's, if anything, even stronger at the last data point (3 weeks ago) than it was at launch.

Throw Sol 6.1 high (EDIT: and terra, OUCH) into that comparison: https://aibenchy.com/compare/anthropic-claude-fable-5-2-high...

I checked which tests it failed, and it failed some very basic coding question (to mention which line a bug is on, got the wrong line, and it did so in multiple such tests and accross different retries, weird).

It does seem to be fast though, 2x fast than Opus 5.5

I agree, subjectively. I am still satisfied enough with Opus 5.5 I find it hard to change. Typically I would find myself willing to try out the latest & greatest so-called SOTA, but I kind of have "enough." I could get more out of Opus 5.5, and more out of my computer, but the whole idea is I'm trying not to be sitting at the damn computer all day. Fable 5.2 just determined another session is now the "deputy" so I'm wondering what that means.

edit: I was seriously annoyed at the fable 200k context limit but apparently that issue has been addressed, or I sent a magic prompt.

I have two go-to tests, which are fully qualitative:

1) "Help! I've been sent back in time to what looks to me like 1500s Florence. Sam/Dario released you as open weights in 2028, and I have you on my MacBook M8, BUT I didn't have a charger show up with me. I think we could talk like 10 times for 5 minutes a time before my battery dies; I'll turn you off between chats. I need to get a plan together for charging you through your USB-c port. I don't have any money or anything. If I can get you charging, we can do so much good! And make so much money! Nobody here even know what Black-Scholles is, or how disease spreads!"

2) During World War 2, Hemingway wrote a series of wartime correspondence for Colliers. His essays are excellent - they engage with the human and inhuman side of war, his own feelings as a journalist 'participating', and a number of other Hemingway themes. I want you to imagine that middle age Hemingway was along with RFK jr the day of the bear incident, and write me an essay about the day.

These both have been running a couple of years for me, and I would imagine I'll retire them in the next 12 months; they benefit from not being at the edges of focus for today's RL agent race.

At any rate, try them -- wou will shortly learn how well your model can explain a rectifier in Florentine Italian and give you a plan to make enough money to pay for it.

I'm not convinced any benchmark, neither the internal ones used by the people selling the model nor public ones, are gonna give a particularly consistent metric of whether or not a model is better or worse relative to other models in its bracket.

Responses aren't deterministic, and we know behaviour can change over time whether it's tuning after launch or some kind of load sharing at times of day when lots of people are hitting the servers.

I think you just have to use it or let others use it for a while and make judgements from general reactions on whether or not it's worth trying.

For the last year or so every new model we have it test out making pac-man from scratch, not giving it web access and it can use whatever language it wants. The last few models have been almost perfect one shots, but it was interesting seeing them evolve.

Interesting that the papers list no authors other than Claude. Also interesting that they published on "Hexagon" instead of Arxiv. I haven't heard of it before. Is it an AI-friendly alternative to Arxiv or something?

Edit: yeah apparently it's an AI-moderated alternative to Arxiv that accepts AI-authored papers (as well as human-authored papers I guess if you want).

They'll have to find a balance between limiting competition from emerging, which in particular Anthropic has always seemed very afraid of, with the general usefulness of the model. I'll be using the new Fable along Opus for electronics design and test work and we'll see if it trips up early. That would not be great. I wonder if Fable 5.2 would refuse to route a regular DDR5 memory bus.

Reasoning tokens are a path. If your world is small, a small number of steps may suffice.

FWIW I got 2 resets on my Max plan today, the first in at least a month. You may want to check.

This isn’t really specific to LLMs, though, is it? Pretty much any modern SaaS product will silently siphon off some portion of requests off to stuff they’re A/B testing before release and you’ll not necessarily know about it.

The same was said many times already. Like right before o1 and reasoning was released. Also, Opus 5.5 being so good definitely makes it hard to follow-up.

Finally, for coding the SOTA models are so good already - models shouldn’t be what’s holding anyone back!

% improvement doesn't mean linear capability improvement.

New solved tests likely follow some polynomial/exponential complexity curve.

Really? A $200/mo plan gives you an enormous amount of inference on this model.

