Claude Fable 5.2: Smarter Than Opus 5.5, and I Still Wouldn't Switch

November 9, 2026, launch evening

For weeks, everyone who works with Claude has been waiting for the same thing: Fable 5.5. After what Opus 5.5 did in September, it's obvious to everyone what should come next. Then Anthropic posts its announcement, and the number in the headline is 5.2.

Within minutes, one of the most common questions on X is a one-liner: wait, why Fable 5.2 and not Fable 5.5?

After a night of working with it and reading every test I could find, I think that number is the most accurate part of the launch.

Claude Fable 5.2 is the newest and most powerful model Anthropic sells, and it took first place on almost every leaderboard that measured it. Still, my recommendation is not to switch to it from Claude Opus 5.5. It's smarter, but by a small to medium margin, it costs more, and it's less polished: in my work it stops mid-task more often than Opus, argues, and sometimes gets stuck on precisely the simple things.

This article draws on three kinds of evidence, and I flag which is which along the way: what Anthropic published itself, what independent organizations and developers measured in the first hours, and what I think. The last part, about what Anthropic is keeping in the drawer, is my interpretation.

In short

Claude Fable 5.2 came out on November 9, 2026 and costs $8 per million input tokens and $40 per million output tokens, exactly double Opus 5.5. It is first on the Artificial Analysis index (65 against 59 for Opus 5.5), on Epoch, on Vals and on CursorBench.
The jump in capability is small to medium. In science and knowledge, the gap is large. In code, which is what most developers work on, it and Opus 5.5 are nearly tied.
It's less polished. Anthropic itself writes that under some conditions the model stops and returns an answer before the work is done, and asks for approvals it doesn't need. And it still recommends that most users start with Opus 5.5.
The real price: at the same effort level, a task on Fable costs 1.6 to 3.4 times as much. My estimate for normal work is at least 2.5x, and on a subscription the weekly usage limit runs out fast.
My opinion: this launch was meant to win back first place on the leaderboards, which Google had taken 12 days earlier. If Opus 5.5 works well for you today, stay with it. And a business owner whose agent already works doesn't need to change anything.

What Claude Fable 5.2 is, and what it costs

Fable is the tier Anthropic places above Opus, for the longest and most open-ended technical tasks. Fable 5.2 came out together with Claude Mythos 5.2, and according to Anthropic the two are "the same underlying model." The difference is that Fable includes safeguards in cybersecurity and biology, so it's open to everyone, while Mythos is available only to verified organizations.

$8 / $40Price per million input and output tokens. Fable 5.1 cost $10 / $50, and Opus 5.5 costs $4 / $20

Up to 50%The share of the weekly usage limit that Max and Ultra subscribers can spend on Fable. On Pro it's paid separately, except during launch week

Up to 30%Faster than Fable 5.1 at generating answers, according to Anthropic. The context window, the amount of text the model holds in mind at once, stays at one million tokens

Sources: Anthropic's launch announcement, the API price list (the API is the direct connection for developers, paid by usage) and the support page on Fable models in subscription plans (checked on November 10, 2026). If the term is new to you, we explained what a token is in a separate article.

Two sentences from Anthropic's announcement sum up the whole story. The first: Fable 5.2 is "a step above Opus 5.5 in almost all the capabilities we evaluated." The second, a few paragraphs later: for most coding and knowledge tasks, Anthropic recommends starting with Opus 5.5, which it calls its strongest daily driver, and moving up to Fable only for the most ambitious, open-ended technical work.

In other words, even the company selling the model recommends that most people start with a different one.

Why everyone was waiting for Fable 5.5

To understand the disappointment with the number, you have to go back two months. Claude Opus 5.5, released on September 22, became within weeks the everyday model for most of the developers and AI people I know. I wrote then that it was the strongest by far of anything I had worked with this year, and my opinion has only grown stronger since.

In my view, what worked in Opus 5.5 wasn't just intelligence, it was the polish. As I understand it, the polish happens in the training stage where the model practices tasks, gets scored on the result and improves (reinforcement learning, or RL). In Opus 5.5 that stage feels almost perfect. It's good to talk to, it takes initiative, it finishes what it starts, and it rambles and gets confused far less than its predecessors.

