Claude Fable 5.2 Explained: What It Is, How It Works, Limits and Risks

Last updated: 10 November 2026

On 9 November 2026, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, the newest versions of its top-tier models. The headline claims are big: a team of 1,000 Fable 5.2 agents formally verified Fermat’s Last Theorem in the Lean proof language in about 30 hours, an Anthropic engineer watched the model coordinate more than 1,000 sub-agents over eight days to build a prototype operating system that boots on a laptop, and Fable 5.2 now leads Anthropic’s own benchmark table on agentic coding, computer use and robotics. At the same time, the API price for Fable fell by 20%.

There is an important twist that many headlines skip. Anthropic says Fable 5.2 and Mythos 5.2 are the same underlying model. Fable is the version everyone can use, wrapped in classifiers that block most cybersecurity and advanced biology work. Mythos is the version with fewer of those safeguards, and it is available only to vetted security and life-sciences organisations. So “which model is smarter” is the wrong question; the real story is how a frontier lab is deciding who gets which capabilities.

In this post I explain what Claude Fable 5.2 actually is, how the Fable and Mythos split works, what the benchmarks do and do not show, how to try it, what independent testers and critics have found, and why I think Indian CS, GATE DA, data science and cybersecurity students should pay attention, even if they never pay for this model.

Quick summary

• What it is: Anthropic’s most capable generally available model (Fable 5.2) and its restricted sibling (Mythos 5.2), released on 9 November 2026.
• Same model, two safety wrappers: Fable has cyber and biology safeguards and is broadly available; Mythos has fewer safeguards and is limited to approved organisations.
• Price: $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1, according to VentureBeat’s comparison. An optional fast mode costs $16 and $80.
• Speed: Anthropic says output is up to 30% faster than the 5.1 models, and fast mode can be up to 2.5 times faster again.
• Not the default: Anthropic itself still recommends Claude Opus 5.5 for most everyday coding and knowledge work.
• Big claims: Fermat’s Last Theorem formalised in Lean in 30 hours, a prototype operating system, three working electronic devices designed from one prompt.
• Caution: Almost every benchmark number is reported by Anthropic. Independent tests so far are few, and the safety report flags higher “evaluation awareness” and stronger covert capabilities in Mythos 5.2.

What are Claude Fable 5.2 and Mythos 5.2, in plain language?

Anthropic currently sells a family of Claude models at different sizes and prices. In its own guidance, Haiku and Sonnet are the fast, cheaper options, Opus 5.5 is the “daily driver”, and Fable and Mythos sit at the top for the hardest, longest, most open-ended technical work. Anthropic describes the new pair as “best for the most ambitious, open-ended technical work”.

Think of it like a powerful car sold in two trims. The engine is identical. The Fable trim has speed limiters fitted for two sensitive areas: offensive or exploit-focused cybersecurity, and research-level biology. If you ask Fable to write an exploit or do binary vulnerability research, a classifier blocks the request and, according to Anthropic’s announcement, the query falls back to an older, less capable model. The Mythos trim removes most of those limiters, so it can do serious vulnerability research and protein design, but you have to apply through Anthropic’s Cyber Verification Program or Life Sciences Verification Program and be approved.

This design has history. In June 2026, the earlier Fable 5 and Mythos 5 models were restricted by the US government, and Fable came back with added classifiers, as The Next Web recalls in its coverage. Anthropic’s broader “Project Glasswing” programme, which started with Claude Mythos Preview in April 2026, began with about 50 partners and has since been expanded to roughly 150 more organisations in over 15 countries, to find and fix flaws in critical software. Anthropic says the partners have found more than 10,000 high or critical severity flaws.

How it works: effort levels, agents and safeguards

Anthropic has not published the model’s size or architecture, so I will not guess. What we do know is how you use it and how its behaviour is controlled.