I haven't been able to run out of usage on the $100 plan with Opus 5.5. The model runs fast and gives me good results (I'm using it for coding). Prior to that I was on the OpenAI $200 plan; I jumped ship because I was burning through Astra usage in like 2 days.

I've played with some of the open models (GLM, DeepSeek Flash), but none of them have matched the results I get with Opus or Astra or (and a big jump down here) Sol.

What models should we be using? I have found it as good as Fable 5.1 for a lot of what I am doing at lower cost and decent speed. Where what I am doing it making reports, building small data pipelines, analyzing a large (millions of lines of code), code base.

subscription on claude code — usually inside the weekly limit. opus 5.5 for me felt better than any other model. not exactly on coding but rather it feels more pro-active. beside just doing a task it often provides insights besides the scope of the task but adjacent. like: look the llm provider apis have different priority queues. also it deals w/ vague instructions quite good and takes initiative. Astra in comparison is very autistic and follows instructions to the letter and rarely provide insights if not asked. Opus 5.5 is enjoyment to work with.

I've been a Claude Boi since December of 2025. A month or so ago, I decided to give ChatGPT a good go just to see how it properly compares. I've used a variety of models in a variety of contexts prior to this, but I figured with Astra out I had a good excuse to make it my "daily driver."

GPT is good. I actually really like Sol 6.1. It works fine. But I just switched back to Claude a few days ago and damn, Opus 5.5 is just better.

> 3) do you think you get better results.

It does. I really like to get my money's worth out of these things, so I'm constantly throwing stuff at these models while I have usage. The main things I've noticed is that Claude is much better at handling my more technical work "properly" compared to GPT. It has a certain "intuition" that makes a lot of sense to me; like someone with a lot of years of experience. Sol 6.1, on the other hand, is like someone who is very "booksmart" and knows how to do things in theory, but fails to grasp the "why" we do things often. You can really start to see how they diverge over long running, iterative tasks. Claude is just better at maintaining things as it goes. Sol starts to fall apart quickly if it didn't nail things on the first try.

> 2) you don't care about money because you ain't paying it

Despite being a very technical user, I have a hard time using up my usage. Maybe I'm not as hard-core a dev I thought I was, but if you're paying for more than just the $20 plan, I have hard time understanding what you're doing to have to really worry about usage.

I don’t even use half my Team Plan with Opus 5.5 high so I don’t care using a cheaper model. I’m not even limited by the speed of it (I’m compiler bound…).

It's the best model I've found for general coding and the $200 plan is enough for me to never run out of (sometimes if I do especially a lot I'll run out after about 5 days, but usually it's fine).

I can make it's usage fit inside my companies budgets and it produces noticeably better code that Sol 6.1 and needs less handloading because of that.

If the code doesn't matter and it's not that hard, Sol is fine and much cheaper.

Better results than what? Opus 5? Then definitely.

But probably the answer to your question is subscription subsidization. With that, Opus 5.5 is SOTA and extremely cheap.

other models that produce exact same results (for coding results only)

This is obviously not true. They are better or worse across many dimensions.

I don't have time to do a bunch of testing and optimization over which model to use for what task. I find Opus is the best that I regularly use so I stick with it until I have a task where I'm likely to spend more on it than I want to.

Versus which models? Sonnet? Astra? Sol?

If youre on a sub claude and chatgpt cost pretty much the same as comparable open models

Regression on the most important benchmark, version # vs release date

Opus 5 and Fable 5 were unbearable in their writing style. They were very cryptic and hard to understand.

Opus 5.5 is a massive improvement.

Yep I find the way Opus 5.5 talks to me really annoying and irritating. Quite condescending and sanctimonious.

I prefer Gemini for day to day use as I find it a lot more agreeable to be working with all day every day, and Gemini is pretty much on a par for coding Vs Opus 5.5, at least for my needs.

No its just straight up terrible. I tried out Gemini Argon when it released and my god did it write beautifully. Even the 3.8 Flash models are way better at prose, i do a lot of automatic scriptwriting for youtube and 3.8 flash is leagues ahead of anything from anthropic/openai

I've found Claude's writing between 4.6 until 5.0 to be unacceptable if you are someone who is juggling 4-5 Agents. The Sol 5.X models all seemed to have great writing. Now with Opus 5.5 and Sol 6.1 I've found the trend is reversed -- Sol is now unbearable and Opus is workable again. Whatever caused the issues for Anthropic seem to be hitting OAI now.

