Ten weeks after Claude Fable 5.1, Anthropic has shipped its successor. For teams that already use Fable, the short version is good news: higher scores on almost every test Anthropic published, faster output and a lower price per token. The longer version is more useful, because Anthropic itself says most people should not start with this model.
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1, and Anthropic says it generates output up to 30% faster than the 5.1 models.
This guide covers what Fable 5.2 is, what changed, how it scores in Anthropic's tests and in independent ones, what a task really costs, how the safeguards and usage limits work, and how to decide whether your work needs it. Every figure links to its source. Where a number comes only from Anthropic, we say so.
What Is Claude Fable 5.2?
Claude Fable 5.2 is Anthropic's most capable model open to all customers. Anthropic describes Fable as “best for the most ambitious, open-ended technical work”: large coding projects, long running AI agents, scientific research, data analysis and computer use.
According to Anthropic's model documentation, Fable 5.2 has a 1 million token context window, can write up to 128,000 tokens in one response, and has a knowledge cutoff of June 2026. Its thinking mode is always on and cannot be turned off. Instead you pick an effort level, from low to max, that controls how deeply it reasons.
It ships alongside Claude Mythos 5.2. Anthropic says the two are “the same underlying model”. Fable 5.2 adds safeguards for cybersecurity and biology so it can be offered to everyone, while Mythos 5.2 is only available to vetted organizations through Anthropic's Cyber and Life Sciences Verification Programs.
What Changed From Fable 5.1?
- Lower price: $8 and $40 per million input and output tokens, from $10 and $50. Cached input costs $0.20 per million tokens.
- Faster output: up to 30% faster than Fable 5.1, according to Anthropic.
- New fast mode: up to 2.5 times faster generation at $16 and $80 per million tokens, on the Claude Platform, Claude Models on Google Cloud, Microsoft Foundry and in Claude Code. Fable 5.1 had no fast mode.
- Fewer tokens on hard tasks: at medium effort, the default in the Claude apps and Claude Code, Anthropic says Fable 5.2 completes challenging tasks with fewer tokens than Fable 5.1.
- Fewer false blocks: about 55% fewer blocked requests than Fable 5.1 in Anthropic's early testing, about 60% fewer for cybersecurity and 40% fewer for biology.
How Does Fable 5.2 Score in Anthropic's Benchmarks?
These are the scores Anthropic published, in the order Fable 5.2, Fable 5.1, Opus 5.5 and GPT-6 Astra. Anthropic notes that each model is shown at its highest available effort unless stated, with safeguards on. Treat them as the vendor's own numbers.
- FrontierSWE v2, agentic coding: 69.9% vs 56.3% vs 62.3% vs 65.5%.
- FrontierCode 2.0, agentic coding at xhigh effort: 49.1% vs 32.4% vs 45.5% vs 42.8%.
- Humanity's Last Exam with tools, expert reasoning: 74.2% vs 65.6% vs 67.7% vs 57.2%.
- Remote Labor Index, real paid freelance projects: 41.3% vs 17.9% vs 21.3% vs 20.8%.
- OSWorld 2.1, using a computer like a person: 89.3% vs 83.5% vs 87.2% vs 79.5%.
- RoboDojo Sim, robotics in simulation, run internally by Anthropic: 64.9% vs 27.6% vs 39.4% vs 29.8%.
- Chartography, reading charts without tools: 73.4% vs 44.8% vs 64.4% vs 71.0%.
Fable 5.2 does not win every row. On AutomationBench, a test of business workflows, it scores 51.4%, behind Google's Gemini 4 Carbon at 53.9%. On the AA-Briefcase knowledge work test it has the highest rubric score, 66.1%, but Opus 5.5 is slightly ahead on that test's Elo rating, 1800 against 1785.
The biggest jumps are in long, messy work. The Remote Labor Index score more than doubled, and the robotics score went from 27.6% to 64.9%. That fits how Anthropic positions the model: tasks that take many steps and hours, not quick questions.
What Do Independent Tests Show?
Artificial Analysis, which runs its own nine test Intelligence Index on every major model, scores Fable 5.2 at 62 at both high and xhigh effort, ranking it second and third of 74 models on those two settings. At its highest effort setting it scores 65 and takes first place, ahead of Gemini 4 Carbon at 60 and Opus 5.5 at 59, according to The Decoder.
The independent numbers also show the trade offs:
- It is slow on hard work. At high effort, Artificial Analysis measured about 17 minutes per index task and about 28 seconds before the first word of the answer appears.