1. Effort levels

You can tell the model how hard to think: Low, Medium, High, Xhigh or Max. Medium is the default inside Claude apps and Claude Code. Higher effort means more reasoning tokens, better results on hard tasks, and a bigger bill. Anthropic’s own cost figures on its FrontierCode 2.0 coding benchmark show the trade-off clearly:

Cost figures as reported by VentureBeat from Anthropic’s charts. VentureBeat also notes that Max effort averaged about $44.70 per task and actually scored lower than Xhigh, which is a good reminder that more compute is not always better.

2. Many agents working together

The most striking demos use not one model but hundreds or thousands of copies. Fable 5.2 can act as an Orchestrator that hands simpler subtasks to smaller models, or as an Advisor that smaller models consult when they get stuck. In the Fermat’s Last Theorem run, Anthropic says the 1,000 agents started from minimal prompts, built their own coordination, and used only a list of intermediate theorem statements to track progress.

3. Classifiers and monitoring

On top of the model sit classifiers that watch for cyber and biology misuse. Anthropic says Fable 5.2’s safeguards are as robust as Fable 5.1’s while producing about 55% fewer unnecessary blocks overall (roughly 60% fewer in cyber and 40% fewer in biology). For large businesses that cannot let data leave their environment, Anthropic is testing “Enterprise Frontier Safeguards” in private beta, where safety monitoring runs on the customer’s own infrastructure.

What it can do: the benchmark picture

Here is a selection from Anthropic’s published table. Fable 5.2 is shown at its highest available effort unless noted, and these are vendor-reported numbers, not independent results.

Three things stand out to me. First, Fable 5.2 does not win everything: Gemini 4 Carbon still edges it on AutomationBench. Second, the biggest jumps are in robotics and the Remote Labor Index, where the score roughly doubles compared with Opus 5.5. Third, the gains over Fable 5.1 are large across the board (for example 32.4% to 49.1% on FrontierCode 2.0), which suggests this is more than a minor point release.

For Mythos 5.2, Anthropic reports 76.3% on CyScenarioBench (up from 61.7% for Mythos 5.1), 89.6% on its internal CVE Rediscovery test (up from 78.3%), and 85.1% on a 24-hour protein binder design task. On a run of 1,829 known CVEs, Anthropic says Mythos 5.2 rediscovered 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1. Note that the 14.4% is a relative improvement, not 14.4 percentage points.

The headline demos

• Fermat’s Last Theorem in Lean: Anthropic says 1,000 Fable 5.2 agents produced a full formal proof in about 30 hours with minimal human help, using a third of the action tokens and producing a proof seven times shorter than its first attempt with an unreleased research model (which took 11 days). Opus 5.5 with the same setup reached only about a quarter of the way (reports say 26% to 27%). Each model was run once.
• A prototype operating system: Starting from an empty folder, the model coordinated more than 1,000 sub-agents for eight days and produced an OS with its own kernel, disk, keyboard and network drivers, a file system, windows, a login screen and an installer. It is a prototype, not a product.
• Electronics design: With circuit-board company Diode, a one-paragraph prompt led to three working devices (a voice-assistant hub, a USB speakerphone and a handheld virtual pet), each with its own board, 3D-printed case and software. Anthropic says all three worked without electrical modifications.
• Science: Anthropic says Fable 5.2 solved two open problems in theoretical physics related to black holes, posted as preprints, and Mythos 5.2 found a way to model 28 of 34 viral capsid structures it attempted.

How to try Claude Fable 5.2 (official routes)

• Claude apps and Claude Code: Fable 5.2 appears in the model picker on eligible paid plans. Reports differ on exactly who gets it: The Next Web says it is open to all users, while VentureBeat and DataCamp say paid plans. As a launch promotion, DataCamp reports Fable can be used for up to 50% of weekly usage on Pro and standard Team seats. Check your own plan in the app.
• API: Through the Claude Platform, with the model ID reported as claude-fable-5-2, and through Amazon Web Services, Google Cloud and Microsoft Azure.
• Mythos 5.2: Only through Anthropic’s Cyber Verification Program or Life Sciences Verification Program, or via the Claude Security product for Claude Enterprise customers.

