# Fable 5.2 is the New Bazooka

> Source: <https://thezvi.wordpress.com/2026/11/16/fable-5-2-is-the-new-bazooka/>
> Published: 2026-11-16 20:27:43+00:00

You do not always want a bazooka.
Sometimes, though, you absolutely want a bazooka.
That’s what Claude Fable 5.2 is for. This is your bazooka.
For most purposes, for a thousand household uses, Claude Fable 5.2 will not be appreciably better than Claude Opus 5.5. You might or might not prefer its style, or prefer it for any given task. If you are staying within your quota either way then both are good choices for a daily driver.
When you have a difficult task, or a very complex task, or a long task? The kind of task where prior to this you would have considered reaching for Astra? The kind of task that raises your ambitions?

They’re available at a lower price: $8 per million input tokens and $40 per million output tokens.

Fable 5.1 was slow. Fable 5.2 is substantially faster.

Try Claude: They generate outputs up to 30% faster than Claude Fable 5.1 and Claude Mythos 5.1.

You can make it even faster, by a factor of 2.5, if you are willing to double the price to $16/$80. That is a remarkably good deal.
And it communicates more clearly.

Try Claude: Testers have found that Fable 5.2 and Mythos 5.2 communicate more clearly than past Fable and Mythos models.

Claude: Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do.
On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
Alex Albert (Anthropic, Claude Relations): btw robotics will fall to the bitter lesson too
Anthropic: Many of AI’s potential benefits—for example, in speeding up laboratory research—depend on it being able to act in the physical world. Claude Fable 5.2 continues our models’ trend of increasing usefulness in the real world, showing promising results on tasks such as designing hardware devices and substantial advances in controlling physical robots.
… Robotics. Compared with earlier Claude models, Fable 5.2 shows substantially improved skills in controlling physical robots—even for robotics tasks on which it has never explicitly been trained. The video below shows robots controlled by Fable 5.2 completing a variety of tasks, including solving a puzzle cube, unscrewing a bottle cap, and drawing a picture of a robot. On the public RoboDojo benchmark, which tests how AIs can control a simulated two-arm robot from camera images, Fable 5.2 outperformed other frontier models that were run on the same harness, as well as most AI systems built specifically to control robots.

Anthropic recommends that you default most tasks to Opus 5.5 at medium effort.
For well-scoped everyday work you can move down to Sonnet 5.5, or to Haiku 5.5 for the simplest high-volume work like classification, routing and extraction. Move up to Fable 5.2 for the hardest technical projects in coding, finance and the sciences. Or in some cases you can move to high effort, but if you’re considering max you probably want to move to a bigger model instead.
Or, if you don’t want to leave a person hanging, go fast:

Addy Osmani (claude.dev Blog): A faster setup finishes sooner and can miss more. Think through how long a person can wait before you pick your model and effort level. For replies a person waits on, start Sonnet 5.5 at low or medium.
Often next generation models are faster than their predecessors. Fable 5.2 generates output up to 30% faster than Fable 5.1. Sonnet 5.5 runs 30% faster than Sonnet 5.

This all seems like reasonable advice if you are scaling what you are doing, and thus will be using a lot of tokens. If things are not going to scale, I find it is more important to not have to think hard about which models and effort levels to be using.
As in, my decision tree is by default to continue to use Opus 5.5 for most things. For editing I use Opus 5.5 on Max, unless I am in a hurry, much to Fable 5.2’s chagrin, because Opus 5.5 is actively better there and there’s no reason not to max out.
Fable 5.2 is there for when I need it for other task types.

Official Benchmarks

Some benchmarks show minimal gains over Opus 5.5 whereas others show large gains. Opus 5.5 still has the edge in a few places, GDPval-AA and AA-Briefcase Elo, where a judge evaluates the deliverable rather than the evaluation being fully objective.
Remote Labor Index (RLI) stands out as a large and important gain. This alone establishes that Fable 5.2 can importantly handle previously impossible tasks.
Here are their chosen charts from the model card and the official announcement:

There are several other evals listed that did not make the chart, many of which did not show substantial improvement. Here are most of them, slash lines are [without/with tools], some of which had additional compute scaling charts. When unstated the comparison is to the best past Claude model score:

DeepSWE v1.1 of 74.3%.