So when the 5.5 versions of Opus, Sonnet and Haiku came out one after another, the conclusion was obvious: next in line was Fable 5.5, the same polish on the big model. Instead, we got 5.2.

From Fable 5 to Fable 5.2: five months

June 9Claude Fable 5A new tier above Opus, $10 / $50 per million tokens
July 24Claude Opus 5Smart, but long-winded and confusing. Many developers didn't like working with it
September 1Claude Fable 5.1
September 22Claude Opus 5.5The model that put Anthropic back on top. Followed by Sonnet 5.5 and Haiku 5.5
October 28Google's Gemini 4 CarbonFirst on the Artificial Analysis index, at 59 according to Google. Open only to a closed groupAnthropic loses first place
November 9Claude Fable 5.265 on the updated version of the index, against 60 for Carbon

Sources: Anthropic and Google announcements, and Artificial Analysis. The description of Opus 5 is my impression, and I wrote about it in the article on Anthropic's comeback.

rezoundous's post on X, November 9, 2026

The question many people asked on launch evening. Anthropic has not published an explanation for the version number.

How much smarter Fable 5.2 is than Opus 5.5

In my use, most of the time you can feel a small to medium step up over Opus 5.5: in visual building, in 3D and in tasks that require planning several steps ahead, and sometimes in code too. The independent measurements from the first hours show a similar picture, with big differences between fields. Every test lists an "effort level" (Low, Medium, High, Max): how long the model thinks before it answers. The higher it is, the better the result usually is, and the more it costs.

Fable 5.2 vs. Opus 5.5 in independent measurements, both at maximum effort

Test	Claude Fable 5.2	Claude Opus 5.5
General indexes
Overall indexArtificial Analysis Intelligence Index	65	59
Professional workVals Index	71.3%	70.1%
Code and agent work
Real coding tasksCursorBench 4.0	60.0%	57.8%
Command-line workTerminal-Bench 4.0	59.6%	59.6%
Long coding tasksFrontierCode 2.0	48.8%	46.4%
Office work judged by expertsAA-Briefcase, Elo	1785	1801
Science and knowledge
Science on the command lineTerminal-Bench-Science	73.3%	59.0%
Expert questionsHLE-Diamond	80.2%	67.9%
Knowledge without making things upAA-Omniscience	67	48

Top score in the row. Sources: Artificial Analysis (index version 5.0.1 and the comparison page for the two models), Vals AI and Cursor, as of November 10, 2026. On Vals, Fable was measured at High. Artificial Analysis has updated its index since October, so the numbers here can't be compared with those in our earlier articles.

Three things stand out in the table:

In science and knowledge, the gap is real. 12 to 19 points on three different tests. If you work on research, scientific analysis or questions that need deep knowledge, you'll probably feel a real difference here.
In code, it's nearly a tie. On Terminal-Bench the scores are identical down to the decimal. On CursorBench, at Max it leads by 2.2 points, and at High by less than a point. At both levels it costs 2.4 times as much.
In office work judged by experts, Opus 5.5 is slightly ahead. Even in Anthropic's system card (its document of tests and safety checks), the only two rows where Fable 5.2 isn't first among the Claude models are Elo ratings for knowledge work: a rating in which judges compare the outputs of two models side by side.

Anthropic's own table paints a rosier picture: Fable 5.2, or Mythos 5.2 on the cyber and biology rows, leads on almost all 15 rows, including 69.9% against 62.3% on FrontierSWE and 89.3% against 87.2% in computer use. These are numbers the maker measured itself, and I haven't yet seen an independent check of them.

What developers saw in the first hours

In visual building and 3D, the most beautiful results I saw came from Fable 5.2. The developer notjazii, who compares models on the same prompts, got a 3D pirate ship in a single attempt and wrote that it was the craziest output he had seen for that prompt. The price: an hour and a half, 35 million tokens and about $34.

notjazii's post and video on X, November 9, 2026

A 3D pirate ship in one attempt. About an hour and a half and about $34.4. His test, not ours.

