Claude Fable 5.2: One Model, Two Doors — and the 20% Price Cut That Isn't About Price

The short version

Anthropic shipped Claude Fable 5.2 and Claude Mythos 5.2 on November 9. On the surface it is a routine upgrade: faster, cheaper, higher benchmark scores. The interesting part is the structure — one underlying model, two different doors into it — and the fact that Anthropic is now pricing by the task, not just by the token.

Three numbers to hold on to:

• Price: $8/$40 per million tokens, down 20% from 5.1's $10/$50
• Speed: output up to 30% faster than 5.1, plus a fast mode at up to 2.5x speed for double the price ($16/$80)
• Benchmarks: 49.1% on FrontierCode 2.0, ahead of Opus 5.5 (45.5%), GPT-6 Astra (42.8%), and Gemini 4 Carbon (41.1%)

Here is the counterintuitive part: Anthropic still recommends Opus 5.5 as the default for most coding and knowledge work. Why isn't the strongest model the default? The answer is in the two calculations below.

One model, two doors

Fable 5.2 and Mythos 5.2 are the same underlying model. The only difference is the guardrails.

• Fable 5.2: ships with cybersecurity and biosecurity guardrails, publicly available
• Mythos 5.2: loosens some of those guardrails, only for organizations that pass Cyber Verification or the Life Sciences Verification Program

Anthropic says Fable 5.2's guardrails falsely block about 55% fewer requests than 5.1 (about 60% fewer in cybersecurity, about 40% in biology), at the same strength. So the same capability comes in two versions: a filtered one for everyone, and a fuller one for verified institutions. Model capability is now explicitly layered by trust.

For agent developers, that means the same prompts and toolchains may behave differently depending on which door you walk through. If you build agents in security or biology, check which door you are on first.

Cheaper per token, but the real bill is per task

Fable 5.2's per-token price is twice Opus 5.5's. Looked at that way, it is the expensive one. Anthropic's argument is a different ledger:

On FrontierCode High effort, Fable 5.2 roughly matches Opus 5.5's best coding score at about half the cost per task. The logic is simple: fewer steps, a higher chance of getting it right the first time, so total tokens go down.

But that conclusion has a precondition: it only holds for long, hard tasks. For short tasks and routine coding, Opus 5.5's per-token advantage is still there — which is exactly why Anthropic recommends it as the default. Before picking a model, look at your own task mix: the higher the share of long tasks, the more Fable 5.2 pays off.

# Model IDs (from Anthropic's official docs)
claude-fable-5-2    # standard mode, $8/$40
claude-mythos-5-2   # restricted mode, same price, verification required
# fast mode: up to 2.5x speed at $16/$80; Mythos fast mode only for verified organizations

How to read the benchmark table

Anthropic's self-reported scorecard (all vendor numbers, not independently verified):

Two independent data points for cross-checking: Artificial Analysis gives Fable 5.2 an Intelligence Index of 65, the highest score to date (Carbon 60, Opus 5.5 59). Mythos 5.2 scores 76.3% on CyScenarioBench and 89.6% on CVE Rediscovery.

A reading note: Fable leads in 13 of 15 vendor comparisons, but Carbon wins one business-automation test (53.9% vs 51.4%), and Opus 5.5 edges ahead in Artificial Analysis's human-to-human Briefcase ranking (1800 vs 1785). "Leading everywhere" is not the fact — "leading in most places" is. Also, Fable 5.2's FrontierCode run was at Xhigh effort while Carbon ran at Medium effort — cross-tier comparisons need a discount.

Four takeaways for agent developers

1. Route by task length first. Long, hard tasks go to Fable 5.2; routine work stays on Opus 5.5. Run your own task distribution through the math before trusting the per-token price.
2. Fast mode is a rescue, not a default. 2.5x speed at double the price fits deadline-driven long tasks, not everyday use.
3. Security and biology work: check your access first. Fable's guardrails will block some requests; false blocks are down 55%, but if your use case sits on the guardrail boundary, go through verification.
4. Benchmarks at max effort will mislead you. Fable 5.2's top tier costs about $26.50 per task, roughly 2.6x Opus 5.5. Test on your own tasks before going to production — don't copy launch-day numbers.

The honest limitations

• Every benchmark above is self-reported by Anthropic; the methodology is in the system card, and nobody had independently reproduced it at launch.
• 1M-token context, 128K max output, knowledge cutoff June 2026.
• Fable 5.1 is now a legacy model; the Pro tier includes 1 Fable reset in the launch week (per Anthropic's help center — details vary by tier).

Would you move your long tasks to Fable 5.2, or keep Opus 5.5 as the default? Tell me how you route.

Facts checked against Anthropic's November 9 announcement and cross-checked reporting (VentureBeat, The Decoder, Artificial Analysis) on 2026-11-10. Benchmark figures are vendor-reported unless stated otherwise.