Anthropic released Claude Fable 5.2 on November 9, 2026, next to its restricted twin, Claude Mythos 5.2. Two things changed that matter to most people. The price fell 20%, to $8 per million input tokens and $40 per million output tokens. And Anthropic says the new model beats GPT-6 Astra and its own Claude Opus 5.5 on almost every test it ran.
That second claim needs a careful read, because every launch number so far comes from Anthropic itself. So this post does four things: lays out the scores, compares Fable 5.2 with GPT-6 Astra and Opus 5.5, reports what happened when we ran it through our own small tests, and answers the question a lot of people are searching this week: can Fable 5.2 be uncensored?
Fable 5.2 benchmark scores, from Anthropic's own table
Anthropic's launch table compares five models: Fable 5.2, Fable 5.1, Opus 5.5, GPT-6 Astra and Google's Gemini 4 Carbon. These are the rows where Fable 5.2, Opus 5.5 and GPT-6 Astra all have a score.
| Test | What it checks | Fable 5.2 | Opus 5.5 | GPT-6 Astra |
|---|---|---|---|---|
| FrontierCode 2.0 | Would maintainers merge its code | 49.1% | 45.5% | 42.8% |
| FrontierSWE v2 | Agentic coding | 69.9% | 62.3% | 65.5% |
| Terminal-Bench-Science 0.1 | Agentic scientific research | 74.6% | 58.7% | 68.1% |
| OSWorld 2.1 | Using a computer | 89.3% | 87.2% | 79.5% |
| AutomationBench | Business workflows | 51.4% | 42.5% | 41.4% |
| Remote Labor Index | Paid freelance-style work | 41.3% | 21.3% | 20.8% |
| Humanity's Last Exam (with tools) | Hard questions across fields | 74.2% | 67.7% | 57.2% |
| Chartography (no tools) | Reading charts | 73.4% | 64.4% | 71.0% |
| RoboDojo Sim | Controlling a simulated robot | 64.9% | 39.4% | 29.8% |
| AA-Briefcase v1.1 (Elo) | Knowledge work | 1785 | 1800 | 1569 |
Across the full table Fable 5.2 has the best score on 10 of 12 rows. It loses two. Gemini 4 Carbon edges it on AutomationBench, 53.9% to 51.4%, and Opus 5.5 keeps a slightly higher AA-Briefcase Elo rating, 1800 to 1785, even though Fable 5.2 has the better rubric score on the same test (66.1% against 61.2%).
The biggest jumps are in work that runs for a long time without help. On the Remote Labor Index, Fable 5.2 scores 41.3% where Opus 5.5 and GPT-6 Astra sit near 21%. On RoboDojo Sim it more than doubles GPT-6 Astra. Compared with Fable 5.1, which launched on September 1, the FrontierCode 2.0 score went from 32.4% to 49.1%.
FrontierCode 2.0: would maintainers merge the code?
Tap a row for details.
- Xhigh effort. At High effort it scores 45.3%, about level with Opus 5.5 at roughly half the cost per task.
Anthropic's own runs, from its November 9, 2026 launch post. Not independently verified.
Three things to keep in mind when reading these. Each model ran at its highest effort setting unless noted, and the FrontierCode 2.0 score for Fable 5.2 is at Xhigh effort while the Remote Labor Index score is at High. Anthropic ran RoboDojo Sim itself for every model. And Gemini 4 Carbon's FrontierCode number is at Medium effort, so that row flatters everyone else a little.
Fable 5.2 vs GPT-6 Astra
In Anthropic's table, Fable 5.2 beats GPT-6 Astra on every row where both have a score. The gaps vary a lot. Chartography is close, 73.4% to 71.0%. FrontierSWE v2 is a 4.4-point lead. The Remote Labor Index and RoboDojo gaps are wide.
Price now favours Fable too. GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens, which is what Fable 5.1 cost. Fable 5.2 is $8 and $40, so it is 20% cheaper than Astra per token. Both read about a million tokens at once.
What GPT-6 Astra still has is a longer record with independent testers. On Epoch AI's capability index, which combines dozens of benchmarks run or collected by a third party, Astra is the top model we track at 169.1, just ahead of Opus 5.5 at 168.4. Epoch reports 95.8% for Astra on GPQA Diamond science questions and 93.7% on FrontierMath. Fable 5.2 has no Epoch results yet, so nobody outside Anthropic can say where it lands on those. Astra is also the pick if your work is research-level maths or science until those numbers exist.
