Claude Fable 5.1: Pay for the Long Jobs, Not Every Prompt

If you operate an agent that loops across many turns — researching, editing code, calling tools, retrying — most of your bill is the same prompt prefix repeated. Claude Fable 5.1 is Anthropic’s generally available model for that workload. This is for engineers shipping agent systems on the Anthropic API who want a concrete recipe rather than a feature tour.

Who this is for

• You run an agent that makes repeated, structured tool calls.
• You already use Opus 5 or Sonnet 4.x and hit a ceiling on long-horizon reasoning or multi-step planning.
• You care about cost per completed task, not cost per request.

API model ID and the specs that matter

Use model ID claude-fable-5-1. The specs to anchor on:

Spec	Value
Context window	1M tokens
Max output	128K tokens
Inputs	Text and images, including vision work on documents such as PDFs, charts, and tables
Output	Text
Reasoning	Adaptive thinking (long-horizon)
Pricing (input)	$10 / MTok
Pricing (output)	$50 / MTok
Prompt-cache write (5-min)	$12.50 / MTok
Prompt-cache write (1-hour)	$20 / MTok
Prompt-cache read	$0.25 / MTok
Batch discount	50% off input and output

Numbers above are Anthropic-reported. Treat them as the vendor’s published list price; confirm against your invoice and any enterprise discount before committing to capacity planning.

How to choose and when to use it

Anthropic’s own guidance is to start with Opus 5 for most workloads and reach for Fable 5.1 for demanding reasoning and long-horizon agent work when high-effort Opus still misses your evals. In practice that means:

• Default to a cheaper Claude tier for short chats, classification, and small edits.
• Route to Fable 5.1 when (a) the task runs across many steps or tool calls, (b) you need a stable plan across long context, or (c) you want adaptive thinking to allocate reasoning budget where it matters.
• Keep Fable off the hot path for single-shot prompts where a smaller model suffices.

Fable 5.1 is on Pro, Max, Team, and Enterprise plans, plus the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Anthropic also ships Claude Mythos 5.1, the same underlying model with different safeguards. Mythos is available only to approved customers in Project Glasswing, not through Fable’s normal distribution. This guide covers Fable 5.1 only.

Anthropic estimates Fable 5.1 costs about 25% less than Fable 5 for typical token-billed workloads and up to about 45% less for highly agentic work, chiefly because cache reads are cheaper. Those are vendor estimates versus Fable 5, not a promise of savings versus another Claude model. Validate them against your own traces.

Implementation recipe

The two things that move the needle on a Fable agent are prompt caching and append-only thinking history. The Python sketch below uses the official anthropic SDK and pins the model ID only — pin a version locally that fits your environment rather than copying a version string you have not verified.

import os
import anthropic

client = anthropic.Anthropic(api_key=os.environ["ANTHROPIC_API_KEY"])

SYSTEM = [{
    "type": "text",
    "text": "You are a long-horizon coding agent. Use tools. "
            "Report progress before each non-trivial step. "
            "When you finish, return a final summary block.",
    "cache_control": {"type": "ephemeral"},  # 5-minute cache
}]

def turn(messages, tools):
    response = client.messages.create(
        model="claude-fable-5-1",
        max_tokens=16000,  # tune this to your task; thinking counts toward the limit
        system=SYSTEM,
        tools=tools,
        messages=messages,
    )
    # Thinking blocks MUST return to the API unchanged on subsequent turns.
    # Do not strip, summarize, or rewrite them. See migration warning below.
    messages.append({"role": "assistant", "content": response.content})
    return response

The explicit breakpoint caches the static system prefix for five minutes by default. Each later request can reuse that prefix at the cache-read rate instead of the full input rate. Fable 5.1 uses adaptive thinking by default; it does not need a thinking field. Anthropic says manual enabled/disabled thinking configurations return 400 on this model.

Cache-aware cost controls

Three levers that compound:

1. Cache the static prefix. System instructions, tool schemas, and project context go above an explicit cache_control breakpoint. User messages and tool outputs stay outside the cache boundary.
2. Pick the write tier that matches your session length. Five-minute caches are the default at $12.50/MTok; one-hour cache writes cost $20/MTok. Pay for the longer tier only when your sessions truly outlast the short window.
3. Use Batch for offline work. Fable 5.1 supports Batch at 50% off input and output. Use it for replanning, backfills, and eval generation — not for interactive loops.

