Anthropic introduced its latest large language model, Opus 5.5, on Tuesday, claiming it establishes a new benchmark for coding and knowledge-based tasks. The company reports that the new iteration outperforms its larger sibling, Fable, across numerous metrics and successfully completes informal assignments where Fable previously struggled.
In addition to performance gains, Anthropic has lowered the pricing for Opus 5.5. Output tokens are now priced at $20 per million, down from the $25 charged for the prior version. Other cost metrics have seen similar reductions, and the model delivers faster inference speeds due to decreased computational requirements.
The release marks a swift follow-up to Opus 5, which launched on July 24. Anthropic anticipates that Sonnet 5.5 and Haiku 5.5 will arrive in the coming weeks, bringing comparable upgrades to those models.
Anthropic also noted behavioral adjustments in Opus 5.5, aiming for clearer communication. The model is designed to reduce technical jargon and prioritize key information at the beginning of its responses.
Safety remains a central focus of this release. Anthropic stated that Opus 5.5’s capabilities in biology and cybersecurity are comparable to those of Mythos. Consequently, it is subject to the same restrictive safeguards applied to Fable, which limit the model’s utility in areas such as discovering software exploits or engineering biological weapons.
This launch aligns with CEO Dario Amodei’s recent commitment to “pacing the frontier.” In a post earlier this month, Amodei argued that fully mitigating AI risks requires slowing the pace of capability advancements to ensure risk prevention measures can keep pace.
While Opus 5.5 underwent standard alignment testing and evaluation by external groups like METR and Frontier Design, Anthropic emphasized that more robust training and security systems are being developed for upcoming releases. The company called for greater public policy involvement in AI safety, stating it is building the necessary infrastructure to support these efforts and expects to share more details soon.
I’m skeptical about these claims until independent labs publish full benchmarks. Anthropic has a history of aggressive marketing language.
Wait, did you say it outperforms Fable? That is completely surprising. I assumed the larger model always held the edge in coding benchmarks.
Great to see safety prioritized over raw speed. Finally, someone in the industry is taking the ‘pacing the frontier’ idea seriously.