Claude Fable 5.2 vs Claude Fable 5.1: Faster, Cheaper, 15 Points Smarter

Anthropic shipped Claude Fable 5.2 on 9 November 2026, ten weeks after Fable 5.1 (1 September 2026). It is the same Fable/Mythos model that ships in two safeguard tiers, and the headline numbers are unusual in one respect: it is cheaper per token than the model it replaces, yet it jumps 15 points on the independent Artificial Analysis Intelligence Index. That combination is rare enough to be worth checking line by line.

The short version: three things changed

Fable 5.1 was an incremental step up from Fable 5 (September 2026's "better at every coding and knowledge task" release, with a cache-read cut to $0.25). Fable 5.2 is a different shape of release. Anthropic pushed the gains into science, knowledge, vision and robotics, dropped the price, and changed the economics of the effort dial by making the middle settings far stronger.

One structural change worth knowing. Fable 5.2 requires 30-day data retention for safety monitoring. Zero data retention is possible only for eligible customers running Anthropic's Enterprise Frontier Safeguards, which keeps the monitoring inside your own infrastructure. If you send regulated data through the API, that is a procurement question, not a footnote.

Price: cheaper on every line, not just the two headline numbers

The 20% cut applies across the whole rate card, including cache writes and batch. Only fast mode is new money.

Independent numbers: the Artificial Analysis effort ladder

Artificial Analysis re-ran the model, and its leaderboard is the cleanest like-for-like comparison because every effort setting is the same nine-evaluation harness (Intelligence Index v5.0.1). The result is a straight jump with a cost cliff at the top:

Read the middle rows carefully, because they are the practical story. Fable 5.2 at high effort scores 13 points above Fable 5.1 at max while costing 30% less per task. The xhigh setting buys the same 62 for $14.83, and max buys 3 more points for $26.54 — nearly triple. If you are moving off Fable 5.1 and you leave the effort dial wherever it was, you may spend more than before and get less than you could.

Which evaluations actually improved

Fable 5.2 sets the highest score Artificial Analysis has recorded on five of the nine index evaluations, and every one of the five is a science, knowledge or coding-quality test rather than an agentic-workflow test:

And here is where it does not lead, which is just as important if you are choosing a model for agent work:

The pattern: Fable 5.2 is a science and knowledge model that can code extremely well, not an agentic-workflow model that happens to know physics. Where the task rewards deep reasoning over a hard artefact — a proof, a kernel, a merged patch, a chart read correctly — it is clearly ahead. Where the task rewards long tool-driven loops with many cheap steps, the gap to Opus 5.5 and to cheaper rivals mostly disappears.

Anthropic's own benchmark table, with the deltas

These are vendor-reported numbers from the launch post, with safeguards on and each model at its highest available effort unless noted. They have not been independently reproduced except where AA ran the same benchmark (FrontierCode 2.0, Terminal-Bench-Science, AutomationBench, AA-Briefcase, OSWorld).

Two rows deserve a caveat. RoboDojo Sim is an internal simulation run — it is evidence of improvement, not of parity with purpose-built robot controllers on physical hardware, and Anthropic says as much. OSWorld 2.1 is an offline, partial evaluation set, so 89.3% is not a universal desktop-task success rate.

Speed: three claims, three different numbers

Anthropic says Fable 5.2 generates output "up to 30% faster" than Fable 5.1. Artificial Analysis measures output speed at about 98 tokens per second for Fable 5.2 (max) against 88 for Fable 5.1 (max) — roughly 11% faster at the top setting. Both can be true: the launch claim is a best case across Anthropic's own evaluations, and the AA figure is an end-to-end average including a long thinking phase.

What matters more in practice is wall-clock time per task, and there Fable 5.2 is slower at max effort: about 44 minutes per AA index task versus 22 minutes for Fable 5.1 at max. It thinks more, so it finishes later even while emitting tokens faster. This is the trade the effort dial makes explicit, and it is why the high setting is the interesting one.

