Anthropic debuts Opus 5.5, matching Fable's power at lower cost and faster speeds

Anthropic unveiled Opus 5.5 this week, a model that achieves top-tier performance across coding and knowledge-intensive tasks—and outperforms the company's larger Fable model on many benchmarks. The release signals Anthropic's pivot toward a more measured approach to AI development.

Performance and capabilities#

Opus 5.5 matches or exceeds Fable's results in numerous evaluations, succeeding at informal tasks where Fable stumbled. The company describes it as "the strongest-performing model we've tested to date." The model also demonstrates capabilities comparable to Mythos in specialized domains like biology and cybersecurity, triggering the same safeguards Anthropic applies to its Fable release—restrictions on discovering software vulnerabilities and designing biological weapons.

Pricing and efficiency#

A significant advantage is the model's lower cost and faster execution. Output tokens now cost $20 per mTok, down from $25 for Opus 5, with proportional savings across other pricing metrics. The reduced computational overhead also speeds up inference, making the model more efficient to deploy at scale.

A new communication style#

Opus 5.5 refines how it interacts with users: it relies less on technical jargon and prioritizes front-loading essential information, making responses clearer and more accessible.

The broader strategy: pacing progress#

The release arrives just eight weeks after Opus 5's July launch and coincides with a strategic shift. CEO Dario Amodei recently called for deliberately slowing AI capability advances to allow safety and alignment research to catch up, arguing that "fully addressing the risks requires even more prudence." Anthropic plans to release Sonnet 5.5 and Haiku 5.5 variants in coming weeks with similar improvements.

While Opus 5.5 underwent the standard alignment testing and external evaluation by groups like METR and Frontier Design, Anthropic signaled that more rigorous safety infrastructure is being built for future releases, including enhanced security monitoring and policy frameworks. The company stressed that as AI systems become more powerful, public oversight will need to expand proportionally—and that building that capacity requires time.