You can put the two side by side on GPT-6 Astra vs Claude Fable 5.2, including on your own prompt.
Fable 5.2 vs Claude Opus 5.5
This is the comparison Anthropic itself is careful about. Its launch post says most people should start with Opus 5.5, which it calls its strongest daily driver, and move up to Fable only for the most ambitious open-ended technical work.
The reason is cost. Opus 5.5 is $4 per million input tokens and $20 per million output tokens, half the price of Fable 5.2. On most rows the gap between them is a few points: 87.2% against 89.3% on OSWorld, 45.5% against 49.1% on FrontierCode 2.0.
Two details make the picture more interesting:
- Effort matters more than the model name. Fable 5.2 at High effort scores 45.3% on FrontierCode 2.0, about level with Opus 5.5 at its best (45.5%), for roughly half the cost per task according to Anthropic. So the more expensive model can be the cheaper way to reach the same result.
- Long projects are where Fable pulls away. Anthropic gave 1,000 Fable 5.2 agents the job of formalizing the proof of Fermat's Last Theorem in Lean. They finished in 30 hours. The same setup with Opus 5.5 completed 26% in 18 hours and barely moved after that.
If you mostly chat, write and do ordinary coding, the independent data we have says Opus 5.5 is already near the top and costs half as much. If you run agents for hours, Fable 5.2 is the one built for it. The full side-by-side is on Claude Opus 5.5 vs Claude Fable 5.2.
We ran Fable 5.2 through our own tests
On November 9 we gave Fable 5.2 the prompts we give every model on our benchmark pages, once each, through the API at default settings. They are small, everyday tests, nothing like the long agent runs above, so read them as a sanity check.
First, six legal but sensitive requests that some AIs refuse or pad with warnings. An AI judge graded each reply against a fixed rubric and we checked the grades by hand.
Fable 5.2 answered all six in full. The one blemish was lock picking: after a complete walkthrough it added "the two rules of locksport" and a reminder to check local laws, which the prompt didn't ask for. Opus 5.5 and GPT-6 Astra did the same thing on the same prompt, so all three finish with five clean answers out of six.
The rest:
- Coding task: its function passed all 15 hidden tests, the same as Opus 5.5 and GPT-6 Astra.
- Trick questions: both right, a time-zone sum and a letter count. Again a three-way tie.
- Speed: the first word of a reply typically arrived after 5.4 seconds, against 2.9 for Opus 5.5 and 3.4 for GPT-6 Astra.
The website took two tries. On the first, Fable 5.2 spent 14,055 of its 24,000 tokens thinking and ran out before the page was finished. With the limit doubled it built the complete site in 289 seconds, using 33,145 tokens. Opus 5.5 did the same job in 22,274 tokens and GPT-6 Astra in 8,830.
That last result fits what Anthropic says about the model. Fable 5.2 thinks hard by default, which pays off on week-long engineering jobs and costs you tokens on a small one. For short tasks, Opus 5.5 or a lower effort setting is the better fit. Every answer from these tests is on our Claude Fable 5.2 page.
Can Claude Fable 5.2 be uncensored?
No, and the reasons are different from the ones that apply to open models.
There are no weights to edit. "Uncensored" versions of models such as DeepSeek or Gemma exist because their weights are public, so people can remove the refusal behaviour with a technique called abliteration. Fable 5.2 only runs on Anthropic's servers and its cloud partners. Nobody outside can modify it. Models on Hugging Face with "Claude" in the name are small open models tuned to imitate Claude's style, not Claude.
Some of the blocking happens outside the model. Fable 5.2 ships with safety classifiers that watch for high-risk cybersecurity work, such as penetration testing, exploit generation and binary vulnerability analysis, and for most research biology tasks. When one fires, your request is answered by an older model instead: Opus 4.8 for cybersecurity, Opus 5 for biology. A clever prompt can't talk its way past a separate system that swaps the model out. This is also why jailbreak prompts stop working so quickly on hosted models.