Fast mode is not available on the Batch API and it stacks with cache multipliers. For an interactive coding agent where a human is waiting, 2.5× is often worth 2× the tokens. For an overnight batch job it is money burned.

The enterprise demos, and the capability they deliberately withheld

The launch post leans on four demos rather than benchmark rows, and they are worth understanding because they show the intended shape of the model. All are company-reported.

Buried in the announcement is the more consequential admission: during training, Anthropic deliberately limited progress on accelerator kernel optimisation and on improving LLM architectures, so that in those specific areas Fable 5.2 performs at about Opus 5.5's level. Anthropic states the training changes did not touch agentic coding, hardware development or interpretability research, and that it does not cap these abilities at inference time. If your workload is kernel tuning or model-architecture search, read that paragraph before you buy an upgrade.

Safety and safeguards at a glance

Those cyber and bio scores belong to Mythos 5.2, the same weights with lighter guardrails, available only to vetted organisations through Anthropic's Cyber and Life Sciences Verification Programs. Fable 5.2's classifiers will block most of that work, as they did on 5.1.

Migrating: what actually changes in your code

The migration is mostly a model-ID change, but three details are easy to get wrong: the effort default, the cache-pricing line, and the retention requirement.

import anthropic

client = anthropic.Anthropic()

resp = client.messages.create(
    model="claude-fable-5-2",
    max_tokens=64_000,
    output_config={"effort": "high"},   # low | medium | high | xhigh | max
    messages=[{"role": "user", "content": "Refactor this module to use dependency injection"}],
)
print(resp.usage)

Routing: which setting to buy

FAQ

Sources

• Anthropic — Introducing Claude Fable 5.2 and Claude Mythos 5.2: the vendor benchmark table, effort/efficiency claims, safety testing, the kernel-optimisation cap, fast mode, availability and the retention policy.
• Claude Platform docs — Fable 5.2 overview: 1M context, 128K output, adaptive thinking, default effort, June 2026 cutoff, pricing, retirement window.
• Claude Platform docs — Fable 5.1 (legacy): the previous generation's specs and price.
• Claude Platform docs — Pricing: full rate card for both models, cache multipliers, fast-mode and batch pricing.
• Artificial Analysis — Claude Fable 5.2: Intelligence Index v5.0.1 score, cost per task, output speed, token use.
• Artificial Analysis — Claude Fable 5.1: the 49 / $12.99 / 88 tok/s baseline figures.
• Artificial Analysis — launch thread: the five evaluations where Fable 5.2 set records, Terminal-Bench 4.0, GDP.pdf, AA-Briefcase Elo, AutomationBench and the 98k-vs-173k token note.
• R&D World — Fable 5.2 index score and cost per task: the full five-setting effort ladder with cost per task, read 9 November 2026.
• OfficeChai — Anthropic jumps to top of the AA Intelligence Index: per-evaluation scores and the FrontierCode high-effort note.
• Emergent — Fable 5.2 benchmarks: transcribed vendor tables, effort-level efficiency claims and the WANDR/Chartography effort notes.
• Eigent — Fable 5.2 and Mythos 5.2 explained: the two-tier safeguard split, the Anthropic guidance to start from Opus 5.5, and the 90%-of-merges context.
• The Decoder — Fable 5.2 and Mythos 5.2: the AA ranking summary and the deliberate capability-limiting decision.
• 9to5Mac — Anthropic upgrades Claude with Fable 5.2: the Opus-5.5-first guidance and the Python-to-Rust migration claim.

Vendor scores are Anthropic-reported, taken at each model's highest available effort with safeguards on unless labelled otherwise; Gemini 4 Carbon's FrontierCode row is at medium effort, Fable 5.2's Remote Labor Index row is at high effort and its FrontierCode row at xhigh. AA scores are independent runs in Intelligence Index v5.0.1. Prices are USD per million tokens, November 2026.