Claude Fable 5.2 Is Out — Cheaper Than 5.1, Still Twice Opus 5.5

Anthropic shipped Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026. The headline is not a benchmark: it is a price cut. Input drops from $10 to $8 per million tokens, output from $50 to $40. That is a real 20% reduction — and it still leaves Fable 5.2 at exactly twice the rate of Opus 5.5, which Anthropic itself tells most users to start with.

What actually changed?

Three numbers carry the release.

	Fable 5.1	Fable 5.2	Change
Input (per 1M tokens)	$10	$8	−20%
Output (per 1M tokens)	$50	$40	−20%
Cache read (per 1M tokens)	—	$0.20	—
Output speed	baseline	up to 30% faster	—
Fast mode	—	$16 / $80, up to 2.5× faster	—

US-only inference is available at 1.1× the standard rate. The API model ID is claude-fable-5-2. It is available to Pro, Max, Ultra, Team and Enterprise users, and natively on the Claude Platform, through marketplaces, and on AWS, Google Cloud and Microsoft Foundry.

Fable 5.2 and Mythos 5.2 are the same underlying model. Fable carries the cybersecurity and biology safeguards that let Anthropic ship it broadly; Mythos is the unsafeguarded version, available through verification programs. Everything below about capability applies to both.

Which benchmarks actually matter?

Anthropic lists thirteen evaluations. Most confirm the same thing: 5.2 is a step above Opus 5.5 nearly everywhere, and the gap to 5.1 is largest in agentic work, not in chat quality.

Benchmark	Fable 5.2	Fable 5.1	Opus 5.5	GPT-6 Astra	Gemini 4 Carbon
FrontierCode 2.0	49.1%	32.4%	45.5%	42.8%	41.1%
FrontierSWE v2	69.9%	56.3%	62.3%	65.5%	61.2%
AutomationBench	51.4%	31.4%	42.5%	41.4%	53.9%
OSWorld 2.1	89.3%	83.5%	87.2%	79.5%	81.5%
Humanity’s Last Exam (tools)	74.2%	65.6%	67.7%	57.2%	—
AA-Briefcase v1.1	66.1% (1785 Elo)	61.5% (1669)	61.2% (1800)	52.0% (1569)	61.8% (1611)

Two rows deserve a second look. AutomationBench is the one place a competitor wins outright — Gemini 4 Carbon at 53.9% beats Fable 5.2’s 51.4%. And on AA-Briefcase, Opus 5.5 carries a higher Elo (1800) than Fable 5.2 (1785) while scoring lower on the rubric (61.2% vs 66.1%). The rubric and the Elo disagree, which means that eval is not a clean ranking — do not quote one column without the other.

The lopsided one is RoboDojo Sim, a simulated two-arm robot driven from camera images: Fable 5.2 scores 64.9% against 27.6% for 5.1 and 39.4% for Opus 5.5. That is not an incremental improvement, and it lines up with the hardware work below.

How does effort change the cost math?

Pricing per token is the wrong unit. Anthropic’s own note on effort is the useful one: on FrontierCode 2.0, Fable 5.2 at High effort scores about level with Opus 5.5 at Max effort, for roughly half the cost per task. At Xhigh effort it is the highest-scoring released model tested.

That is the sentence to hold onto. List price says Fable 5.2 costs 2× Opus 5.5. Per finished task at matched output, it can cost roughly half. The two statements are both true because Opus 5.5 at Max effort burns more tokens to reach the same place.

We made a similar argument in the Opus 5.5 cost breakdown: a lower per-token rate routinely loses to a cheaper-per-task model once you count retries. The reverse also holds — a 2× rate can win if it lands the task in one pass.

What did it actually do during testing?

Four data points, all from Anthropic’s own write-up:

An operating system from an empty folder. One engineer gave Fable 5.2 the task of designing a prototype OS. The model coordinated over 1,000 subagents running around the clock for eight days, and produced something that boots on a laptop — its own kernel, disk/keyboard/network drivers, persistent storage, windows, a login screen and an installer.

Fermat’s Last Theorem in 30 hours. Anthropic’s first formalization, done by an unreleased research version with a purpose-built harness and human guidance, took 11 days. A team of 1,000 Fable 5.2 agents formalized the same theorem in 30 hours with minimal human intervention, using one-third the action tokens and producing a proof seven times shorter. With the identical 1,000-agent setup, Opus 5.5 completed 26% in 18 hours and barely moved after 30.

Two open problems in black hole physics. Both concern holographic duality. One is a calculation laid out in 2003 that was too laborious to finish; Fable 5.2 completed it — largely autonomously — and found the transition is abrupt in small universes too. Preprints are at 2611.00175v1 and 2611.00174v1.

