For Developers/Models/Claude Fable 5.2
Closed SourceAnthropicReleased 2026-11

Claude Fable 5.2

Anthropic's top Claude tier for long-horizon agentic work, 20% cheaper than Fable 5.1 at $8/$40

Context window

1M

Input / 1M tokens

$8.00

Output / 1M tokens

$40.00

Provider

Anthropic

20% below Fable 5.1 ($10/$50). Cache reads $0.20 per 1M (0.025x input), 5-minute cache writes $10 and 1-hour cache writes $16 per 1M; Batch API $4/$20. Optional fast mode (up to 2.5x faster output) costs $16/$80 per 1M on the Claude API, Claude Code, Claude Models on Google Cloud and Microsoft Foundry (Anthropic hosted, as claude-fable-5-2-fast); it does not work with Batch. The full 1M context is billed at standard rates; US-only inference adds a 1.1x multiplier. · Data verified 2026-11-11

Claude Fable 5.2 (November 9, 2026) replaced Claude Fable 5.1 as Anthropic's Fable model, the tier above Claude Opus 5.5 built for demanding reasoning and agent sessions that run for hours. It costs $8/$40 per 1M input/output tokens, 20% less than Fable 5.1, and Anthropic says it generates output up to 30% faster. In Anthropic's system card it scores 93.0% on SWE-bench Pro and 67.6% on Terminal-Bench 4.0, ahead of Opus 5.5 (89.9% and 66.4%) and Fable 5.1 (81.2% and 55.8%). The same model ships as Claude Mythos 5.2 with fewer safeguards for verified organizations. Anthropic still recommends starting with Opus 5.5 for most coding and knowledge work and moving up to Fable 5.2 when Opus at higher effort falls short.

How to access it

Available on the Claude API (claude-fable-5-2), Amazon Bedrock (anthropic.claude-fable-5-2), Claude Platform on AWS, Claude Models on Google Cloud and Microsoft Foundry. In the Claude apps and Claude Code (version 2.1.321 or later) it is on every paid plan: included up to 50% of weekly limits on Max, Ultra and premium Team and Enterprise seats, and billed through usage credits on Pro and standard seats. It is not on the Free plan. As a launch offer, Pro plans and standard Team seats get Fable models included up to 50% of weekly limits through November 16, 2026.

Strengths

  • ✓SWE-bench Pro 93.0% vs 89.9% for Opus 5.5 and 81.2% for Fable 5.1 (Anthropic system card)
  • ✓Terminal-Bench 4.0 67.6% at max effort vs 55.8% for Fable 5.1 (Anthropic-reported)
  • ✓FrontierCode 2.0 49.1% vs 45.5% for Opus 5.5 and 32.4% for Fable 5.1
  • ✓20% lower list price than Fable 5.1, with cache reads at $0.20 per 1M
  • ✓About 55% fewer safety-classifier blocks than Fable 5.1 in Anthropic's early testing
  • ✓Optional fast mode with up to 2.5x faster output

Best for developers who...

Agent sessions that run for hours, such as large code migrationsMultistep deep research carried through to a finished document, spreadsheet or deckOrchestrating smaller Claude models, or acting as an advisor model they consultHard coding and agentic tasks where Opus 5.5 at higher effort falls short

When to choose it (and when not to)

Reach for Claude Fable 5.2 when...

  • →Long-running agentic coding or research where Opus 5.5 at higher effort still falls short
  • →You run Fable 5.1 today: 5.2 costs less ($8/$40 vs $10/$50) and scores higher on Anthropic's coding and agent benchmarks
  • →You want one strong model to orchestrate or advise cheaper Claude models

Look elsewhere if...

  • ✕Everyday coding and knowledge work: Anthropic recommends starting with Opus 5.5 at half the price ($4/$20)
  • ✕You need zero data retention: Fable 5.2 requires 30-day retention unless Anthropic authorizes otherwise
  • ✕Your harness forces tool use (tool_choice any or tool) or sets temperature, top_p or top_k, which return a 400 error
  • ✕You need Priority Tier, which Fable 5.2 does not support
  • ✕Penetration testing, exploit work or research biology: classifiers route blocked requests to Opus 4.8 or Opus 5

How to use it

  • ›The API default effort is high; step up to xhigh for the hardest long-running coding work and keep max for ceiling tests
  • ›Lines such as "answer directly" do not stop thinking at medium effort and above, so lower the effort for easy turns
  • ›On long multistep tasks, ask it to finish the whole task: turns can end early at low and medium effort
  • ›Send each assistant turn back exactly as returned; edited history invalidates preserved thinking and returns a 400

Quickstart

Python
import anthropic

client = anthropic.Anthropic()
message = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=16000,
    messages=[{"role": "user", "content": "Your prompt here"}],
)
print(next(block.text for block in message.content if block.type == "text"))

Adaptive thinking is always on, so the reply can start with a thinking block; the code prints the first text block. Omit temperature, top_p and top_k. On Amazon Bedrock the ID is anthropic.claude-fable-5-2.

API model id: claude-fable-5-2

Benchmarks

BenchmarkScoreNotes
SWE-bench Pro93.0% (max effort)Anthropic system card, average of 5 trials; Opus 5.5 89.9%, Fable 5.1 81.2%
Terminal-Bench 4.067.6% (max effort)Anthropic system card, Claude Code --bare without internet access, safeguards on, 10 trials per task; Opus 5.5 66.4% (xhigh), Fable 5.1 55.8%
DeepSWE v1.174.3% (max effort)Anthropic system card, average of 5 trials; Fable 5.1 67.4%
FrontierCode 2.049.1% (xhigh effort)Anthropic-reported; Opus 5.5 45.5%, GPT-6 Astra 42.8%, Fable 5.1 32.4%
Terminal-Bench-Science 0.174.6%Anthropic-reported; Mythos 5.2 75.3%, Opus 5.5 58.7%, Fable 5.1 52.6%
Humanity's Last Exam (with tools)74.2%Anthropic-reported; Opus 5.5 67.7%, Fable 5.1 65.6%, GPT-6 Astra 57.2%
OSWorld 2.189.3%Anthropic-reported, offline set, partial score; Opus 5.5 87.2%, Fable 5.1 83.5%

Source: Anthropic Fable 5.2 / Mythos 5.2 system card

Tools powered by Claude Fable 5.2

Compare Claude Fable 5.2

Compare Claude Fable 5.2 with any other model

Build a comparison →
All model comparisons →

Learn the concepts