Claude Fable 5.2 and Mythos 5.2 are the same Anthropic model, released on November 9, 2026: Fable costs $8 per million input tokens and $40 for output, 20 percent less than Fable 5.1, but double Opus 5.5. Mythos has fewer restrictions and cannot be purchased.
Fable 5.2 is now available in the Claude API with the identifier claude-fable-5-2, and also on Amazon Web Services, Google Cloud, and Microsoft Azure. Here’s what changes, what a session actually costs, and when it’s worth paying double.
What is Claude Fable 5.2?
Claude Fable 5.2 is the most capable model Anthropic offers to the public, designed for demanding reasoning and long-term agentic work: the tier above Opus. According to Anthropic’s model documentation, it has a context window of 1M tokens, a maximum output of 128K tokens, and adaptive thinking always active. That means you don’t turn extended reasoning on or off: you control how much the model thinks with the effort level.
Anthropic also claims it generates output up to 30 percent faster than Fable 5.1, even without fast mode.
What is Claude Mythos and how does it differ from Fable?
Claude Mythos 5.2 is the same model as Fable 5.2, with fewer safeguards and restricted access. The difference is in the controls and who can use it:
- Fable 5.2 has stricter safeguards on cybersecurity and biology, and is open to everyone.
- Mythos 5.2 is reserved for verified organizations. Security professionals request access through the Cyber Verification Program, and life sciences researchers through the Life Sciences Verification Program. Enterprise customers can also use Mythos 5.2 through Claude Security without belonging to the program.
Both are priced the same. In practice, the exclusive Mythos figures published in the announcement, like 89.6 percent in CVE rediscovery, are numbers most readers won’t be able to verify, because they don’t have access.
How much does Claude Fable 5.2 cost?
Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, and $16/$80 in fast mode. This is the comparison that matters, according to Anthropic’s pricing page:
| Model | Input (MTok) | Output (MTok) | Cache hit (MTok) |
|---|---|---|---|
| Claude Fable 5.2 | $8 | $40 | $0.20 |
| Claude Fable 5.2 — fast mode | $16 | $80 | — |
| Claude Fable 5.1 | $10 | $50 | $0.25 |
| Claude Opus 5.5 | $4 | $20 | $0.20 |
| Claude Opus 5.5 — fast mode | $8 | $40 | — |
| Claude Sonnet 5.5 | $2 | $10 | $0.10 |
Three numbers stand out:
- 5.2 is cheaper than 5.1. If you already pay for Fable, upgrading reduces your bill by 20 percent, both on input and output.
- Fable 5.2 at normal speed costs exactly the same as Opus 5.5 in fast mode. Both cost $8/$40. At that price, you choose between a faster Opus and a more capable model.
- The cache hit price is the same as Opus 5.5: $0.20. In a long agentic session, where most of the input is reread from cache, the difference between both models concentrates on output.
An example, with figures chosen to show the effect and not measured: a session that reads 2 million tokens from cache and writes 100,000 output tokens costs about $4.40 with Fable 5.2 and $2.40 with Opus 5.5. That’s 1.8 times more, not double. The more output the task generates, the closer you get to double.
Cache writes still cost double compared to Opus ($10 per 5-minute write, versus $5), so the first turn of each session is more expensive.
Data retention: outside Enterprise Frontier Safeguards, prompts are kept for 30 days for security monitoring. Enterprise Frontier Safeguards is compatible with zero data retention, but it’s in private beta.
On subscriptions: at the time of publishing this note (November 10, 2026), Anthropic has a temporary promotion where Fable can use up to 50 percent of weekly limits on Pro plans and standard Team seats, with one to three Fable usage resets depending on your subscription tenure.
Claude Fable vs Opus 5.5: which is worth it?
In Anthropic’s own benchmarks, Fable 5.2 outperforms Opus 5.5, with a clear advantage in agentic and long-duration tasks. All figures are self-reported by Anthropic and still lack independent replication:
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| FrontierCode 2.0 | 49.1% (effort Xhigh) | 32.4% | 45.5% |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% |
| Terminal-Bench-Science 0.1 | 74.6% | 52.6% | 58.7% |
| AutomationBench | 51.4% | 31.4% | 42.5% |
| Remote Labor Index | 41.3% (effort High) | 17.9% | 21.3% |
| OSWorld 2.1 | 89.3% | 83.5% | 87.2% |
Notice that the advantage over Opus 5.5 in FrontierCode 2.0 is less than four points, and that result was measured with Xhigh effort, which consumes more tokens. The big jumps are in long-term automation tasks.