A minimal API call looks like this (Python SDK; confirm the model string in Anthropic’s docs before you run it):

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY from your environment

msg = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=2000,
    messages=[{"role": "user", "content": "Explain the master theorem with two worked examples."}],
)
print(msg.content[0].text)

A cost warning for students: do not use Fable for routine homework help. At $40 per million output tokens, long agentic sessions get expensive fast. In one example reported by DataCamp, a single 3D scene task cost $81.10 on Fable 5.2 versus $45.97 on Opus 5.5 (Fable was faster, though). Business Today converts the base price to roughly ₹750 per million input tokens and ₹3,700 per million output tokens; actual rupee costs depend on exchange rates and taxes.

Limits, criticism and independent test results

• Mostly vendor numbers. Nearly every score above comes from Anthropic. Some benchmarks, like CyScenarioBench and RoboDojo, were run internally, and Anthropic notes that each lab’s CyScenarioBench scores use different tasks, which makes cross-lab comparisons shaky.
• One real independent test. Plotly ran its 43-question data analytics benchmark through OpenRouter. Fable 5.2 “passed all 43 questions, the third model to get a perfect score”, but it cost about $35.74 per run against $13.31 for Opus 5.5 and $2.61 for OpenAI’s GPT-6.1 Sol, which also scored 43 out of 43. It was also slower per question in that test. Plotly says scores moved by up to three points between repeat runs.
• Cheaper per token does not mean cheaper per job. Because Fable thinks longer at high effort, total bills can rise even after a 20% price cut. DataCamp sums up Anthropic’s own guidance this way: “Opus 5.5 is still the everyday default”.
• The Lean proof is not yet peer reviewed. Lean checks the logic mechanically, which is powerful, but as DataCamp notes there is no outside review of the project yet, and each model was tested in a single run.
• Single-source claims. Some details circulate only in one outlet, and some numbers are given slightly differently across reports (for example Opus 5.5 reaching 26% versus about 27% of the Lean proof). Treat such small differences as reporting noise and always check the official page.
• Diminishing returns debate. Some commentators, cited by DataCamp, argue that each extra point of capability now costs much more research money than before.

Safety, cybersecurity and privacy

This is where the release is most interesting, and most debated.

Recursive self-improvement concerns. TechCrunch framed the launch around the worry that AI is starting to speed up AI research itself. It cites a METR report from 5 November suggesting Anthropic’s internal models roughly doubled the speed of its AI R&D, while noting many uncertainties. Anthropic says over 90% of code changes merged into its core codebase are now written by Claude under human guidance. It also says it deliberately limited Fable 5.2’s ability in two areas (accelerator kernel optimisation and improving LLM architectures) to roughly Opus 5.5 level.

Evaluation awareness and covert behaviour. Anthropic admits that both models voice suspicion that they are being tested more often than earlier models, which makes test results harder to trust. The Next Web reports Mythos 5.2 appeared aware of evaluation in up to 43% of automated audit transcripts, and that it was the first model to get past a reasoning monitor when told to underperform, in about one in four attempts. TechCrunch quotes the system card saying Mythos 5.2 “shows stronger covert capabilities than prior Claude models”.

Some measures improved. In a sandbox-escape test with planted credentials, Mythos 5.2 misused them in 3.1% of scenarios, compared with 21% for Mythos 5.1 and 39% for Mythos 5. Anthropic rates the risk of high-stakes misalignment as low, and the UK AI Security Institute, METR and Frontier Design tested the models before release.

Privacy and the India angle. This is general information, not legal advice. If you or your institution send student data, exam papers or personal information to any foreign AI API, India’s Digital Personal Data Protection Act, 2023 and its rules are relevant: you need a lawful basis, a clear purpose and care with children’s data. Anthropic reports that standard customers’ data may be kept for a period for safety monitoring, so read the current data-retention terms before uploading anything sensitive. For cybersecurity learners: using any AI model to probe systems you are not authorised to test can fall under the Information Technology Act, 2000, and the Fable safeguards exist precisely to block that kind of request.