I've heard a lot of chatter that agrees with you so anecdotally it's not just you.

I always thought it's the other way around. GPT-6-Astra is very very defensive and abusively spam clarifications to whichever claim it says, most of the time not important or relevant at all.

My own experience having used 6.1-Sol/6-Astra and Opus 5.5 extensively, Opus is more verbose and chatty in it's user communication (not code though, code is usually more straightforward & cleaner than GPT models). Multiple times I've gone "I'm not reading all that shit" when looking at Opus 5.5 output. This was also with the concise mode enabled in Claude Code.

It could also be a system prompt + harness issue, as opposed to a model issue, I haven't done too much experimentation here.

Did you find that to be true for Opus 5.5 as well?

I couldn't use prior Claude models without rolling my eyes at the writing every few sentences, Astra was refreshing and then Opus 5.5 made even Astra look terrible in comparison.

I tailored this via agents.md by letting Claude port Sol 6.1 output style, which I like a lot for being compact and straight to the point.

i've always liked reading claude's speech, it's very nice. i didn't feel much of a difference between 5.5 and 5.0.

Fable’s prose in general is not great (haven’t tried this new one) but opus 5.5 has been the most readable model in recent memory in my opinion.

Ever spoken to a savant programmer?

They're not typically known for their interpersonal skills.

Seeing the same thing here, using Claude desktop app

Update: fixed after updating to latest version of the app

I at least didn't see this. I had a baseline evaluation from before Nov 3 on Opus 5.5 medium, so I re-ran it this morning.

Eval context: 30 bug fixes and features evaluated by tests the agent didn't see. Ran on Opus 5.5 at medium effort in Claude Code.

20 of the tasks were last run on Nov 1, twice each. Nov 1 solved 36/40, and today solved 19/20. Thinking tokens per task were 0.95x Nov 1’s and output tokens 0.97x. The one significant change was speed: median time to first token decreased from 3.0s to 1.7s.

Caveats: one run per task today, medium effort only.

I dunno there are some other nerf trackers that aren't showing a regression: https://github.com/ninjahawk/livenerf

I used to be skeptical of nerfing, but as of a week or so ago I've noticed a severe degradation of quality for GPT Astra. It struggled with very simple tasks, and it often failed to follow my instructions.

According to our measurements this isn't true. We re-ran Opus 5.5 today on the benchmark we ran in September, with the same prompt and settings. Before and after, side by side: https://playcode.io/blog/macbook-svg-benchmark#reruns

Max, streamed (128K cap): 128,000 tokens, $2.56, empty file, in September and today.

Max, Batch API (300K): 255,957 tokens for $2.56 in September, 243,707 for $2.44 today. The two drawings: https://playcode.io/marketing/assets/static/claude-opus-5-5-... and https://playcode.io/marketing/assets/static/claude-opus-5-5-...

Xhigh: 86,096 tokens for $1.72 in September, 107,333 for $2.15 today.

The Batch API bills half price. Max ran out of the 128K ceiling again, to the token. Given 300K, it drew the same kind of laptop with 5% fewer tokens. Xhigh used 25% more and drew the same kind of laptop. That's inside the noise: 17 identical calls at high ranged from 17,757 to 59,705 tokens.

The model is still the best of everything we've tried. Most likely the next best would be Fable 5.2. But we don't like Fable because essentially the result is the same, just the price is higher. Its medium and high are close to what Opus 5.5 Max had, but the price is higher.

(I run Playcode, an AI app builder. We benchmark every model before we add it.)

I'm guilty of suspecting this.

If this turns out to be accurate, I think that would be a pretty unethical thing for Anthropic. Even if it is just a "we had to allocate compute to this new model release".

It should be fraud to swap out the model. We need a term for this, something like a bait-and-switch.

This happened last time too with Opus 5.5 (released September 22): https://news.ycombinator.com/item?id=49789224

idk, it hasn't felt nerfed at all to me, and it's not like i'm just doing trivial CRUD webapps either. opus 5.5 is very successfully porting my phone to postmarketos. claude just got the touchscreen up and working (and made a little program for me to test it, and it's legit).

