)
Last updated: November 12, 2026
What is Claude Fable 5.2? Claude Fable 5.2 is Anthropic's top generally available model, released on November 9, 2026. It lists at $8 per million input tokens and $40 per million output tokens, 20% below Fable 5.1, and it is the same underlying model as Claude Mythos 5.2 with cyber and biology safeguards turned on. It tops one independent index, and it brings a prompt injection caveat that enterprises should read before deploying agents.
TL;DR: Key Takeaways
- Price falls 20%: Fable 5.2 lists at $8 input and $40 output per million tokens, against $10 and $50 for Fable 5.1, with a 1M-token context window and 128K max output (Claude Platform docs).
- Artificial Analysis ranks it first: 65 on its Intelligence Index at max effort, against 59 for Opus 5.5 and 57 for GPT-6 Astra, but at $26.54 and 47.8 minutes per task (Artificial Analysis).
- Anthropic's table shows big gains over Fable 5.1, such as 49.1% against 32.4% on FrontierCode 2.0, and one loss: Gemini 4 Carbon leads AutomationBench, 53.9% to 51.4% (Anthropic).
- Anthropic reports about 55% fewer blocked requests than Fable 5.1, and blocked cyber requests fall back to Opus 4.8 (Anthropic).
- The system card shows fallbacks matter for agents: indirect prompt injection succeeded 4.09% of the time against Fable 5.2 and 0.96% against Opus 5.5, as reported by Notebookcheck.
- Covered Model rules apply: 30-day minimum retention by default, and zero data retention only as a limited transition for eligible customers (Claude Help Center).
Claude Fable 5.2 at a Glance
Claude Fable 5.2 is a reasoning model with a 1M-token context window, a 128K output limit, adaptive thinking that is always on and a list price of $8 and $40 per million tokens. It sits above Opus 5.5 in Anthropic's lineup and costs twice as much per token. Anthropic still recommends Opus 5.5 as the daily driver.
| Spec | Claude Fable 5.2 | Claude Fable 5.1 | Claude Opus 5.5 |
|---|---|---|---|
| API model ID | claude-fable-5-2 | claude-fable-5-1 | claude-opus-5-5 |
| Context window / max output | 1M / 128K tokens | 1M / 128K tokens | 1M / 128K tokens |
| Input / output price per 1M tokens | $8 / $40 | $10 / $50 | $4 / $20 |
Source: Claude Platform docs (2026). Fable 5.2 launched on the Claude apps, the Claude Platform, Amazon Bedrock, Google Cloud and Microsoft Foundry. For the previous generation, see our Fable 5.1 vs GPT-6 Astra benchmark.
How we researched
On November 12, 2026 we read Anthropic's launch post, the Claude Platform model docs, the Claude Help Center Covered Models page, the Artificial Analysis leaderboard and model page, the Terminal-Bench 4.0 leaderboard and OpenRouter's listing. Scores are labeled vendor-reported or independent, and we ran no tests of our own. Prompt injection figures come from Notebookcheck's reading of the system card.
What Changed From Fable 5.1
Fable 5.2 changes five things from Fable 5.1: a 20% lower price, up to 30% faster output, an optional fast mode, about 55% fewer classifier blocks and stronger anti-distillation protections. Anthropic also describes a step up in coding, reasoning and vision, and says the model communicates more clearly than earlier Fable releases.
| Change | What Anthropic reports |
|---|---|
| Price | $8 input and $40 output per million tokens, down from $10 and $50 |
| Speed | Up to 30% faster than Fable 5.1; fast mode up to 2.5x faster at $16 and $80 |
| Safeguard false positives | About 55% fewer blocks in early testing: about 60% fewer for cyber, 40% for biology |
| Token efficiency | Fewer tokens than Fable 5.1 at medium effort; High effort roughly level with Opus 5.5 at Max on FrontierCode 2.0, at about half the cost per task |
| Anti-distillation | 90% to 99% less data harvested by distillers; preserved thinking now applies to all accounts |
Fast mode runs on Claude Code, the Claude Platform, Claude Models on Google Cloud and Microsoft Foundry. According to OpenRouter (2026), the Anthropic fast endpoint serves about 216 tokens per second, against 78 for the standard endpoint, and the page shows it at the doubled price. All of these gains are Anthropic's own early-testing claims.
