Claude Fable 5 and Mythos 5 for AI Agents: Model IDs, Prices and API Changes

この記事は英語版のみです。

Short answer: Claude Fable 5 has two successors. The current one, Claude Fable 5.2 (claude-fable-5-2), is Anthropic's most capable widely released model, with a 1M-token context, 128K output tokens and a price of $8 input and $40 output per million tokens, lower than Fable 5's $10 and $50. Claude Mythos is the same model as the Fable release of the same number, offered only to organizations that pass Anthropic's verification programs. For agent code, Fable always thinks, controls depth with effort, rejects forced tool choice from 5.1 on, and can stop with a refusal that you should handle.

Key takeaways

• Use claude-fable-5-2 for new work. Fable 5 and 5.1 are still served, at higher prices.
• Mythos needs verification. Mythos 5.2 matches Fable 5.2 in capability, limits and price, for verified organizations such as members of the Cyber Verification Program.
• Opus 5.5 is the cheaper default. At $4 and $20 per million tokens, it is half the price of Fable 5.2. Use Fable where measured quality justifies the cost.
• Agent harnesses need changes. No disabled thinking, no sampling parameters, no prefill, no forced tool choice, and append-only history.

The Fable and Mythos models

From Anthropic's models overview and pricing, read on 2026-11-15. All have a 1M-token context and 128K maximum output.

Model	Model ID	Input per MTok	Output per MTok	Access
Claude Fable 5.2	claude-fable-5-2	$8	$40	Generally available
Claude Mythos 5.2	claude-mythos-5-2	$8	$40	Verified organizations
Claude Fable 5.1	claude-fable-5-1	$10	$50	Still served
Claude Mythos 5.1	claude-mythos-5-1	$10	$50	Verified organizations
Claude Fable 5	claude-fable-5	$10	$50	Still served
Claude Mythos 5	claude-mythos-5	$10	$50	Verified organizations
Claude Opus 5.5 (for comparison)	claude-opus-5-5	$4	$20	Generally available

Cache reads cost $0.20 per million tokens on Fable 5.2, $0.25 on Fable 5.1 and $1 on Fable 5. Fast mode on Fable 5.2 runs at $16 and $80 per million tokens on the Claude API.

Fable or Mythos

Anthropic describes each Mythos release as the same model as the Fable release of the same number, with the same limits and prices, offered to organizations verified through programs such as the Cyber Verification Program. If your organization does not have that access, use Fable. One difference in the older pair: Claude Mythos 5 ran no safety classifiers, so it never returns stop_reason: "refusal". Mythos 5.1 and 5.2 do.

Fable or Opus

Fable 5.2 is for the hardest reasoning and long agentic work. Anthropic recommends Claude Opus 5.5 as the default for most work. It costs half as much and has the same context and output limits. Start with Opus 5.5 at a higher effort, and move a route to Fable 5.2 when your own evaluation shows the gain.

API changes that affect agent code

These apply to Fable 5.2 and, unless noted, to the earlier Fable and Mythos models. Anthropic's migration guide has the details.

Change	What to do
Thinking is always on	Omit thinking or send {"type": "adaptive"}. {"type": "disabled"} and budget_tokens return a 400.
Effort controls depth	Set output_config.effort to low, medium, high, xhigh or max. Start at high on Fable 5.2.
No sampling parameters	Remove temperature, top_p and top_k; they return a 400.
No assistant prefill	Use structured outputs or instructions to control the format.
No forced tool choice (5.1 and 5.2)	tool_choice any or tool returns a 400. Use auto, name the tool in the prompt, and set strict: true on the tool.
Refusals	Check stop_reason before reading content. A refusal comes with stop_details and a category.
History must be append-only (5.2 on every account)	Pass thinking blocks back unchanged and do not edit earlier turns, or later thinking blocks become invalid.
Data retention	Fable needs 30-day retention. Organizations on zero data retention get a 400 unless Anthropic authorizes it.
Long turns	A hard request can run for minutes. Stream the response and set timeouts to match.

A Fable 5.2 request with a refusal fallback

On the Claude API, the server-side fallback beta re-runs a declined request on another model inside the same call (refusals and fallback):

import anthropic

client = anthropic.Anthropic()

response = client.beta.messages.create(
    model="claude-fable-5-2",
    max_tokens=16000,
    betas=["server-side-fallback-2026-07-01"],
    fallbacks=[{"model": "claude-opus-4-8"}],
    output_config={"effort": "high"},
    messages=[{"role": "user", "content": "Plan the migration of our billing service to the new queue."}],
)

if response.stop_reason == "refusal":
    print("Declined:", response.stop_details.category)
else:
    print(response.model, response.content[-1].text)

response.model tells you which model answered. A refusal on the final response means every model in the chain declined.

Prompting Fable for agent work

From Anthropic's migration and prompting guidance for Fable 5.2:

• Set effort explicitly. The default is high, where most of the gain arrives. Gains above high are uneven and depend on the task. At high and above the model thinks before almost every reply, so use low or medium for easy turns.
• Set length in the prompt. Effort barely changes how long replies are.
• Remove lines that tell the model to think less. They do not work at medium and above. Lower the effort instead.
• Watch for early stops at low effort. At low and medium the model is more likely to stop before the task is done. Anthropic's Fable 5.2 prompting guide gives text to add for agentic work at those levels.
• Size max_tokens for thinking. It caps thinking plus the reply, so a small limit can stop the turn before any text.

Use Claude Fable on AIsa

AIsa lists claude-fable-5-2, claude-fable-5-1 and claude-fable-5 in the model catalog, with Fable 5.2 at $8 input and $40 output per million tokens, the same as Anthropic's API on 2026-11-15. AIsa does not list the Mythos models.

AIsa serves Claude through the Anthropic Messages format, so the official SDK works with a different base URL. The base URL has no /v1, because the SDK adds /v1/messages (AIsa Claude docs):

import anthropic

client = anthropic.Anthropic(
    base_url="https://api.aisa.one",
    api_key="YOUR_AISA_API_KEY",
)

response = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=16000,
    output_config={"effort": "high"},
    messages=[{"role": "user", "content": "Summarize the three biggest risks in this plan."}],
)

if response.stop_reason != "refusal":
    print(response.content[-1].text)

The same key reaches GPT, Gemini, Qwen and Kimi models and 6,000+ data APIs, so a Fable agent can search the web, read social data or check markets without another account. To run Claude Code on AIsa, see Use AIsa in Claude. To get a key from Anthropic instead, see How to get a Claude API key. Create a key in the AIsa console.

FAQ

What is the Claude Fable 5 model ID?

claude-fable-5. The current release is claude-fable-5-2, and claude-fable-5-1 is still served.

What is the difference between Claude Fable and Claude Mythos?

They are the same model at each version number, with the same limits and prices. Mythos is offered only to organizations verified through Anthropic's programs, such as the Cyber Verification Program.

How much does Claude Fable 5.2 cost?

$8 per million input tokens and $40 per million output tokens, with cache reads at $0.20. Fable 5 and 5.1 cost $10 and $50.

Can I turn off thinking on Claude Fable?

No. Thinking is always on, and {"type": "disabled"} returns a 400. Lower output_config.effort to reduce thinking.

Can I use Claude Fable through AIsa?

Yes. AIsa lists Fable 5, 5.1 and 5.2 on its Anthropic Messages endpoint at https://api.aisa.one. It does not list the Mythos models.

Sources

• Models overview, pricing, migration guide, Fable 5.2 prompting guide and refusals and fallback, Claude Platform Docs, read 2026-11-15
• Claude models on AIsa and the AIsa model catalog, read 2026-11-15