When the same developer asked both models, at the same effort level, to build an Xbox controller, the picture was less clear-cut. In his words, the results came out fairly similar, Fable added an impressive amount of detail, and for the price Opus is the better deal. Opus finished in about 15 minutes, for about $8.2. Fable thought for about 20 minutes before it started, finished after about an hour in total, and cost about $18.

notjazii's post and comparison video on X, November 9, 2026

Same prompt, same effort level: Opus 5.5 in about 15 minutes for $8.2, Fable 5.2 in about an hour for $18.

Theo, a developer and content creator with a large developer audience, ran his regular game test and came to the same conclusion: one of the best runs he had seen, but "not significantly better than Opus 5.5 (esp. for the price)." 76 minutes, about $42.

Theo's post and video on X, November 10, 2026

A very good result, without a big gap over Opus 5.5. His opinion.

My problem with it: it isn't polished

This is where I get to the reason I don't recommend it. Anyone who worked with Opus 5 over the summer knows the feeling: a smart model that isn't pleasant to work with. Fable 5.2 reminds me of that. Much milder than Opus 5, but from the same family.

In my work, it looks like this: more often than with Opus 5.5, Fable doesn't finish a task in one go. It stops midway to ask for clarification on something with an obvious answer, writes a lot along the way, and sometimes gets stuck or fails "for no reason" on a simple task that Opus 5.5 handles without a second thought. On one hard task it's brilliant, and on the easy task right after, it misses.

That's my impression, so I checked whether it shows up elsewhere. Part of it does, and in a place I didn't expect: the prompting guide Anthropic published for the model.

What Anthropic writes in the model's guide

Official documentation, November 9

On long tasks, the model sometimes stops and returns an answer before the work is done, with a list of what remains or an offer to continue. This happens more at low and medium effort
It may ask for approvals it doesn't need and hold back on routine work, mainly when there are lots of caution instructions or several agents working together
In one test scenario it didn't retry in about a third of cases, where Fable 5.1 and Opus 5.5 did
At high effort it thinks before almost every reply, even before replying to "hello"

What developers wrote on day one

Posts on X and blogs, November 9 and 10

Theo: the model is making "a lot of stupid mistakes I would not have expected from Opus"
Gegam: "On some tasks it's a genius; on others, it's as dumb as a shrimp," and "the most pedantic model yet"
Simon Willison: at maximum effort the model spent 128K tokens thinking and returned no drawing at all. He says this failure is new for the Fable series
Sholto Douglas of Anthropic: "it is so bad at animation." A colleague's reply: "working on this"

Anthropic also offers a fix: a prompt wording that asks the model to finish the whole task. In its testing, the wording halves the number of tasks that stop midway, and raises token use by about 30%. So it can be improved, and you pay for the improvement.

Theo's post on X, November 10, 2026

A few hours after launch. One developer's impression.

Two caveats, to be fair. First: there's no independent, structured measurement of mid-task stops yet. What exists is Anthropic's documentation, developers' posts and my own experience, after one night. Second: the common complaint on X isn't exactly mine. People write that it's pedantic, argues and thinks a lot before it starts. I didn't find a single post complaining that it stops to ask. Anthropic presents the arguing as a strength: the model "will push back" when your idea isn't good.

Bottom line, Opus 5.5 is reliable. You give it a task, and it finishes it. That's what matters most to me in a work tool, and it's exactly where Fable 5.2 falls short.

What Fable 5.2 really costs

On paper, the answer is simple: double. $8 and $40 per million tokens against $4 and $20. In practice, the price per task depends on how many tokens the model uses to finish it, and that's where the argument starts. Anthropic writes that on the FrontierCode test, Fable 5.2 at High matches Opus 5.5 at Max "for roughly half the cost per task." On X, too, I saw a test where it finished the same task with 30% fewer tokens.

The test that best shows what's going on is CursorBench, because Cursor publishes both a score and an average cost per task for every model, at every effort level:

Source: Cursor, CursorBench 4.0 leaderboard, as of November 10, 2026. Cost is calculated from each model's official price list and the tokens it used on each task. The chart and markings are ours.

There are four ways to look at the cost, and they're all true at the same time:

I have $4 per task. Which model gives me more?