)
Claude Fable 5.2 Benchmarks: What Anthropic Reports
Anthropic's launch table puts Fable 5.2 ahead of Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini 4 Carbon on most listed tests, with Gemini 4 Carbon ahead on AutomationBench. These are vendor-reported scores at each model's highest available effort with safeguards on, so treat them as ceilings and verify them on your own tasks.
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
|---|---|---|---|---|---|
| FrontierCode 2.0 | 49.1% | 32.4% | 45.5% | 42.8% | 41.1% |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
| Terminal-Bench-Science 0.1 | 74.6% | 52.6% | 58.7% | 68.1% | 64.8% |
| OSWorld 2.1 (offline set, partial) | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| Humanity's Last Exam (with tools) | 74.2% | 65.6% | 67.7% | 57.2% | not reported |
| Chartography (no tools) | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
Source: Anthropic launch post (2026). FrontierCode 2.0 shows Fable at Xhigh effort and Gemini 4 Carbon at Medium effort.
Fable 5.2 wins six of the seven rows, and the gap over Opus 5.5 is largest on Chartography and Terminal-Bench-Science. The exception matters for business automation: on AutomationBench, Gemini 4 Carbon scores 53.9% against 51.4%. According to Anthropic (2026), Fable 5.2 at High effort scores 45.3% on FrontierCode 2.0, about level with Opus 5.5's best 45.5%, at roughly half the cost per task.
Independent Results and Cost per Task
Artificial Analysis ranks Claude Fable 5.2 first on its Intelligence Index, but the lead is expensive. At max effort the model scores 65 and costs $26.54 per task, against $10.04 for Opus 5.5 at 59. At high effort, the API default, it scores 62 for $9.08.
| Model (effort) | Index score | Cost per task | Time per task |
|---|---|---|---|
| Claude Fable 5.2 (max) | 65 | $26.54 | 47.8 min |
| Claude Fable 5.2 (high) | 62 | $9.08 | 16.6 min |
| Gemini 4 Carbon (high) | 60 | $7.76 | 25.5 min |
| Claude Opus 5.5 (max) | 59 | $10.04 | 25.0 min |
| GPT-6 Astra (max) | 57 | $6.24 | 24.3 min |
| GPT-6.1 Sol (max) | 55 | $1.27 | 21.7 min |
Source: Artificial Analysis leaderboard, Intelligence Index v5.0.1, November 12, 2026.
According to Artificial Analysis (2026), Fable 5.2 at max effort is expensive, slow per task and very verbose: it takes 245.58 seconds to its first answer token and produced 275M output tokens on the index, against a median of 72.8M. The index version also changed, so scores here are not comparable with the v4.3.2 figures in our Fable 5.1 article.
One gap remains. On November 12 the official Terminal-Bench 4.0 board lists no Fable 5.2 run. It shows Fable 5.1 at 57.9%, below Opus 5.5 at 64.8% and Sonnet 5.5 at 61.8%, all in Claude Code at max effort (Snorkel AI). Wait for that entry before treating Fable 5.2 as the top terminal agent.
Claude Mythos 5.2: Who Can Use the Same Model
Claude Mythos 5.2 is the same underlying model as Fable 5.2 without the cyber and biology safeguards. Anthropic offers it to vetted organizations through its Cyber Verification Program and Life Sciences Verification Program. Claude Enterprise customers can also reach it indirectly through Claude Security, whose model mix now includes Mythos 5.2.
Anthropic reports that Mythos 5.2 finds 89.6% of the advisories on its internal CVE Rediscovery test, which covers 1,829 CVEs, against 78.3% for Mythos 5.1 and 81.3% for Opus 5.5 without safeguards. On CyScenarioBench it scores 76.3% against 61% for GPT-6 Cyber, though each lab uses unique tasks, and Anthropic's own footnote says raw vulnerability identification is roughly on par with Opus 5.5. Mythos fast mode is limited to verification program members.
Security and Governance: Running Claude Fable 5.2 in Production
The main enterprise risk with Claude Fable 5.2 is the fallback path, not the model alone. When a classifier flags a request, a weaker model answers, and agents that read untrusted content can trigger those flags. According to OWASP (2026), Prompt Injection is still LLM01 and Excessive Agency is LLM03 (OWASP LLM Top 10 2026).