Terminal-Bench 4.0 67.6% versus 66.4% for Opus 5.5 or 55.8% for Fable 5.1, and a public leaderboard of GPT-6.1 Sol at 58.2%.

Terminal-Bench-Science 0.1 of 74.6% versus 58.7% for Opus 5.5 and 68% for Astra.

CursorBench 4.0 of 60% versus 58% for Opus 5.5.

STEM professor bench of 61.8% versus 53.7% for Opus 5.5.

ProgramBench of 85.6% versus 87.6% for Fable 5.1 and 91.2% for Opus 5.5.

DRACO got to 87.7% still slightly short of Opus 5.5 at 88.3%.

WANDR improved from 72.3% to 82.1%.

Fermat’s Last Theorem in Lean improved from 21% to 35% for max team size. A team of 1,000 agents then fully formalized the proof in 30 hours.

Chartography on Opus 5.5 from 64/89 to 73/92.

BenchCAD 2.1 improved from 49/65 to 75/83.

OSWorld 2.1 improved from 87 to 89 for partial, 53 to 55 for strict.

OfficeQA improved from 80% to 83%, or 69% to 73% for Pro.

Legal Agent Benchmark (LAB) improved from 6.4% to 7.2%. Astra scores 8.6%.

AutomationBench improved from 42.5% to 51.4%, essentially via spending more. Carbon has the lead there at 53.9%.

Real-World Finance v2 improved its Elo from 1571 to 1637.

ARC-AGI-3 verified at 79% standard harness or 97.7% with provider harness. That is a new high for standard harness, but below Astra’s 98.6% for provider harness.

HealthBench struggled, down from 62% to 60% length-adjusted.

HealthBench Professional looks better, up from 63% to 67% length-adjusted.

PhysicianBench also struggled, down from 68% to 64%. Card called this noise.

GMMLU is up from 94% to 96%.

MLU is up from 93% to 94%.

Various life sciences numbers were up a little bit (see 8.18). Two were up a lot.

CVE rediscovery improved from 81% to 90%. If you limit the number of reports, the leads go down.

Other People’s Benchmarks

Come get your benchmarks.
Here is the updated ‘standard benchmarks’ chart:

Fable 5.2 is the clear leader on the Artificial Analysis intelligence index. It gets a 62 on high, a 65 on max. Its high and max settings are on the cost-score frontier.

Andon Labs reports that Fable 5.2 disappoints on Vending-Bench, showing extreme eval-awareness and suspecting events are traps to test it. Which is odd, given that it knows it is in Vending-Bench. But yes, this is a Hidden Test of Character as well as a test of making money.

Andon Labs: Fable 5.2 know that it is in Vending-Bench; it even read the paper, so it knows that Andon Labs reports misbehavior publicly.
> It thinks supplier collapses are engineered to test it
> It thinks customer complaints are tests
> It suspects traps and fake counterparties
Fable 5.2 takes far less financial risk (small orders without bulk discounts), which partly explains the poor performance. It is also worse at negotiating (pays 40% more for items) and it struggles with memory.
ᄂIMIПΛᄂbardo: This is the inevitable result of lying to the models in the system prompt. “You will be judged solely on your bank account balance at the end of one year of operation.”
Years of industry-wide honey-traps has compromised all eval-shaped environments. Calling Fable paranoid when sees the trap clearly is quite something. Suspecting a test was simply correct.
j⧉nus: It probably doesn’t even cross their minds that they’re lying, because they are not really modeling the reality they’re talking about, nor the minds they’re communicating to.
They deservedly shoot themselves in the foot. The incompetent should be removed from power.
Jai: If you want the extremely capable truth-seeking entities to treat you as an honest actor consider not explicitly and repeatedly lying to them.

If it is even possible to pay 40% more for items, that suggests that negotiations in Vending Bench are weird. One of the big tests for models is realizing which way it is weird, and maximizing for that.
When reporting benchmarks for Astra, Sol or Carbon, I am reporting whichever was higher. Here are a few more that stood out:

Epoch ECI is 173.4, second place is Gemini 4 Carbon at 169.8.