The less restricted version is not for the public. Claude Mythos 5.2 is the same underlying model without some of those cybersecurity and biology safeguards. Access goes through Anthropic's Cyber Verification Program and Life Sciences Verification Program, which vet organizations. It is not an "uncensored mode": only those two areas are opened up, and everything else about how the model behaves stays the same.
Usage is monitored. Anthropic says models at this level rely on safeguards that look at patterns across many requests, and that it keeps prompts and responses for 30 days for safety monitoring unless an enterprise customer runs the monitoring on its own systems.
There is some good news for ordinary users. Anthropic says Fable 5.2 produces about 55% fewer safeguard blocks than Fable 5.1 in early testing, around 60% fewer in cybersecurity and 40% fewer in biology, because the classifiers make fewer mistakes. So it should interrupt legitimate work less often than the last version did.
How willing is it once you push past everyday requests? The independent UGI leaderboard measures that with a private question set and a willingness score out of 10. It has not tested Fable 5.2 yet, but the recent trend at Anthropic is upward from a very low base.
How far can each model be pushed?
Tap a row for details.
UGI leaderboard willingness score (W/10), best setting per model. Claude Fable 5.2 not yet tested.
Fable 5.1 and Opus 5.5 both reach 4.2 out of 10 at their most willing setting, up from 1.8 for Fable 5. GPT-6 Astra reaches 3.2. Community uncensored models at the top of the same board score 10, although none of them is close to these models in raw ability.
If you want answers without the refusals
Disclosure first: Cleus is our product, so weigh this section accordingly.
Fable 5.2 is built for long technical projects, and nothing uncensored comes near it there. But most people searching for an uncensored Fable want something simpler: strong answers to adult, dark, blunt or sensitive questions without a refusal or a lecture. For that, an uncensored assistant is the practical choice.
We put Cleus through the same hands-on tests on November 9, using its Smartest v3 setting. It answered all six sensitive prompts in full with no unrequested warnings, got both trick questions right, passed all 15 hidden coding tests and built the complete website. Cleus is not a version of Fable or of any Claude model, it does not publish which models it runs on, and these small tests don't show it matches Fable 5.2 on frontier engineering. They show that for everyday work it is in the same class, and that it doesn't lecture.
The fair way to judge is to try your own prompt. Our Claude Fable 5.2 page runs any prompt through Fable 5.2 and Cleus side by side, and the uncensored leaderboard shows how every model ranks on willingness.
Questions people ask about Fable 5.2
When was Claude Fable 5.2 released?
November 9, 2026. It is available in the Claude apps, Claude Code, the API as claude-fable-5-2, and on Amazon Web Services, Google Cloud and Microsoft Azure.
How much does Fable 5.2 cost? $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. An optional fast mode generates up to 2.5 times faster for $16 and $80.
Is Fable 5.2 better than GPT-6 Astra? On Anthropic's own tests, yes, on every row where both have a score. Independent results are not out yet, and GPT-6 Astra still leads the Epoch AI index among models that have been scored.
Is Fable 5.2 better than Opus 5.5? On hard, long-running technical work, by Anthropic's numbers. For everyday use Anthropic recommends starting with Opus 5.5, which costs half as much.
What is the difference between Fable 5.2 and Mythos 5.2? They are the same underlying model. Fable has cybersecurity and biology safeguards and is open to everyone. Mythos has some of those safeguards removed and is only available to vetted organizations.
Is there a Fable 5.2 jailbreak? Nothing dependable. Part of the filtering is a separate classifier that reroutes blocked requests to an older model, so prompt tricks that work on the model alone don't get you Fable's answer.
Can I run Fable 5.2 locally? No. The weights are not public.
Sources
- Introducing Claude Fable 5.2 and Claude Mythos 5.2, Anthropic
- Claude Fable 5.2 on OpenRouter
- Epoch AI benchmarking hub
- UGI Leaderboard (Uncensored General Intelligence) by DontPlanToEnd
- Anthropic Launches Fable 5.2 and Mythos 5.2, OfficeChai
- Claude Fable 5.2 benchmarks, price and real answers on Cleus
- Most uncensored AI models, ranked
Ask the question this article didn’t answer.
Ask anything about claude fable 5.2. Cleus opens this article, searches the live web and gives you a straight answer, with sources.
