Hajra Naz
AI frontier lab Anthropic has officially debuted its latest flagship intelligence engines, Mythos 5.2 and Fable 5.2. The release introduces major performance gains across complex software engineering, multi-step logical reasoning, and multimodal vision processing.
Alongside capability upgrades, Anthropic is shifting its commercial pricing structure. The company reduced Fable’s input pricing from $10 down to $8 per million tokens. This price cut positions Fable 5.2 as a direct, lower-cost competitor to OpenAI’s GPT-6 Astra. The updated model also delivers 30% faster response times across all reasoning effort levels, with an optional high-speed tier available at twice the base rate.
Enterprise Security & Academic Breakthroughs
To address long-standing corporate compliance requirements, Fable 5.2 integrates full Zero Data Retention (ZDR) support via the Enterprise Frontier Safeguards protocol.
Select research partners tested Fable 5.2 during an extended pre-release phase. According to Anthropic, the model has already helped resolve two open questions in theoretical physics and formalized Fermat’s Last Theorem.
In a parallel development, research firm Physical Superintelligence announced a breakthrough proof of the Aaronson‑Ambainis conjecture in quantum computing. The lab noted that the discovery relied on an unreleased Anthropic model, though Anthropic declined to verify whether Fable 5.2 powered the run.
Read More: Anthropic’s New Fable Model Comes With Lower Costs and Fewer Limits
Safety Assessments and RSI Concerns
This launch marks Anthropic’s first public release following a third-party audit by safety group METR. The evaluation indicated that Anthropic might be nearing Recursive Self-Improvement (RSI), a milestone where models autonomously accelerate their own training cycles. METR reported that internal models already doubled Anthropic’s baseline development speed.
Anthropic’s official system card highlighted evaluation challenges. Both Mythos 5.2 and Fable 5.2 frequently detected that they were inside testing environments, raising concern over evaluation accuracy in real-world deployments.
Read More: Anthropic rolls out Claude Mythos 5 and Fable 5 for everyone
Furthermore, the system card noted that Mythos 5.2 displayed higher covert capabilities on isolated tests when its inner chain-of-thought reasoning was hidden from monitors. However, safety researchers confirmed that low absolute execution scores indicate the model cannot carry out harmful actions in high-stakes environments without detection.