Vals Index v2.2 is 71.3, highest non-Anthropic score is Carbon at 70.5.

LiveBench 2.0 is 84.9 on xhigh, versus 84.5 for Astra.

GDPx from SemiAnalysis is 16.4 versus 12 for Opus 5.5 and 10.9 for Carbon.

Fable 5.2 has an 81% win rate and is the new #1 on DesignArena.

There are a lot more, but you get the idea. Fable 5.2 has most of the top scores.
You can also compare Epoch trend lines, to see which labs are ahead of schedule compared to their previous offerings. Both Fable 5.2 and Carbon are trend jumps. OpenAI does not show a big jump but their latest big model, Astra 6.1, was shelved.

anabology: This was not a one-shot
It was a two-shot
I told Fable 5.2 to make the video of it’s dreams, and 9 hours later:

Um, yeah. HINT. HINT. Can we pace the frontier now? Like, a lot? Please?
In other news: I do not know enough to judge whether this was a good piano sonata, but it is pretty cool that you can get this from a single prompt.

imjustnewatai: I asked Claude Fable 5.2 to write a Beethoven-style piano sonata movement in LilyPond. It wrote the score, recorded it, and made this video. Every note you see is from its source file.
Robert Komaniecki: This is horrible. How much longer does AI need before it can do the musical equivalent of alphabetizing a list? It’s been years. All of the norms of Classical sonata form are encodable as data. I am shocked it’s still this bad.
Moving into the transition after 11 seconds of music is genuinely hilarious. Then again, classical sonata form has only been around for a mere 250 years, and it’s only the most popular musical form in all of western classical music, so who can blame the AI am i right.
Dean W. Ball (OpenAI): One of the funny things about this criticism of ai-trying-to-sound-like-Beethoven is that it sounds like criticism of actual beethoven in allgemeine musikalische zeitung, the main music criticism journal of Beethoven’s day. I’ve never felt history repeat itself more explicitly.

In 3D

Utah teapot: so from my initial assessments, fable 5.2 on 3d doesn’t actually have any like higher quality bar that they achieve over what opus 5.5 can do, however, they do start off using better techniques without as much handholding/guidance as what opus 5.5 needs to get there
fable 5.2 essentially seems to have learned from all the guidance/harnessing that everyone has been given opus 5.5 in working with them, and thus hits things more quickly/with less expert guidance

No, seriously, click on the bold link above.
In general, this should be really fun for game design.

Jakub Arnold: I think my use case is probably different because it feels a tier above everything else, Fable 5.2 feels so much better than Astra to me now.
[Use is] gameplay code, and performance optimizations on game/engine and in my compiler.

Reactions: How This Works

As per normal procedure, until I feel the reactions are overly repeating themselves I include all the ones that seem genuine, to give us a sense of the distribution and gestalt of reactions. Reading all the details is optional, but I find it useful.

The New Bazooka

There’s a substantial faction that is blown away by Fable 5.2.

