Anthropic Ships Claude Fable 5.2 With a 20% Price Cut and Keeps Its Cyber-Capable Mythos 5.2 Behind Vetting

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9 and describes them as its most capable models for coding, finance and the sciences. Most of the numbers come from Anthropic's own launch table. On FrontierCode 2.0, Fable 5.2 scored 49.1%, against 45.5% for Claude Opus 5.5, 42.8% for OpenAI's GPT-6 Astra and 41.1% for Gemini 4 Carbon. Anthropic reports 69.9% on FrontierSWE v2, 89.3% on OSWorld 2.1 and 74.2% on Humanity's Last Exam with tools, where GPT-6 Astra scored 57.2%.

Independent testing confirms the ranking and adds a cost figure. Artificial Analysis gave Fable 5.2 at maximum effort a 65 on its Intelligence Index, first among 74 models listed. The same testing measured about 232 seconds to first answer token, 103 output tokens per second and $26.54 per index task. Fable is not ahead everywhere. On the Elo-rated AA-Briefcase knowledge-work test (a ranking system similar to chess ratings), Opus 5.5 leads 1800 to 1785. On AutomationBench, Gemini 4 Carbon leads 53.9% to 51.4%.

The API price fell 20% to $8 per million input tokens and $40 per million output, from $10 and $50 for Fable 5.1. Even after the cut, Fable costs twice as much as Opus 5.5. Anthropic says output speed improved by up to 30%, and that claim has not been independently checked.

Mythos 5.2 matters more for governance. Anthropic describes it as the same underlying model as Fable 5.2 with certain cyber and biology safeguards removed. Access is limited to vetted organizations in its Life Sciences and Cyber Verification Programs. On the CVE (Common Vulnerabilities and Exposures) Rediscovery benchmark, Mythos 5.2 found 89.6% of known vulnerabilities, up from 78.3% for Mythos 5.1. Because both products share one set of weights, the access boundary depends on serving-time safeguards and customer vetting. The weights themselves do not differ.

New research addresses the monitoring layer that this kind of split relies on. A paper posted to arXiv on November 11 defines topic monitoring as labeling the topic of every token in a response while the model generates it. The authors note that activation-based monitors running at generation time have so far been trained mainly to detect harmful content. Their method trains sparse autoencoder probes on prefill activations and applies them at decode time. If such probes scale, a lab could enforce finer access rules than a binary refusal.