Skip to content

Model Card|Claude Fable 5.2

Nov 14, 20261 min
TL;DRClaude Fable 5.2 (API ID: claude-fable-5-2) replaced Fable 5.1 on 2026-11-09: 1,000,000-token context, 128,000 max output tokens, input $8.00 / output $40.00 per 1M tokens (down 20% from the predecessor), closed-source; FrontierSWE v2 (agentic coding) hits 69.9%, 13.6 points above Fable 5.1's 56.3%, and RoboDojo Sim robot control jumps from 27.6% to 64.9%. The restricted Claude Mythos 5.2 variant finds 10.2% more vulnerabilities than Opus 5.5 on CVE Rediscovery.

Look up & save

Loading dictionary…

🌏 中文版

Model Information

FieldValue
Model IDclaude-fable-5-2
VendorAnthropic
ParametersNot disclosed
Context Window1,000,000 tokens (128,000 max output tokens)
Input pricing (USD/1M tokens)$8.00 (cache read $0.20, 5m cache write $10.00, 1h cache write $16.00)
Output pricing (USD/1M tokens)$40.00
Open sourceNo
Release date2026-11-09
Official announcementAnthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2
FamilyClaude 5.x (Fable/Mythos tier; predecessor: Fable 5.1 / Mythos 5.1)

Highlights

  • FrontierSWE v2 (agentic coding) hits 69.9%, up 13.6 points from predecessor Fable 5.1's 56.3%, and ahead of GPT-6 Astra's 65.5%
  • RoboDojo Sim (simulated robot control) jumps from the predecessor's 27.6% to 64.9% — more than double, and ahead of Anthropic's own Opus 5.5 (39.4%)
  • Output generation is 30% faster than Fable 5.1 (fast mode adds another 2.5x), while pricing drops 20%, from $10/$50 to $8/$40
  • The restricted Mythos 5.2 variant finds 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than predecessor Mythos 5.1 on CVE Rediscovery, while the safety classifiers' false-block rate drops 55%

Benchmark Results

BenchmarkFable 5.2Predecessor (Fable 5.1)Best competitor
FrontierSWE v2 (agentic coding)69.9%56.3%GPT-6 Astra: 65.5%
Terminal-Bench-Science 0.1 (agentic scientific research)74.6% (Mythos 5.2: 75.3%)52.6%GPT-6 Astra: 68.1%
OSWorld 2.1 (computer use, offline set)89.3%83.5%Opus 5.5: 87.2%
Humanity's Last Exam (with tools)74.2%65.6%Opus 5.5: 67.7%
RoboDojo Sim (robot control, internal eval)64.9%27.6%Opus 5.5: 39.4%

⚠️ All figures come from Anthropic's own hand-picked benchmark suite in its official announcement. The "best competitor" scores for GPT-6 Astra and Opus 5.5 were also run by Anthropic itself and may use a different methodology from OpenAI's or Google's own self-reported numbers — treat them as reference points only.

Comparison with Predecessor/Competitors

Compared with Fable 5.1, the biggest gains land in tasks requiring long-running autonomous execution and physical control: RoboDojo Sim jumps from 27.6% to 64.9% (more than double), Remote Labor Index rises from 17.9% to 41.3% (2.3x), and Terminal-Bench-Science 0.1 gains 22 points — this release continues Fable 5.1's trajectory of pushing agentic execution capability rather than broad knowledge.

Against same-generation competitors, Fable 5.2 leads Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon on most of Anthropic's own benchmark suite, but not every one: it loses to Gemini 4 Carbon on AutomationBench (enterprise workflow automation, 51.4% vs. 53.9%) and on Chartography (visual chart recognition, 73.4% vs. 72.5%, a narrow margin), and also trails GPT-6 Astra on Chartography (71.0%). On pricing, Fable 5.2's $8/$40 is 20% below its predecessor and still above Anthropic's own Opus 5.5 ($4/$20, exactly half), but it now opens a 20% price gap against GPT-6 Astra, which holds its $10/$50 sticker price.

Implications for Agent Development

Fable 5.2 and Mythos 5.2 continue the Fable 5.1 pattern: the same underlying weights, differing only in the safeguard tier applied at inference. Anthropic positions Fable 5.2 for "the most ambitious, open-ended technical work," while recommending most coding and knowledge tasks start with Opus 5.5 and scale up to Fable or down to Sonnet/Haiku as complexity demands. Fable 5.2 can also act as an Advisor (consulted by smaller models on harder problems) or an Orchestrator (delegating simpler tasks to smaller models).

  • If you're building multi-agent, long-running systems (research, planning, cross-tool automation): Fable 5.2 is well-suited as an orchestrator or advisor — Anthropic's own case study shows a team of 1,000 agents formalizing a proof of Fermat's Last Theorem in 30 hours, down from the 11 days a manual research effort previously took. Coordination capability is the core selling point of this release
  • If you're building hardware or robotics applications: the jump in RoboDojo Sim (27.6% → 64.9%) and Anthropic's own circuit-board design demos are worth a look — "the model directing the physical world" hasn't historically been a Claude strength, and this release marks a clear inflection point
  • Not a fit: workloads requiring zero data retention (ZDR) — Fable 5.2 requires 30-day data retention unless Anthropic grants an explicit exception. Also not a fit for distilling Claude's outputs into smaller custom models — this release ships Anthropic's strongest-ever anti-distillation protections (cutting data harvested by distillers 90-99%), and preserved thinking now applies to all accounts, closing off the path of editing prior context to extract reasoning transcripts

Today's Takeaway

One counterintuitive detail in this release: Anthropic shipped its strongest-ever anti-distillation protections while simultaneously cutting the safety classifiers' false-block rate by 55% (60% fewer false blocks in cybersecurity, 40% fewer in biology). It's easy to assume "stricter safeguards" and "fewer false positives on legitimate use" trade off against each other, but this classifier update achieved both at once — a reminder that the real bottleneck on guardrail quality usually isn't a single "loose vs. strict" dial, it's how accurate the classifier itself is. Treating those as two separate axes, rather than one, is what reveals the room still left to optimize.

References