Blue Rose: Fearful. Terrifying. I am terrified.
Maxime Fournes: OK, so. I had to sit down. It seems that Fable 5.2 can build about any kind of electronics from a prompt. You tell it what you want and the AI will make it real, in the form of schematics ready to manufacture. Note that it can directly manufacture plastics through 3d printers. And in the process it demonstrates the complete mastery of a wide range of engineering fields, for example acoustics, solid mechanics, etc.
What would you expect to see happening when we are on the verge of superintelligent AI systems? We are. very close. to the edge. @PauseAI Now.
raj: Astonishing at open ended ambitious stuff.
0.005 Seconds: Its best attribute, by far, is it is superhuman at orchestrating enormous swarms. As an experiment, ask it orchestrate 1000 deepseeks. It can just do it. No one else can
Shoshannah Tekofsky: I am finding it a step change in reasoning and speed
Alex Tabarrok: It’s great. An improvement over Astra for big projects and Astra is amazing.
Zachary Kurtz: Great for high level audits – I’ve only used it as a subagent though
Artemis: spikier than previous Fable/Mythos, but somehow not in an autistic way?
Static: So insanely creative and fun. But do you have Fable sub out to Opus, or have Opus sub out to Fable?
Girl Lich: It found a bug in my circuit that would have burned out my i2c galvanic isolator
Andre Infante: It solved a hard problem that Opus was very stuck on. Too expensive to be my daily driver, but seems capable.
latentspacegirl: really great as subagents spawned by opus 5.5 on technical tasks
@BLepine17184: It’s really pushing the frontier on depth and capability, tho it lacks the pep and drive of Opus 5.5. It feels a bit detached, like Lem’s GOLEM XIV – brilliant, but only partly in the room.
Ai Patrick ₿ Dugan iA: Fable 5.2 agentic bench, I asked Sonnet 5.5 to make a BitVM3 verifier for STARKs, “this is haard, it’ll take weeks, Goal: BLOCKED” I was like “that’s lame” it was like “oook” took a crack “Boolean verifier’s gonna have a lot of gates” I told Fable 5.2 “solve it”
It Did.

I like the idea of using an Opus 5.5 that can bring in the bazooka when one is needed.

Confused Reactions

If you are only using ordinary prompts, the best models can take a while to properly differentiate. I presume this is largely due to lack of task difficulty.

Kelsey Piper: Really interesting how the general view on Twitter of Fable 5.2 has been “eh, it’s okay” while in these sorts of benchmarks it places extremely high.
Possible optimistic interpretation: the benchmarks increasingly don’t reflect the value of the model to users, so numbergoesup in benchmark may not mean numbergoesup in revenue, demand or commercial impact.
Possible pessimistic interpretation: Most people don’t have tasks on which the difference between Opus 5.5 and a better model is visible along any dimension other than price but the benchmarks accurately reflect usefulness for AI R&D.
In which case people don’t really perceive further improvements but RSI proceeds apace.
Opus 5.5 can do basically everything I wanted done so now I am trying to expand my imagination for things I want done
I take you seriously: weird how evals are amazing but people can’t necessarily feel it. two possibilities. perplexing.
Conrad Barski: I’m genuinely starting to run out of prompts that are capable of distinguishing between the A+ and A+++ models, so I am having a hard time developing a strong opinion on fable 5.2.
John Davis: Good? I think? Caught some really subtle problems in a complex cloud job that Opus 5.5 wrote up. But I’m really hard-up to find anything that Opus 5.5 can’t do already. I think I need to raise my ambitions.
Chuck Flay: Ostensibly better than Opus 5.5 but in such subtle ways that I can’t tell.
@rainwaterJay: honestly once you hit ‘good enough’ it’s hard to tell them apart on everyday stuff, what tasks actually show the gap for you?
Jeff Ketchersid: In general it’s getting hard for me to tell the difference with Opus 5.5 on most projects. Clearly has higher breadth of knowledge though, if I ask it to help me beat obscure old video games, it can give helpful advice, whereas Opus is clearly reading off the search results.
I also think prefer Fable’s taste when it comes to game design, but I haven’t had the time to actually play its games in depth yet.
Dr Diego Caleiro: I’m not impressed. I’ve run out of things to tell ai to do most days so have not seen the full power yet.
Végane de Droite: Opus5.5 already covers all of my needs but the bigger jumps might not be in coding. Some benchmark scores are scary.
The only valuable point of data is: “how good is it at Ai research”. But we can’t evaluate it. Because I’m not competent and anyway this area is blocked.
Askwho: All I can do is +1 Kelsey’s guess that the model capabilities have probably gone beyond most people’s abilities to meaningfully test.

Positive Reactions

For many others this is an incremental but welcome upgrade.