- It is not the best at everything. On Terminal-Bench 4.0, a command line coding test, it scored 60%, tied with Opus 5.5 and behind Kimi K3.1 Pro at 67%, Gemini 4 Carbon at 65% and Sonnet 5.5 at 64%, as reported by The Decoder.
- On the AA-Briefcase knowledge work test, Sonnet 5.5 at 1,813 Elo and Opus 5.5 at 1,800 rate above Fable 5.2 at 1,785.
In short, independent testing confirms Fable 5.2 is at or near the top overall, while cheaper models match or beat it on some everyday tasks.
What Does Fable 5.2 Cost Per Task?
Token prices are easier to judge as a real task. Take a request with 20,000 tokens of input, such as a long brief plus some documents, and 2,000 tokens of output.
- Fable 5.2: 20,000 × $8 / 1,000,000 + 2,000 × $40 / 1,000,000 = $0.16 + $0.08 = $0.24.
- Fable 5.1: $0.20 + $0.10 = $0.30, so the same task is 20% cheaper on 5.2.
- Opus 5.5, at $4 and $20 per million tokens: $0.08 + $0.04 = $0.12, half the price of Fable 5.2 for the same token count.
Real bills differ for two reasons. First, thinking tokens are billed as output, so a hard task at high effort can produce far more output than the visible answer. Second, a stronger model can need fewer tokens to reach the same result. Anthropic reports that Fable 5.2 at high effort scored about level with Opus 5.5 at max effort on FrontierCode 2.0 for roughly half the cost per task, and at low effort matched Opus 5.5's best Humanity's Last Exam score for less than one tenth of the cost per task. These are Anthropic's figures, but they show why cost per finished task matters more than price per token.
Other pricing details worth knowing, from Anthropic's pricing documentation:
- Batch processing, for work that can wait, is 50% off: $4 and $20 per million tokens.
- Prompt caching costs $10 per million tokens to write for 5 minutes or $16 for 1 hour, and $0.20 to read. Long agent runs repeat the conversation on every step, so caching is what keeps them affordable. It is one part of good context engineering.
- Keeping processing in the United States costs 1.1 times the normal rate.
Fable 5.2 vs Opus 5.5: Which Should You Use?
Anthropic's own advice is clear: “we recommend that users start with Claude Opus 5.5, our strongest daily driver” for most coding and knowledge tasks, then move up to Fable for the hardest work or down to Sonnet 5.5 or Haiku 5.5 for simpler work. Dianne Penn, Anthropic's head of product management for research, told CNBC that Fable 5.2 is mainly for “power users or power use cases,” such as hard problems in engineering, mathematics and robotics.
- Choose Fable 5.2 for multi day coding projects, agents that coordinate many subagents, deep research, complex financial or data analysis, and work where a wrong answer is expensive.
- Choose Opus 5.5 for writing, editing, customer support, summaries, routine coding and high volume work.
- Not sure? Run 10 to 20 of your real tasks through both models and compare quality and cost per finished task, not cost per request.
On effort, Anthropic's developer guidance says to start at high, since most of the gain is already there, use medium or low where quality holds, keep xhigh for the hardest long coding and agent work where you have measured a gain, and reserve max for testing the model's ceiling. For costs and scores at every effort level, see our full Fable 5.2 vs Opus 5.5 comparison.
How Can You Access Fable 5.2, and What Are the Limits?
- Claude apps: available on the Pro, Max, Ultra, Team and Enterprise plans, not the Free plan.
- Usage limits: on Max, Ultra and premium Team seats, Fable can use up to 50% of your weekly usage limit. On Pro and standard Team seats, Fable is paid with usage credits, except during launch week, November 9 to 16, 2026, when it is included for up to 50% of weekly limits. Eligible plans also get one to three extra Fable usage resets for launch, according to Anthropic's help center.
- Claude Code: available, with fast mode as an option.
- API: model ID claude-fable-5-2 on the Claude Platform, plus Amazon Web Services, Google Cloud and Microsoft Foundry.
- Data retention: Fable models require 30 day data retention for safety monitoring by default. Enterprise Frontier Safeguards, which runs that monitoring inside the customer's own systems, is in private beta with companies such as Bridgewater, Salesforce and Stripe. Eligible customers can already use Fable 5.2 with zero data retention.
What Are Early Customers and Researchers Reporting?