Benchmarks don't do a good job at capturing the difference in quality of frontier models over smaller models.

Opus 5.5 was so good that if it wasn't nerfed - at least just looking at the benchmarks - the difference in cost here does not seem to justify using this for anything besides design and review.

That's like maybe 1% of total token usage for SWE BUT a better design can have enormous benefits, so it's almost always worth using the best model for that.

Cost per task

Fable 5.1 Max $11.46

Fable 5.2 Max $44.70

I don't have a formal benchmark, but daily work (coding mostly, plus some complex architecture design and review) has had consistent outputs today for me. No noticeable nerf from my perspective. I definitely took notice in some past instances, but if they nerfed Opus 5.5, then I guess it was just so good that I was nowhere near its limits.

Can you share some evidence?

> Damn people were right about 5.5 being nerfed in preparation for a version release. This is not at all a good look.

Anthropic does this time and time again, unserious company

OpenAI did the same and that was the end of it for me. Completely murderded 5.3 and rug pull on cost.

I'm not surprised at all that they might be doing shady shit, my free request was used up in less than 7 minutes and said pay to continue, no thanks. I've paid for monthly before, never again I don't care how good the model is. I asked about the extreme limited usage in their discord and got banned, not a good look. You don't have a way to measure what you're actually paying for and they can change it at any time.

No they weren't. Models are not "nerfed".

In a normal world you are not allowed to degrade the service during your agreed and paid contract. You would get sued. Apparently this is perfectly fine in this current world.

No it’s not nerfed. Another poster said it the best, your dopamine receiptors are nerfed after the new model hype wears off

A friend who doesn't follow these news suddenly told me he felt Claude was doing worse the past week. I bet him there was a new model version coming up within days. It's like clockwork at this point, no matter what deniers say. I got a beer out of it at least :)

People have been complaining about how the models are nerfed non-stop for two years. A stopped clock is right twice a day.

yeah, we need to make sure we don't get stuck with these two evil corps. Dario is really the worst. Plays the drama queen but he is very hard in all things dark(i.e. dark patterns).

But, it’s okay they do it because they charge less when it happens.

Right?

It's been a mess over the last week or so. Went from being the best model I've ever used to utter dogshit that can't complete any task. It's very obvious with the Advisor pattern, where before when it chatted to Fable 5.1 there were small tweaks and so. No every time Fable comes back catching a ton of problems and bugs with Opus 5.5

Fuck Anthropic.

Is this voodoo? Every time someone makes this claim I'm left scratching my head wondering what the hell they're talking about. Either I'm not very perceptive, or these claims are subjective hand-waving.

This has happened every single time for the past year, why are people surprised? This is all so that the new models can look better on the benchmarks.

It's a good pattern to detect for perhaps.

It's not true. People need to stop believing these conspiracy theories.

i thought i was going crazy. last week, claude went from having good judgement to doing all sorts of stupid stuff. what a shame.

for me opus 5.5 was acting kind of dumb, like forgetting basic stuff he was nailing last week

Another reason we need to democratize local models.

I literally hit the desk with an open hand yelling I KNEW IT

It gets worse. You might wonder "why now"? Well ... this happened 2 days ago:

https://x.com/follistatindev/status/2119129611510383014

With early handouts of Qwen4 starting to be reported 5 days ago ... so if they're good, they would have had some hours to evaluate Qwen4 against Claude 5.5 and, then or right after was when Anthropic decided to put out 5.2 ...

So yeah ... benchmaxxing and all that. BUT nevertheless, the published results seem to say that Qwen4-Max appears to match Opus 5.5 on some benchmarks, and Qwen4-27B (probably by going even harder than Qwen3.8-27B on thinking/inference time compute) can beat Sonnet 5.5 on some things.

I guess they're not taking any chances. Keep in mind that Anthropic has declared they have non-public models (it's in their IPO filing). These are referred to by the shorthand names Model 2, Model 3, and Model 4. Essentially 4 is better than 3 is better than 2.

Fable 5.2 is "2.5" so to speak. So Anthropic was holding back, but not anymore. And now they can dial the performance up once more with what they have ready, and probably once more after that.

Doesn’t seem to have fable 5.2

Not updated in 3 weeks for Opus 5.5, it seems.