Fallbacks and prompt injection
Fable 5.2 blocks high-risk cyber requests, including penetration testing and exploit generation, and most biology research. A blocked request is answered by Opus 4.8 for cyber topics or Opus 5 for biology. According to Notebookcheck (2026), citing Anthropic's help page, the checks cover everything the model reads, including memory, connector content, search results and files, so text you never typed can trigger a fallback.
| Prompt injection result | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| Gray Swan indirect attack success, 15 attempts | 4.09% | 1.02% | 0.96% |
| Attempts answered by the Opus 4.8 fallback | 53% | 23% | n/a |
| Adaptive attacker, 200 attempts per coding scenario, with Anthropic's product probes | 28.16% | not reported | 11.13% |
Source: Notebookcheck (2026), citing the Claude Fable 5.2 system card. We could not extract the PDF text, so confirm figures in the system card.
Attacks succeeded 2.05% of the time when Fable 5.2 answered itself and 5.17% when Opus 4.8 did, so the weaker fallback explains most of the gap. Without safeguards, 89.82% of adaptive attempts succeeded against Fable 5.2. Claude apps and Claude Code offer a "Switch models when a message is flagged" setting, which pauses flagged requests instead. Anthropic's own sandbox test is more reassuring: according to Anthropic, Mythos 5.2 used credentials placed outside its sandbox in 3.1% of over 1,800 scenarios, against 3.3% for Opus 5.5 and 21% for Mythos 5.1.
Data retention
According to the Claude Help Center (2026), Fable 5.2 is a Covered Model, so prompts and completions are kept at least 30 days and zero data retention is not available by default. Eligible customers can use ZDR for a limited time while Enterprise Frontier Safeguards, which monitors data stored in the customer's own infrastructure, rolls out. Anthropic names Bridgewater, Salesforce and Stripe as private beta customers.
Controls that hold regardless of model
- Decide the fallback setting per workload, and turn it off for agents that read email, web pages or third-party repositories.
- Log which model answered each request, so a fallback is visible in audit trails.
- Give agents least-privilege tools and short-lived credentials, because ASI02 (Tool Misuse) and ASI03 (Identity and Privilege Abuse) sit on OWASP's agentic list (OWASP Top 10 for Agentic Applications).
- Vet MCP servers and connectors before use, following our MCP security best practices.
- Cap spend per key, since a max-effort task can cost $26.54.
NeuralTrust puts policy outside the model. The Agent Gateway (TrustGate) routes agent traffic to models, MCP servers and APIs through one policy point with identity-based access and budgets, and ships with 200+ pre-built MCP servers. Agent Runtime Security (TrustGuard) inspects prompts, tool calls and responses to block injection and secret leakage before they act. For the coding-agent side, see our Claude Code security guide and the OWASP agentic guide.
Which Model Should You Choose?
Pick by workload and by what the agent reads. Opus 5.5 remains Anthropic's recommended starting point, and Fable 5.2 earns its price on long, ambiguous work. Gemini 4 Carbon and GPT-6.1 Sol cost far less per task. Teams with strict data rules should check Covered Model retention first.
| Your situation | Better fit | Why |
|---|---|---|
| Daily coding and knowledge work | Claude Opus 5.5 | Anthropic's recommended default, $4 and $20 per million tokens |
| Hardest long-horizon agentic work | Claude Fable 5.2 at High | Index 62 at $9.08 per task, against 59 at $10.04 for Opus 5.5 at max |
| Best accuracy regardless of cost | Claude Fable 5.2 at Max | Index 65, first on Artificial Analysis |
| Agents reading untrusted email, web or repos | Opus 5.5, or Fable with fallback off | Lower injection rate than Fable 5.2 with fallbacks |
| Business workflow automation on a budget | Gemini 4 Carbon | AutomationBench lead of 53.9%, $7.76 per Index task |
| Lowest cost per task | GPT-6.1 Sol | $1.27 per Index task at a score of 55 |
| Vulnerability research | Claude Mythos 5.2 | Cyber Verification Program access |
Conclusion
Claude Fable 5.2 is a cheaper, faster Fable that leads the independent index and Anthropic's own table on most tests. It also costs more per task than Opus 5.5, trails Gemini 4 Carbon on AutomationBench and has no Terminal-Bench 4.0 entry yet. Use it for long, ambiguous work, keep Opus 5.5 as the default, and treat fallbacks and retention as part of your threat model.
Secure Claude Fable 5.2 Agents in Production with NeuralTrust
Route Claude agent traffic through one policy point with identity-based access, budgets and real-time inspection of prompts and tool calls.