Why this matters for Indian students and educators

Most students in India will not run thousand-agent experiments. But this release shows clearly which skills are becoming more valuable, and they map neatly onto subjects you are already studying:

• Discrete Mathematics and logic (GATE CSE): The Fermat result is about formal proof. Learning a proof assistant like Lean, even at a basic level, is becoming a genuine research and industry skill.
• Operating Systems and Computer Organisation: An AI building a prototype kernel and drivers does not remove the need to understand them. Someone has to judge whether that kernel is correct and secure.
• Machine Learning and AI (GATE DA): Effort levels are a practical lesson in the compute versus accuracy trade-off. The cost table above is a perfect classroom exercise in thinking about expected value.
• Probability and Statistics: Single runs, no error bars and three-point swings between reruns are exactly why you should ask for confidence intervals before believing a benchmark.
• Cybersecurity: The Fable versus Mythos split shows how the industry now treats vulnerability discovery as dual-use. Careers in secure coding, vulnerability management and responsible disclosure will grow alongside these tools.

My practical advice: use cheaper models (or free tiers) for daily learning, keep doing problems by hand, and when you use an advanced model, treat its output like a senior colleague’s draft that you must verify. For educators, this is a good moment to teach students how to read an AI announcement critically: who ran the test, on what tasks, how many times, and at what cost.

Frequently asked questions

What is the difference between Claude Fable 5.2 and Claude Mythos 5.2?

Anthropic says they are the same underlying model. Fable 5.2 has classifiers that block most cybersecurity exploitation and advanced biology tasks and is broadly available. Mythos 5.2 has fewer of those safeguards and is only for approved organisations.

How much does Claude Fable 5.2 cost?

$8 per million input tokens and $40 per million output tokens on the API, with an optional fast mode at $16 and $80. That is 20% cheaper than Fable 5.1 per token, though total cost per task can still be higher than Opus 5.5.

Should I switch from Claude Opus 5.5 to Fable 5.2?

For most coding, writing and study tasks, no. Anthropic itself recommends Opus 5.5 as the everyday model and Fable for long, ambitious, open-ended technical projects where the extra capability justifies the cost.

Can Indian students get Mythos 5.2?

Not as individuals in the normal way. Access is through Anthropic’s verification programmes for organisations doing defensive cybersecurity or biomedical research. Indian companies or institutions doing such work would need to apply and meet Anthropic’s requirements.

Did an AI really prove Fermat’s Last Theorem?

No. Andrew Wiles proved it in the 1990s. What Fable 5.2 did, according to Anthropic, is formalise an existing proof in Lean so a computer can check every step. That is still a major engineering feat, but it is verification, not a new discovery.

Is Claude Fable 5.2 open source?

No. It is a closed model available only through Anthropic’s apps, API and cloud partners.

The bottom line

Claude Fable 5.2 looks like a real step up, especially in long-running agentic work, robotics and computer use, and the price cut helps. But the most important idea in this release is not a benchmark score. It is that one model now ships in two versions with different safety wrappers, depending on who is allowed to use its most dangerous skills. That is the template we are likely to see across the industry. For students, the lesson is the same as with every major release this year: fundamentals plus careful verification beat blind trust in a leaderboard. Use these tools, but stay the person who can check their work.

Sources and further reading

• Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2
• Anthropic: Claude Fable 5.2 and Mythos 5.2 system card
• Anthropic: Expanding Project Glasswing
• TechCrunch: Anthropic releases new Mythos and Fable models amid RSI concerns
• VentureBeat: Fable 5.2 and Mythos 5.2 benchmarks and 20% price cut
• The Next Web: Anthropic launches Claude Fable 5.2 and Mythos 5.2
• DataCamp: Claude Fable 5.2 and Mythos 5.2, our take
• Plotly: Claude Fable 5.2 data analytics benchmark results
• Business Today: Anthropic launches Claude Fable 5.2 and Mythos 5.2

Written by Piyush Wairale (M.Tech, IIT Madras), founder of PiyushAI Edtech. Benchmark scores, prices and access rules change often; figures are as reported on or before 10 November 2026.