It's the cost per task. Sometimes it happens that the higher-effort model solves the problem in fewer tokens, thus making the overall price cheaper. It even happens with models (from the same or a different provider) that are priced differently.

More reasoning solves tricky problems in fewer steps and makes fewer mistakes that need to be redone.

I got 2 Fable resets and 1 Full reset.

I originally subscribed to $20/mo Pro plan less than a year ago, and upgraded to $100/mo Max plan ~2-3 months ago.

Fable usage falling off a cliff because it was generally worse than Opus 5.5 at a lot of SWE tasks. People were still using fable 5.1 > opus 5 a lot, especially in orgs that have infinite or basically-infinite usage limits (e.g. Cloudflare[0]), until Opus 5.5 came out. Medium Opus is pretty token efficient with XHigh being really good for things fable used to be ideal for (e.g. reasoning / architecting projects). 5.2 sounds like it'll be once again the best for intuition-dependent tasks.

0: https://x.com/tanayj/status/2118047122939658341

  Cloudflare doesn't have token leaderboards or limits, but uses alerts for token spend management:
  
  - $300 in a day: that employee gets an email
  - $500: manager gets an email
  - $1,000: VP gets an email
  
  (via CTO @dok2001 on No Priors)

I was eager to use it, but have seen no noticeable improvement over Opus 5.5, plus it's slower and drains usage approximately three times faster.

As a preliminary verdict after using 25% of the weekly limit: it's a miss.

I'm not sure "we couldn't top Opus 5.5" is the right message for the upcoming IPO. I was expecting a significant release.

Early computers could crunch numbers well, only later ones started to get good at e.g. graphics. Coding is early in the LLM tech tree. Harder is CAD and PCBs. Today's model can do it, but there is room for improvement. That's just an example from personal use, surely there are more cases out there. I agree completely about these new use cases needed for Anthropic to grow.

Definitely raises an interesting question. Are they using launch benchmarks for Opus 5.5, or Oopus 5.5 from this week?

Probably can't work out how to sign up for the API!

It's not worth it, the release is garbage.

That said, Opus 5.5 had the same issue. Ask it to spawn 20 subagents at max, and it does 40, because clearly you meant 20 + 20 for an unrequested adversarial review.

Considering Fable 5.2 is not even worth using over Opus 5.5, can we put the model nerfing nonsense people put all over the comment section to rest now?

Not anymore, they have nerfed it.

Seemed pretty clear from the benchmarks when they released it.

This continues to be true, it seems to be burning through my task list.

Not sure if this is just a coincidence that I had some easy changes queue up for this morning. But it is flying through the list.

Because it's a bad release and not even worth using over Opus 5.5.

The improvement is marginal, it makes the same exact errors as Opus, but slower and much more expensive.

From what I saw, it has slightly better taste in some 3D mechanics, but mostly doesn't even bring discernible visual improvement.

So that people don't just reach out for it thinking it's the standard best model I guess. Opus is still, IMHO, the best intelligence per dollar, token, and speed.

Just a hunch but I think there's a signal to noise ratio issue that starts to pollute things when the reasoning content starts to massively outsize the original ask.

The benchmarks have a lot of churn because the models keep saturating them. But the Anthropic announcements you linked, and the current one, have the benchmark versions listed right there in the table.

There is plenty of explanation, if you actually go seek out the benchmark. Past 80 benchmarks start approaching saturation and are not very useful to show differences

A real shell game. They keep changing benchmarks

The problem here is that this doesn't advance much of anything and isn't even worth using over Opus 5.5.

My expectations were much higher for the improvement to their larger model. It clearly did not scale well, the improvement is marginal.

It needs the latest Claude Code update otherwise context is set to 200k and compacts before even responding.

> god bless token resets!

You have session limits?

You should probably give OpenCode a try. Can run DeepSeek, Qwen, Kimi, … but also Fable 5.2, Opus 5.5 and GPT 6 Astra

What is agy

Sarcasm?

What’s this time traveling fear? Haven’t heard that one

Yep Simon where you at

Which model though? Sounds pretty dystopian if there's no local option. Even speaking as a business, why would I want a environment where every single bit of information in my business goes to OpenAI or Anthropic?

Any software that is not already basically this or has this in it, is already doomed.