Opus 5.5.

At $3.97, Opus at High gets 56.0%. At $3.94, Fable at Low gets 53.2%. Up to about $9 per task, Opus is better. Between $10 and $13, Fable at High leads by less than a point. Fable beats Opus's best score only at Max: 60.0%, at $32.79 per task.

Both models on the same setting

Fable costs 1.6 to 3.4 times as much.

On CursorBench: 1.6x at Medium, 2.4x at High and Max, 3.4x at Low. On the Artificial Analysis index, at Max: $26.5 per task against $10, or 2.6x. There Fable also used more output tokens, 274K per task against 173K.

Fable at High vs. Opus at Max

Here Fable is cheaper, at roughly the same quality.

On CursorBench: $9.70 against $13.43, with a score 0.9 points lower. On Artificial Analysis: $9.08 against $10.04, with a score 3 points higher and about 40% fewer output tokens. On Vals: $17.94 against $24.77 (Opus's effort level wasn't stated there). This is the comparison behind Anthropic's "half the cost," and behind the claim that it does the work with fewer tokens.

A monthly subscription, not the API

The weekly usage limit runs out fast.

Anthropic only writes that Fable models use up the limit "faster than other Claude models," without a number. Theo calculated that the $200-a-month Ultra plan gets you about $15,000 worth of work on Opus 5.5, and only about $3,000 on Fable. A Hacker News user reported burning through two Fable limit resets in under three hours.

Sources: Cursor, Artificial Analysis, Vals AI, Anthropic's support page, Theo on X and the Hacker News launch thread. The calculations by Theo and the Hacker News user are theirs, not measurements of ours.

Theo's post on X, November 10, 2026

The post says "Fable 5.5" by mistake, and Theo corrected it in a follow-up post: he meant 5.2. The calculation is his.

So does Fable cost the same as Opus? Only if you run it in a lower gear and give up part of its advantage. If you just swap the model name and leave everything else as it is, you pay more. Also worth knowing: in the API, Fable 5.2 defaults to High effort and Opus 5.5 to Medium, so a "clean" swap also raises the effort level without you noticing. In the apps and in Claude Code, Fable defaults to Medium.

My estimate, from everything I've seen and used: in normal work, the cost will come out at least 2.5 times that of Opus 5.5. That matches the Artificial Analysis index, where at the same effort level Fable costs about 2.5 times as much, and CursorBench, where at High and Max it costs 2.4 times as much. In some cases the gap can be much bigger, 5x or even 10x, mainly on projects the two models approach in completely different ways, with one of them taking the long road. In developers' tests, the range is wide. In 52 runs on real code on the paddo.dev blog, with both models at Medium, Fable cost 1.63 times as much. In Atomic Agent's test, swapping only the model name in the API cost 7.9 times as much, on a single task. I found no documentation of a 10x gap.

Who Fable 5.2 is for, and who it isn't

Worth trying

Research, science and analysis that need deep knowledge, where the gap over Opus 5.5 is large
Visual building and 3D, when the result matters more than time and price
One long, open-ended task that other models have failed at
As an "advisor" that plans while a cheaper model executes

Check first

The effort level: most of the advantage is already there at High. Max costs 3 times as much or more
The subscription limit: up to 50% of the weekly limit on Max and Ultra plans, and it runs out fast
Data retention: in the API, Anthropic keeps the data for 30 days, and there's no option for it not to be stored at all

Stay with Opus 5.5

Everyday coding, where the task needs to finish without supervision
An agent that runs on its own for hours, with no one there to answer its questions midway
Office and document work, where Opus is slightly ahead in the judges' ratings
Any workflow with a fixed budget

For a business owner whose agent already works, the practical takeaway is short: nothing to change this week. An agent that answers customers, manages a website or prepares reports needs, above all, to finish what it started, and there Opus 5.5, Sonnet 5.5 and even Haiku 5.5 do a good job at a much lower price. Fable 5.2 is worth testing only if you have one especially hard task that your current model fails at.