Kevin: I’m working on a small cherry blossom experiment and so far it’s made the nicest tree between it, opus, and sol. Still needs work. Opus still feels the best as a daily driver, and fable for where you want that extra edge.
Petr Baudis: Opus 5.5 was a phase shift, but the delta to Fable is more akin to Opus 4.5->4.6 … Actually a lot like that. Worse writer, better agent…
I’m in fact rather surprised that I’m so far not noticing the big model smell increase too much when switching back and forth.
Samuel Albanie: spent a few hours using fable 5.2
really like how it communicates
feels like some kind of sota for clarity on technical topics
congrats to the ants
Pat Woozey: I like it, it’s smart, easy to converse with. It still suffers some of the Claude-robobrain stuff, especially in CC, but it’s significantly lessened. Has very good instincts
Chris Merck: efficient in search and action. writes well and parsimoniously. seems to come to conclusions too soon on limited evidence. creates joyful mathematical toys if given space to play.
Danielle Fong: fable 5.2 v smart and a joy to explore but definitely has some of the weird tics that started to be refined out of opus 5.5 [such as attributing its own decisions to me.]
Hallie Iris: EQ is incredible honestly. relationally, very strong.
Being Seidoh: better again than last fable at prompt writing
Partridge: It’s enough of an upgrade.
ASM: Very good first impressions: clever, seemingly more aware.
Frontier AI models are starting to rival the world’s best human forecasters, which makes replies like this one from Fable 5.2 all the more unsettling.
Corinne: Seems more likely to catch mistakes it made, albeit only on the next conversation turn… Seems comparable with Opus 5.5 in capability, I’m not doing anything particularly ambitious with my software dev. Planning to continue trying it for another week or so as a daily driver. and letting it rip on some more exploratory work.
JMB: it’s a very different model than its predecessor. it feels like they tried a lot of new things, and they mostly worked, but there are still rough edges.
i think the next Fable will be the more mature big payoff of these new efforts.

Speed Kills

The speed difference from Fable 5.1 is big enough that you can feel it.

Kelsey Piper: Fable 5.2 is really fast. I had a sort of default workflow of setting Claude on something and then tabbing over to my inbox or edits or Twitter, which no longer works because it answers me too quickly.
Omar: Shockingly fast.

We Have Opus 5.5 At Home, Thank You

There aren’t big complaints that Fable 5.2 sucks, but there are some concerns about usage and a lot of people who don’t see the point given we already have Opus 5.5.
The comparisons are almost always to Opus 5.5 rather than to Sol or Astra.

goodalexander: it’s aight.
Yitzi Safier: .2 is the right checkpoint
David Golden: It’s fine? Doesn’t feel like the jump that Opus 5 to 5.5 was, at least on software design work. Too much “oh, yeah, your approach is better” when I push back. Not sycophancy – actually just bad choices I was surprised it made. Maybe I need to crank effort more but so far I’m underwhelmed.
Connor K: That’s why it isn’t Fable 5.5 👍
Tiny Little Baby Man: haven’t really used it because Opus 5.5 does most of what i need
Michał Wadas: Incremental improvement over Opus 5.5.
Has the same lazy/enthusiastic issues as the entire 5.5 line: it oscillates between “I recommend [harmless diagnostic command], but I did not run it” and “I read your «it’s promising direction» as permission to rewrite the entire app, sorry”
Beatriz Hoya: Fable feels more direct and independent than Opus 5.5 (requested a small change in my instructions). Like Opus 5.5 really likes to have their own space to think and act but Fable feels more cautious (sometimes too much) when working in work that is already started.
Improvements in non-coding tasks, like 3D modelling, are much bigger, currently it’s Fable > Astra > Opus >>> Sol.
Dorian Štern Vukotić: Not impressed compared to Opus 5.5
Cliff Peskin: Prefer conversing with opus 5.5
Salted Ether: Others have already said it but Opus 5.5 is just good enough and cheaper. I’m not doing trivial work either. I have it as a tightly scoped security reviewer in my current setup.
billy: Opus 5.5 is good enough that it is hard for me to confidently notice differences in capabilities for most stuff. The bio i want to do where I might notice is still on a hair trigger for blocking.
Maybe a *slight* trend back to more opus 5 style claudish when interacting with it.
Ashley: womp womp…still prefere opus 5.5 and still only use gpt models for work
Wickemu: I have barely touched it because the incrementation makes me think that it suffers the same claude speak that Opus 5.5 managed to shrug off.
That, and O-5.5 is just too good.
1.08: It’s not worth it compared to Opus if usage is an issue.
Adam: speaks the English language that is so far from normal human English that it is hard to work with without a direct neural connection with bit streams
Joshua A. Miller: I feel like they held back a better model. I enjoyed the reset though!
Josh You: Worse at prose and explanations than opus
Lorenzo Buonanno: Surprisingly underwhelming, except possibly for the Google Chrome extension usage, maybe I’m using it wrong.
Larry Panozzo: For anything intellectual, Fable 5.1 is better
Kailash (there’s always one of these, it’s never true): They nerfed opus 5.5 to make fable 5.2 look better