A worked example: a warm-cache turn with 800K cached prefix tokens, 50K new tokens, and 4K output costs $0.90 versus $8.70 without caching. The first turn that writes that 800K prefix adds a $10.00 five-minute cache-write charge, so do not mistake a cold-cache turn for the steady-state cost. Cache only pays back when the reusable prefix clears Anthropic’s 512-token minimum. Re-run the math on your own traces — your mix of cached versus uncached tokens determines actual savings.

Safeguards, routing, and fallback handling

Fable 5.1 has safety classifiers. A refusal is an HTTP 200 with stop_reason: "refusal", not a server error. If you opt in to Anthropic’s Fallback API, flagged cybersecurity requests can be served by Opus 4.8 and flagged biology requests by Opus 5; Anthropic says a rerouted request is not billed at Fable prices. Two operational consequences:

• Plan for the served model in your logs. response.model identifies the model that answered after a fallback.
• Instrument stop_reason == "refusal" and fallback_message entries in usage.iterations separately. Refusals are HTTP 200 responses, so an error-rate dashboard will miss them.

Fable 5.1 requires 30-day data retention. A request from an organization or workspace without that boundary returns a 400. Zero-retention access requires Anthropic’s express authorization; Enterprise Frontier Safeguards, which Anthropic says will let enterprise customers keep data in customer-controlled cloud infrastructure, is rolling out in phases. Confirm your contract before sending sensitive material.

Migration warning: append-only preserved-thinking history

Fable 5.1 uses adaptive thinking. The contract Anthropic documents is explicit and easy to get wrong:

• Return thinking blocks to the API unchanged on every subsequent turn.
• Treat message history as append-only. Do not edit, truncate, summarize, or “clean up” previous turns.
• Ask the model to produce explicit progress updates before each non-trivial step; do not infer progress from its scratchpad.

If your existing agent strips or rewrites prior content — for example, a summarizer pass between turns, or a context-pruning routine that drops old tool outputs — it can invalidate preserved thinking. On accounts created on or after August 31, 2026, Anthropic enforces the check with a 400. Test from any account with the thinking-binding-controls-2026-08-01 beta header and prefix_mismatch_behavior: "drop_block", then log any input_transformations before production.

Failure modes to expect

• Cache misses on long pauses. The five-minute TTL expires; a session resuming after idle pays full input price on the first turn back. Either keep it warm or accept the re-prime cost.
• Silent safeguard down-routes. Flagged prompts are answered by a smaller model and billed differently. Track response.model against what you requested.
• Output truncation. 128K output is large but not infinite. Plan tool-output boundaries and final summaries so a long thinking run does not exhaust the budget mid-response.
• Thinking-block drift across forks. If you branch an agent (A/B tool choice, replan), copy the full prior history verbatim into both branches. Summarizing at the fork seam is a common break.
• Forced tool choice carried from Fable 5. Fable 5.1 rejects tool_choice: {"type": "any"} and forced named-tool calls. Migrate to automatic tool choice plus clear instructions, or use a documented structured-output path.
• Eval regression from prompts tuned for older Fable. Pricing-sensitive prompts that thrashed cache boundaries on Fable 5 will behave differently here.

Done means

• A working Fable 5.1 path that uses model ID claude-fable-5-1 with cache_control: ephemeral on the static prefix.
• Logs that record response.model on every turn, so down-routes are visible.
• An eval that explicitly checks multi-turn behavior with thinking blocks preserved end to end.
• A fallback path tested against a deliberately flagged prompt.
• Cost numbers recomputed from your own traces, not quoted from this guide.

What this guide does NOT cover

• Claude Mythos 5.1 access, pricing, or availability.
• Self-hosted or local-quantized Fable weights. There are none documented for this release.
• General prompt-engineering patterns beyond cache boundaries and adaptive thinking.
• Embeddings, vision fine-tuning, or fine-tuning of Fable 5.1.
• Legal or compliance review of your specific zero-retention boundary.
• Net-new breaking-change claims beyond what the migration guide documents.

Related guides

• /guides/abs-model-claude-opus-5/ — Opus 5 entry, the right default for non-agentic workloads.
• /guides/claude-sonnet-vs-haiku-vs-opus-when-to-pick-which/ — model-tier triage across the Claude family.
• /guides/hermes-agent-provider-anthropic-api/ — wiring Hermes Agent to the Anthropic API.

Sources

• Claude Fable 5.1 and Claude Mythos 5.1 research dossier: Johnny5/Websites/Agentic Botsitting/research/claude-fable-mythos-5-1/claude-fable-mythos-5-1-research-2026-09-01.md
• Anthropic launch announcement
• Anthropic Fable product page
• Anthropic Fable 5.1 model overview
• Anthropic prompt caching
• Anthropic preserved thinking
• Anthropic refusals and fallback