Platin Digital is an OpenAI Select Partner, and we build the agents we set up with the models and tools that best fit each business, from OpenAI, Anthropic, Google and others. Fable 5.2 is going into my testing for research tasks and visual building. For everyday work I'm sticking with Opus 5.5, and I'll test again when Anthropic releases an update, or when the model in the drawer comes out.

Why now: the leaderboards, in my opinion

From here on, this is my interpretation.

If the model isn't polished, and Anthropic itself recommends staying with Opus 5.5, why release it now? My answer is in the timeline. On October 28, Google announced Gemini 4 Carbon, a big sibling of Gemini 4 Argon, and said it was first on the Artificial Analysis index. Google had passed Opus 5.5 by one point, and the top model on the table was no longer Anthropic's. Twelve days later, Fable 5.2 came out:

Ranking	First before November 9	First after
Artificial Analysis, overall index	Gemini 4 Carbon, 60	Fable 5.2, 65
Epoch Capabilities Index	GPT-6 Astra	Fable 5.2, 174
CursorBench 4.0	Opus 5.5, 57.8%	Fable 5.2, 60.0%
Arena, website building	Opus 5.5	Fable 5.2, 1832 against 1818, within the margin of error. Automated evaluation, no votes yet
Arena, text	Gemini 4 Carbon, 1526	Gemini 4 Carbon (Fable 5.2 not measured yet)
Arena, agents	Opus 5.5	Opus 5.5 (Fable 5.2 not measured yet)

Sources: Artificial Analysis, Epoch AI data, Cursor and Arena, as of November 10, 2026.

And here's a detail that's easy to miss: on launch day, Carbon wasn't yet available to the public. It was open only to a closed group of cyber defense organizations, and it had no API. Even so, in the comparison table Anthropic published at launch, Google's column is Gemini 4 Carbon. Anthropic compared itself to a model almost no one can run. In my opinion, because that was the model sitting in first place.

So I read this launch as a leaderboard move: release something that wins back first place, even if it isn't ready yet to be the everyday work tool. And I think this is a recurring pattern. Opus 5 in July was also smart on paper and unpleasant to work with. This is my theory. Anthropic has not linked the timing to Google anywhere.

There's also a competing explanation, and it doesn't contradict mine: according to a report in The Next Web, Anthropic filed paperwork on October 26 ahead of an IPO. A company on its way to the stock market wants to be first on the leaderboards.

What I think Anthropic is keeping in the drawer

I think Fable 5.5 exists. That Anthropic knows exactly how good it is, and that it's waiting for the right moment to release it.

Part of this is no longer a guess. In the Fable 5.2 system card, a 264-page document, Anthropic writes for the first time that it has "a separate, unreleased internal research model that is somewhat more capable than Claude Mythos 5.2," and even publishes its score on Anthropic's internal index: 178.78, against 177.12 for Mythos 5.2 and 174.78 for Opus 5.5. And in the launch announcement, it says that a formal proof of Fermat's Last Theorem was produced by "an unreleased research version of Claude."

Verified

Anthropic has an internal model, stronger than Mythos 5.2, that hasn't been released. The system card says so
During training, Anthropic deliberately capped Fable 5.2's capabilities in areas that speed up AI development, at roughly Opus 5.5's level
Its protections against competitors copying the model are the strongest it has deployed to date

Rumors

Researchers who read the card wrote that the tested version is called "Model 3 Early" there, and that an earlier report also mentioned a "Model 4". I haven't checked this myself
A claim that "Mythos 5.5" briefly appeared in the website's code. Not verified
Several posts that said "Fable 5.5" were simply typos

My hypothesis

The model in the drawer is Fable 5.5, meaning the same large model with the polish of the 5.5 generation
It has existed for months already, maybe half a year
If that's true, what Anthropic is training today is even further ahead of what we see

Peter Wildeford's post on X, November 10, 2026

A researcher's inference from the system card: Anthropic has stronger internal models, and in his view they're equivalent to Fable 5.5. Anthropic has not said when, or whether, they will be released.

This ties into a theory I wrote about in the article on Anthropic's comeback: labs train their next models with the help of the strongest model they have, so the model outside is always a step or two behind the model inside. There I speculated that Fable served as a teacher for the 5.5 generation. If that's true, then even when Opus 5.5 came out, people inside Anthropic were working with something stronger. And what we're seeing now as 5.2 is, in my opinion, a card the company was willing to show, not its strongest card.