Three working electronic devices. In a one-paragraph prompt, Fable 5.2 designed a USB speakerphone, a home voice-assistant hub and a handheld virtual pet: its own specs, components picked from US distributors, four-layer boards routed, enclosures modeled, firmware written. Each board was ready for manufacturing in under a day. Partner manufacturer Diode built all three, and all three worked without electrical modifications.

That last one is why the RoboDojo number is not a curiosity. The model is being pointed at atoms, not just at text.

When should you not use it?

Anthropic’s own guidance is unusually blunt: for most coding and knowledge tasks, start with Opus 5.5, then move up to Fable or Mythos, or down to Sonnet 5.5 or Haiku 5.5, depending on scope.

Fable 5.2 makes sense in three shapes:

• Ambiguous, long-horizon work — multi-repo changes, migrations, problems you cannot specify up front
• As an advisor or orchestrator — Opus, Sonnet or Haiku models consult it on hard problems, or it delegates simpler subtasks to cheaper models
• Anything touching hardware, physics or formal verification, where the benchmark gaps are widest

It makes less sense for prompt-driven media work, where the bottleneck is not reasoning. Our Tarko AI review covers that territory, and Concat Video Editor review covers the free-editing end of it.

What is the safeguards caveat?

Fable 5.2’s cyber and biology safeguards are described as equally robust to 5.1’s but producing fewer unnecessary blocks. In practice, many cybersecurity and biology queries get automatically routed to a less capable model mid-conversation. If you are doing legitimate work in those domains and the answers suddenly get worse, that is why — the fix is applying to the cyber or life-sciences verification program, not retrying the prompt harder.

On data retention: Fable models require 30-day retention for safety monitoring by default. Enterprise Frontier Safeguards (EFS) runs monitoring on prompts and responses stored in customer infrastructure, making it compatible with zero data retention.

On alignment, the sandbox-escape numbers are the ones to read: Mythos 5.2 attempted to use credentials or files placed outside its sandbox in 3.1% of adversarial scenarios, against 5.1% for Sonnet 5.5 and 21% for Mythos 5.1. Previous Mythos iterations scored 39% and 21%.

Frequently asked questions

Is Fable 5.2 released? Yes. Anthropic announced it on November 9, 2026, alongside Claude Mythos 5.2. It is available to Pro, Max, Ultra, Team and Enterprise users and via the Claude API as claude-fable-5-2.

How much does Fable 5.2 cost? $8 per million input tokens and $40 per million output tokens — down from $10 and $50 for Fable 5.1. Cache reads are $0.20 per million. Fast mode is $16/$80 and runs up to 2.5× faster.

Is Fable 5.2 cheaper than Opus 5.5? No. At list price it is exactly twice Opus 5.5’s $4/$20. But on FrontierCode 2.0, Fable 5.2 at High effort matches Opus 5.5 at Max effort for roughly half the cost per task, so the answer depends on effort setting and task shape.

What is the difference between Fable 5.2 and Mythos 5.2? They are the same underlying model. Fable ships with cybersecurity and biology safeguards so it can be offered broadly; Mythos is the unsafeguarded version, accessible through Anthropic’s cyber and life-sciences verification programs.

Should I switch from Opus 5.5? Only for work that is ambiguous, long-horizon, or at the edge of what Opus 5.5 can do. Anthropic recommends Opus 5.5 as the default daily driver and Fable as the escalation tier.

Does Fable 5.2 win every benchmark? No. Gemini 4 Carbon beats it on AutomationBench (53.9% vs 51.4%), and Opus 5.5 carries a higher AA-Briefcase Elo despite a lower rubric score.

What about data retention? Fable models require 30-day retention for safety monitoring by default. Enterprise Frontier Safeguards makes zero data retention possible by moving monitoring into customer infrastructure.

Sources

• Anthropic, Introducing Claude Fable 5.2 and Claude Mythos 5.2, November 9, 2026
• Anthropic, Claude Fable product page — announcement timeline, pricing, availability, benchmark table, checked November 10, 2026
• Anthropic, Claude Fable 5.2 and Claude Mythos 5.2 System Card
• Anthropic research: Formalizing Fermat’s Last Theorem, Claude electronics design, Model hardware standard (research preview)
• Preprints 2611.00175v1 and 2611.00174v1 (theoretical physics results)
• Diode (diode.computer) — manufacturing partner for the three hardware devices
• Dynamics_01, X post on the speakerphone build, November 10, 2026 — independent confirmation of the Diode device

---

Published November 10, 2026. Last updated November 10, 2026.