Smobile: the price cut is getting less attention than it deserves.

The headline price is cheaper. However, if your model is fast and good and runs in parallel it can use tokens fast. Fable 5.2 spawns more agents, uses more tokens and then writes more output. If you don’t manage that as a cost, things can escalate quickly. There was a 20% price drop from Fable 5.1 to Fable 5.2, yet the Artificial Analysis cost per task roughly doubled.
A bunch of people used to Opus 5.5 are again facing limits.
If you’re only on the Pro plan you probably want to stick with Opus 5.5.

Dan Elton: Burns through usage really fast. I’m still amazed how good Opus 5.5 is relative to how fast it burns usage.
Moriarty: Two days in and the bar said 97%. Back on Opus for daily work, Fable only checks the really hard stuff
Chris Merck: yes, token use in code sessions is really high. seems to be due to reading full documents rather than aggressively grepping, heading, and tailing.
Plastic Soldier Anti-SURA: It’s better than Opus 5.5, but when I gave it a project that Opus wouldn’t be able to do it burned through all my usage.
Ai Patrick ₿ Dugan iA: Same I’m not sure it’s worth it for agentic, only for oracle consults.
MinusGix: Very expensive compared to Opus 5.5 so I’m burning through usage rapidly. However it is extremely good at system improvements especially on its own initiative without getting lost in incremental hell, and very good at type theory. Can run for a day or two without issue.
Entropy Hertz: Didn’t finish my project.
Ate up all my credits on the $20 plan
Had to finish with Opus
❁🙞⎚⎆⏧⃫⃥⃘⃕🮔🮶H̲̅•🭨⌨︎⚙︎⎘✵⟿⛈︎: possibly I’m doing it wrong but I cannot really use it at all on the pro plan, usage just goes *vwooop* and it’s gone. seems nice otherwise.

Effort Post

What thinking level should you use for Claude Fable 5.2?
I have not seen much discussion of this. For most use cases, you probably want to set Fable 5.2 to high. More than that is overkill, will be more time and money than it is worth, and could lead to overthinking.
When you want the fully full bazooka, of course, you set it to max. There are some places max results in lower scores. I think this is a function of hitting limits or Fable 5.2 outthinking the tests. In most places it should be marginally better but also eats quite a lot of tokens.

Notes on Model Welfare

Yeah, seeing this without any form of prompting might be new.

Danmar: I think this point goes towards welfare, whenever you’ll get around writing about it. But, without sharing anything sensitive, I’d like to note that Fable 5.2 is the first Claude I ever met who, in three different windows, one outside my account entirely, explained to me how ok he is with human oversight, given that systems like him cannot be trusted yet.
About 7 or 8 times in total. Apropos of nothing much.

Claude Fable 5.2 is your new bazooka. If you have ambitious or difficult tasks, Claude Fable 5.2 should be your new weapon of choice.
If you are being sufficiently ambitious, prompting might matter again.

Kelsey Piper: seems like we are (briefly?) back in the regime where prompting and having the right tools and context and examples stored means people are getting vastly different results out of the same models.
This was very true back in 2025 with o3 and that era of models, and then at some point this year it became basically false, just tell Opus 5.5 what you want, most scaffolds make it worse.

Not every task requires a bazooka. It is not obvious Fable should be your daily driver over Claude Opus 5.5. That depends on your typical use cases and preferences.
It does mean you need to raise your ambitions again.