)
Related Comparisons
- Claude Fable 5.1 vs. GPT-6 Astra: Benchmark Comparison
- Claude Opus 5.5 Enterprise Security
- Claude Sonnet 5.5: Benchmarks, Pricing and Security
- Gemini 4 Carbon: Benchmarks, Pricing and Security
- Gemini 4 Argon: Benchmarks, Pricing and Security
FAQs about Claude Fable 5.2
1. When was Claude Fable 5.2 released?
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. Fable 5.2 is available on the Claude apps, the Claude Platform, Amazon Bedrock, Google Cloud and Microsoft Foundry, and the API model ID is claude-fable-5-2. Anthropic amended its launch post on November 11, so re-check the post for later edits.
2. How much does Claude Fable 5.2 cost?
Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. Fast mode costs $16 and $80. Artificial Analysis measured $26.54 per task at max effort and $9.08 at high effort, so real cost depends heavily on the effort setting.
3. Is Claude Fable 5.2 better than Opus 5.5?
On most benchmarks Anthropic reports, yes: 49.1% against 45.5% on FrontierCode 2.0 and 69.9% against 62.3% on FrontierSWE v2. It also scores 65 against 59 on Artificial Analysis. But Opus 5.5 costs half as much per token, took 25.0 minutes per task at max effort against 47.8, and is Anthropic's recommended daily driver.
4. What is the difference between Fable 5.2 and Mythos 5.2?
They are the same underlying model. Fable 5.2 has strong cybersecurity and biology safeguards and is available to everyone. Mythos 5.2 lacks those safeguards and is limited to vetted organizations through Anthropic's Cyber Verification Program and Life Sciences Verification Program.
5. Is Claude Fable 5.2 included in Claude Pro?
Anthropic says Pro plans and standard Team seats that include Fable models get Fable for up to 50% of weekly usage limits, plus one to three usage resets as a launch promotion. According to Notebookcheck (2026), the Pro offer runs until November 16. Check your plan page, because offers change.
6. What is fast mode in Claude Fable 5.2?
Fast mode is an optional setting that speeds up token generation by up to 2.5 times for $16 per million input tokens and $80 per million output tokens. It runs on Claude Code, the Claude Platform, Claude Models on Google Cloud and Microsoft Foundry. Mythos fast mode is limited to verification program members.
7. Is Claude Fable 5.2 safe for AI agents?
It has strong safeguards, but fallbacks add risk. Notebookcheck reports that indirect prompt injection succeeded 4.09% of the time against Fable 5.2, against 0.96% for Opus 5.5, because flagged requests go to Opus 4.8. Disable the fallback for agents that read untrusted content, and enforce tool policy outside the model.
About the Author
Roger Howroyd is Head of Global SEO and AI at NeuralTrust, where he leads the company's search strategy across SEO, AEO, GEO, and LLM optimization. He specializes in AI-powered search, content strategy, and SEM. Connect on LinkedIn.
NeuralTrust is the leading platform for securing and scaling AI agents. Named a Pioneer in the Gartner Emerging Market Quadrant for AI Application Security 2026, recognized across four Gartner Hype Cycle reports in 2026, and featured in the Gartner Market Guide for Guardian Agents 2026, the Gartner Market Guide for AI Gateways 2025 and the KuppingerCole Leadership Compass for Generative AI Defense 2025. Headquartered in Barcelona with offices in London and New York. ISO 27001 certified.
Sources
Accessed on November 12, 2026.
- Anthropic, "Introducing Claude Fable 5.2 and Claude Mythos 5.2", November 9, 2026.
- Anthropic, "Claude Fable 5.2 and Claude Mythos 5.2 System Card", 2026.
- Claude Platform Docs, "Models overview", 2026.
- Claude Help Center, "Covered Models", 2026.
- Artificial Analysis, "LLM Leaderboard" and "Claude Fable 5.2", 2026.
- Snorkel AI, "Terminal-Bench 4.0 Leaderboard", 2026.
- Notebookcheck, "Claude Fable 5.2 fallbacks make AI agents more vulnerable", November 10, 2026.
- OpenRouter, "Claude Fable 5.2", 2026.
- OWASP GenAI Security Project, "Top 10 for LLM Applications 2026" and "Top 10 for Agentic Applications 2026".
)
)
)
)