Anthropic's Opus 5.5 AI model arrives with price cuts and top-tier benchmark results
Anthropic's new Opus 5.5 model tops its own benchmarks in coding and knowledge tasks, surpassing the larger Fable model in several tests. The company also cut output token pricing from $25 to $20 per million, and the model runs faster. It introduces a more direct communication style, placing important details at the start of responses.
Opus 5.5 arrives just two months after Opus 5, with the company confirming that the mid-tier Sonnet 5.5 and budget-tier Haiku 5.5 will follow in the coming weeks with similar performance gains. The pricing drop on output tokens — from $25 to $20 per million — reflects reduced compute requirements for serving the model.
The release is the first since CEO Dario Amodei publicly endorsed slowing frontier AI progress to match alignment work. Safety evaluation for Opus 5.5 followed the same pattern as prior models, with external groups METR and Frontier Design conducting pre-release testing, while Anthropic says more advanced training and monitoring systems are already in development for future releases.
The combination of lower pricing, faster inference, and top benchmark results could accelerate enterprise adoption of AI for coding and knowledge work, potentially reshaping labor markets in technical fields. However, the company's own acknowledgment that safeguards limit capabilities in biology and cybersecurity suggests these systems carry real dual-use risks. As Anthropic deliberately paces advancement, the broader industry may face pressure to adopt similar restraint, though competitive dynamics could complicate that.