Anthropic's Opus 5.5 Shakes Up AI Market with Cheaper Prices and Fable-Level Performance

Anthropic has officially launched its newest artificial intelligence model, Opus 5.5, marking a significant advancement in the fields of coding and knowledge work. According to the company, this release sets a new state-of-the-art benchmark, positioning Opus 5.5 as the most capable and expensive tier within Anthropic’s Claude lineup, which also includes the mid-tier Sonnet and the fast, cost-effective Haiku models.

A notable achievement of Opus 5.5 is its superior performance over the larger Fable model in numerous benchmarks, successfully completing several informal tasks that Fable previously failed. Beyond its enhanced capabilities, the new model also boasts significant cost reductions. Output tokens for Opus 5.5 are priced at $20 per million tokens, a decrease from the $25 per million tokens charged for its predecessor. Similar price drops are observed across other metrics, complemented by a faster operational speed that reflects a reduced computational requirement for serving the model.

The latest iteration of Opus introduces substantial changes to its communication style, making Opus 5.5 less prone to using jargon and more inclined to present crucial information at the beginning of its messages. This launch follows closely on the heels of the Opus 5 release in July, just two months prior. Anthropic has also announced that Sonnet 5.5 and Haiku 5.5, the subsequent tiers in the Claude series, are slated for release “in the coming weeks,” promising comparable performance enhancements.

Regarding safety and ethical considerations, Anthropic states that Opus 5.5 shares similar biology and cybersecurity capabilities with its Mythos model. Consequently, its release is subject to the same rigorous safeguards applied to the company’s Fable model. These safeguards specifically limit the extent to which the models can be utilized for discovering exploits in compiled programs or for developing identifiable biological weapons, among other restricted activities.

Opus 5.5 is the first model released since Anthropic CEO Dario Amodei publicly endorsed the strategy of “pacing the frontier.” This philosophy advocates for a deliberate deceleration of progress in AI capabilities to allow the rate of progress on AI alignment and risk prevention to catch up. Amodei elaborated on this stance, stating that “fully addressing the risks requires even more prudence,” emphasizing the need not only to invest in risk prevention but also to synchronize the advancement of capabilities with the development of adequate safety measures.

The safety training for Opus 5.5 broadly mirrors that of its predecessors, incorporating alignment testing and pre-release evaluations conducted by external organizations such as METR and Frontier Design. Anthropic has underscored that more advanced training and evaluation systems are already under development for future models, which will include enhanced security and monitoring features. The company also highlighted the growing importance of public policy in ensuring the safety of AI systems as they become more capable, acknowledging that building such capacity takes time and requires foundational infrastructure, with more details on these efforts expected to be shared soon.