Where Fable 5.2 stands in Sigma AI Ranking

Fable 5.2 entered Sigma AI Ranking, our ranking that combines several open sources into one score, on November 10, marked "Preliminary score." The reason: so far only three sources have measured it, and our rule is at least five before a model gets a place in the table.

Sigma AI Ranking: weighted score, as of November 10, 2026

Claude Fable 5.2 · Preliminary score

91.7

Claude Opus 5.5

90.2

Gemini 4 Argon

87.6

GPT-6.1 Sol

86.4

Claude Fable 5.1

85.8

Claude Sonnet 5.5

85.3

Source: Platin Digital's Sigma AI Ranking, which weighs open sources (Epoch AI, Arena, Terminal-Bench and official price lists). Under the ranking's rules, a gap of less than 2 points from the leader counts as a tie.

What's behind the 91.7:

General capability: first place. On Epoch AI's index it scores 174, a record, against about 169 for Opus 5.5 and GPT-6 Astra. On the SimpleQA fact test, too, it answers 94% of questions correctly, against 72% for Opus.
Cost: low score. The ranking calculates by price per token, and Fable costs double what Opus does. That lowers its score, as it should.
Agent work and writing: not measured yet. It has no Terminal-Bench result yet, and Arena hasn't published text and agent rankings for it. For now, the ranking fills in these two areas with an estimate based on general capability, and marks it with ≈.

In other words, the very area where I feel it's weaker than Opus, long agent work, is the area the ranking doesn't measure yet, so the estimate there leans in its favor. Under our rules, 1.5 points above Opus 5.5 is a tie at the top. When the Arena and Terminal-Bench results come in, the score will update, and I expect the gap to narrow.

Gemini 4 Carbon has also entered the ranking. When we wrote this article it had no public price yet, so its score was preliminary and didn't include cost. On November 10, Google opened it to everyone, with a price, and we wrote about what that means here. Why we built our own ranking, and why it relies only on open sources, we explained in the article on Sigma AI Ranking.

Further reading: Anthropic's comeback and Opus 5.5, Gemini 4 Argon: top marks on familiar tests and why we built Sigma AI Ranking.

Frequently asked questions

What is Claude Fable 5.2?

Claude Fable 5.2 is the most powerful model Anthropic offers to the public, one tier above Claude Opus 5.5. It came out on November 9, 2026 together with Claude Mythos 5.2, which is the same model without some of the safeguards and is open only to verified organizations. It has a one-million-token context window and is built for long, open-ended technical tasks.

How much does Claude Fable 5.2 cost?

In the API: $8 per million input tokens and $40 per million output tokens, double Claude Opus 5.5 and 20% less than Fable 5.1. On subscriptions: Max and Ultra subscribers can spend up to 50% of their weekly usage limit on it. On Pro it's paid separately, except during launch week, through November 16, 2026.

Is Fable 5.2 better than Opus 5.5?

In capability, yes, by a small to medium margin: 65 against 59 on the Artificial Analysis index, and a large gap on science and knowledge tests. On coding tests the two are nearly equal, and in office work judged by experts Opus 5.5 is slightly ahead. In everyday work, Opus 5.5 is more reliable at finishing tasks, and Anthropic itself recommends that most users start with it.

Why did Anthropic call it 5.2 and not 5.5?

Anthropic hasn't published an explanation. My interpretation: the model didn't get the same polish that Opus, Sonnet and Haiku got in the 5.5 generation, and the number reflects that. According to the system card, Anthropic has a stronger internal model that hasn't been released, and I suspect it will be Fable 5.5.

Does Fable 5.2 use up more of my Claude subscription limit?

Yes. Anthropic writes that Fable models use up the limit faster than other Claude models, without publishing a ratio. By the developer Theo's calculation, the $200-a-month Ultra plan gives you about $15,000 worth of work on Opus 5.5, and about $3,000 on Fable. A limit reset given at launch restores only Fable usage, not the whole weekly limit.