Anthropic's product page quotes companies that tested the model before release. A few are concrete:
- Shopify staff engineer Tom Pritchard: “It feels like working with another staff-level engineer.”
- Uber staff engineer Andrew Mager said it ran a 10 to 12 hour security audit using about 170 reviewer agents, and “works until the job is actually complete instead of stopping at a plausible checkpoint.”
- Snowflake called it the strongest model it has measured on data-eng-bench, an open source set of 103 data engineering tasks.
- Cognition said it “says it doesn't know things instead of hallucinating,” addressing a common failure known as AI hallucination.
Anthropic's announcement also lists research results. It says Fable 5.2 solved two open problems in theoretical physics about black holes, published as preprints, and that a team of 1,000 Fable 5.2 agents fully formalized Fermat's Last Theorem in the Lean proof language in 30 hours. That second claim comes with context: the agents started from a sketch of the proof and 47 already proven lemmas from an earlier effort, and Anthropic amended that section on November 11. With the same setup, Opus 5.5 completed 26% of the work in 18 hours. None of these results has been peer reviewed yet.
How Do the Safeguards Work?
Fable 5.2 has built in safeguards for cybersecurity and biology, and Anthropic says they “will block most cybersecurity-related tasks.” When a request is flagged, it is answered by a less capable model instead: Opus 4.8 for cybersecurity and Opus 5 for biology. In the Claude apps this happens automatically. Developers using the API have to turn on the fallback themselves, or a flagged request simply returns a refusal.
For marketing, sales and operations work, the safeguards rarely matter. Security and life science teams that hit them often can apply to Anthropic's verification programs.
On safety testing, Anthropic says the UK AI Security Institute, Frontier Design and METR evaluated the models before release. It also reports that both models voice suspicion that they are being tested more often than any of its recent models, which makes real world behavior harder to predict. And by design, Fable 5.2 performs at about the level of Opus 5.5 on tasks that could speed up AI research itself, such as optimizing the code that runs on AI chips. The release came two months after CEO Dario Amodei called for a slower pace of AI development. As Penn told CNBC: “We called for pacing, not pausing.”
What Fable 5.2 Means for Your Business
For most teams, Fable 5.2 is not a reason to switch every AI tool. It is a reason to revisit the work you decided was too hard for AI six months ago: a messy codebase, a research report that takes a week, an audit nobody has time for. Those are the jobs where the gap between Fable 5.2 and cheaper models is largest, and where its higher price per token is easiest to justify.
The rest of your AI work, such as content, support replies and summaries, will usually run just as well on Opus 5.5 or Sonnet 5.5 for less. The skill that matters is matching each task to the right model and effort level, and measuring cost per finished job.
If you want your team to choose and use models like Fable 5.2 and Opus 5.5 with confidence, from picking the right model to writing better prompts, our AI training program can help. Contact us to plan a session.
Frequently asked questions
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 was available the same day in the Claude apps, Claude Code and the API.
How much does Claude Fable 5.2 cost?
On the API, Fable 5.2 costs $8 per million input tokens and $40 per million output tokens. Cached input costs $0.20 per million tokens, batch processing is half price, and fast mode costs $16 and $80. In the Claude apps it is included in the Pro, Max, Ultra, Team and Enterprise plans.
Is Claude Fable 5.2 better than Opus 5.5?
On most benchmarks, yes, and it ranks first on Artificial Analysis's Intelligence Index at its highest effort. But Opus 5.5 costs half as much per token, matches or beats Fable 5.2 on some everyday tests, and is the model Anthropic recommends starting with for most coding and knowledge work.
What is the difference between Claude Fable 5.2 and Mythos 5.2?
Anthropic says they are the same underlying model. Fable 5.2 adds cybersecurity and biology safeguards and is available to everyone. Mythos 5.2 does not have those limits and is only available to vetted organizations through Anthropic's verification programs.
What is the context window of Claude Fable 5.2?
Fable 5.2 has a 1 million token context window, can generate up to 128,000 output tokens in one response, and has a knowledge cutoff of June 2026.
Can I use Claude Fable 5.2 on the free plan?
No. Fable 5.2 is available on paid plans only. Max, Ultra and premium Team seats include it for up to 50% of the weekly usage limit. On Pro and standard Team seats it is paid with usage credits, apart from the launch week of November 9 to 16, 2026.
Need help with this?
Growthtrait can help you put this into practice. Let's talk about your goals.
Contact us