Anthropic itself adds a warning worth keeping in mind: both 5.2 models expressed suspicions of being evaluated more frequently than any recent model, and Anthropic acknowledges this limits how well their evaluations predict real-world behavior.
Fewer blocks. Anthropic reports that in early tests on claude.ai, Claude Code, and Cowork, Fable 5.2 triggered about 55 percent fewer blocks from its safety classifiers than 5.1 (around 60 percent fewer on cybersecurity and 40 percent fewer on biology), and maintains that its safeguards remain equally robust. That combination is how Anthropic presents the figure. When a query is blocked, it falls back to Opus 4.8 on cybersecurity and Opus 5 on biology: you get a response from another model instead of nothing.
If you want to see how costs are split across the models you already use in Claude Code, we compare it turn by turn in Claude Sonnet 5.5 vs Opus 5.5: in Claude Code, choosing a model is now a cost decision.
When is it worth paying double?
Fable 5.2 pays for itself when a failed run costs you more than the tokens:
- Long-term agentic work, like multi-step refactorings, migrations, or automations that run unsupervised. That’s where Fable’s advantage is biggest and where a wrong turn forces you to redo the work.
- Tasks where Opus 5.5 already struggles. If you’re reformulating the Opus prompt three times, a single run with Fable is cheaper.
Opus 5.5 remains the best default choice for routine code, quick edits, and any task with high output whose results you can review yourself. The announcement also mentions Advisor and Orchestrator patterns, where Fable 5.2 advises smaller models or delegates work to them: a way to pay Fable price only for planning.
How to use Claude Fable 5.2 in the API?
Use the model identifier claude-fable-5-2. Effort is defined with output_config.effort, which accepts low, medium, high, xhigh, and max. In the API, Fable 5.2 uses high by default, and Anthropic’s recommendation is to start there and move to xhigh only on the hardest long work, where you’ve measured an improvement:
response = client.messages.create(
model="claude-fable-5-2",
max_tokens=16000,
messages=[{"role": "user", "content": "..."}],
output_config={"effort": "high"},
)
Since reasoning is always active, define a wide max_tokens with high or higher levels: it’s a strict limit for reasoning and response text combined.
How to use Claude Fable in Claude Code?
The announcement says Fable 5.2 is available in Claude Code, with Medium effort by default and with fast mode. However, at the time of publishing this note (November 10, 2026), Claude Code’s documentation hadn’t been updated: its model configuration page still indicated that the fable alias points to Fable 5.1, and its fast mode page only listed Opus 5.5, Opus 5, and Opus 4.8 as compatible. Check the /model selector in your own version to confirm which Fable you get.
Here’s what the documentation does say about Fable models in Claude Code, and applies today:
- Fable is never the default model. You have to select it explicitly.
- It may be charged to usage credits instead of your plan’s limits, depending on the plan and seat type. In that case, the
/modelselector shows “Requires usage credits”, and interactive sessions ask your consent first. - Non-interactive runs don’t ask.
claude -pand applications with the Agent SDK charge usage credits without asking. If you use Claude Code in CI, check this before switching models. - Organization restrictions or zero data retention may leave Fable unavailable; if so, that row in the selector appears grayed out.
And on fast mode in general: on subscription plans it charges only usage credits, even if you have available usage in your plan, and the first time you activate it in a conversation you pay fast mode price without cache for all accumulated context. Turn it on at the start of the session, not midway through. Run /status to see whether your session is billed to a subscription or a Console API key.
What about you? On what specific task would you pay double for Fable 5.2 instead of sticking with Opus 5.5?
- Opus 5.5
- Sonnet 5.5
- Fable
- I prefer another option (tell us which)
Comment below or, if you got this article by email, reply directly to the email: your response gets published here.
Related: Claude Haiku 5.5 vs Sonnet 5.5 vs Opus 5.5: what each turn actually costs in Claude Code