Claude Fable 5.2: 20% cheaper, a 2.5x fast mode, and fewer requests handed to older models

On November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, ten weeks after the 5.1 pair. The headline numbers are a 20% price cut, outputs up to 30% faster, and a new fast mode. The more useful change for anyone building on the API is narrower: the safeguards that quietly hand some requests to older models now fire about 55% less often, and API customers must now configure how that handoff works.

Anthropic’s launch page, dated 9 November 2026. Screenshot: anthropic.com

One model, two names, two doors

Anthropic is direct about the naming: “Claude Fable and Claude Mythos are the same underlying model.” Fable is that model with cybersecurity and biology classifiers on top, which is what lets Anthropic sell it to everyone. Fable 5.2 is on Pro, Max, Ultra, Team and Enterprise plans and on the API as claude-fable-5-2, through the Claude Platform, AWS, Google Cloud and Microsoft’s cloud, according to Anthropic’s Fable page.

Mythos 5.2 is the unfiltered version and stays behind an application. Organizations get it through Anthropic’s Life Sciences and Cyber Verification Programs, and Enterprise customers can use it inside Claude Security without joining the cyber program, as Fortune and MacRumors also report. If you are not in one of those programs, Fable 5.2 is the model you can call.

Pricing: per token, per task, and a fast mode at double the rate

Both models cost $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for the 5.1 generation. Cache reads are $0.20 per million, and pinning inference to US-only infrastructure adds a 1.1x multiplier. The new fast mode costs $16 and $80 and generates tokens up to 2.5 times faster. It runs in Claude Code, on the Claude Platform, on Google Cloud and on Microsoft Foundry; for Mythos it is limited to verified organizations.

Chart: GetAIBook, from Anthropic’s published prices.

The price in Anthropic’s own words. Screenshot: anthropic.com

The per-token cut matters less than Anthropic’s per-task claims, because reasoning models are billed for how much they think. On its FrontierCode 2.0 coding test, Fable 5.2 at High effort scores 45.3%, close to Opus 5.5’s best of 45.5%, “for roughly half the cost per task.” On the WANDR search benchmark it matches Opus 5.5’s top score at Medium effort for about 40% of the cost, and on Humanity’s Last Exam it does so at Low effort for under a tenth. Medium is the default in Claude apps and Claude Code, and Anthropic says Fable 5.2 finishes hard tasks there with fewer tokens than 5.1 did. Those are vendor numbers on vendor-chosen tasks, so the honest test is your own workload at each effort level.

Fortune puts the standard price 20% below OpenAI’s GPT-6 Astra and cached reads 80% below it. Those comparisons rest on Fortune’s figures for Astra; Anthropic’s pages do not list OpenAI’s prices.

Benchmarks: where it leads, and where it does not

Anthropic’s table compares Fable 5.2 with Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini 4 Carbon, each at its highest available effort and with safeguards on. Fable 5.2 leads most rows. The large gaps are in long, open-ended work: 41.3% on the Remote Labor Index against roughly 21% for Opus 5.5 and Astra, and 64.9% on the RoboDojo robotics simulation against 39.4% for Opus 5.5, though Anthropic ran that last one internally.

Chart: GetAIBook, from the scores Anthropic published.

Two rows go the other way. Google’s Gemini 4 Carbon leads AutomationBench, a business-workflow test, at 53.9% to Fable’s 51.4%, and Opus 5.5 keeps a narrow lead on the AA-Briefcase Elo rating, 1800 to 1785. On computer use (OSWorld 2.1) the margin over Opus 5.5 is only 2.1 points. Note also that the cybersecurity and protein-design rows report Mythos 5.2, not the Fable model most developers will call.

Anthropic’s full benchmark table. Screenshot: anthropic.com

The Fallback API and the 55% fewer blocks

Fable’s classifiers still block penetration testing, exploit generation, binary vulnerability work and most research biology. When they fire, the answer comes from a different model: Opus 4.8 for cybersecurity and Opus 5 for biology. Anthropic also routes a small set of AI research topics, such as accelerator kernel optimization, to Opus 5. In the apps this happens transparently. On the API, Anthropic says customers “must configure their settings with our new Fallback API” (Fable page).

That makes fallback a design decision for anyone shipping Fable in a product. If your traffic touches security tooling or life-science data, some responses will come from an older model with different strengths, and your evaluation, logging and user messaging should account for which model actually answered. The good news is the rate: Anthropic reports about 55% fewer blocks than Fable 5.1, roughly 60% fewer in cybersecurity and 40% fewer in biology, with safeguards it describes as equally robust.

The benchmark table is honest about this too: scores for Fable 5.2, Fable 5.1 and Opus 5.5 are reported with safeguards on, so the numbers above are what a filtered model delivers. For the previous release’s pricing changes, see our coverage of Fable 5.1’s cache price cut.

How Anthropic wants you to route work

Anthropic is not pitching Fable 5.2 as the default. Its advice is to start with Opus 5.5, “our strongest daily driver,” then move up to Fable for the hardest problems or down to Sonnet 5.5 and Haiku 5.5 for cheaper ones. It also suggests using Fable 5.2 as an “Advisor” or “Orchestrator” over smaller models, which fits the per-task numbers: a model that plans well at Low or Medium effort can direct cheaper workers without paying for High effort on every step. CNBC quotes Dianne Penn, Anthropic’s head of product management for research, describing Fable 5.2 as being for “power users or power use cases” (CNBC).

One constraint belongs in that routing decision. Using Fable models requires 30-day data retention for safety monitoring by default. Zero data retention is available only to eligible enterprise customers through Enterprise Frontier Safeguards, which is in private beta. Teams under strict retention rules may find Opus 5.5 is the model they can actually deploy.

What to watch

Independent results come first. Every number above is Anthropic’s, and the system card and third-party leaderboards will show whether the per-task savings hold outside its own harness. Second, the timing: CNBC reports Anthropic’s IPO could come as soon as next week, and this is its first frontier release since Dario Amodei called for a slower pace in September. Third, Anthropic says this release carries its strongest protections against distillation, cutting data harvestable through outputs by 90% to 99%. For developers who use frontier outputs to train smaller models, that is a policy change worth reading closely before building on it.