Anthropic Releases Claude Sonnet 5.5 with Enhanced Cyber Safeguards
Anthropic has unveiled Claude Sonnet 5.5, its latest model, with impressive performance and cost-effectiveness. It scores 70.6% on the Terminal-Bench 4.0 agentic coding test, surpassing its predecessor, Sonnet 5 (10.3%), and even the more expensive Opus 5.5 (66.4%).
Key Features:
- Faster and Cost-Efficient: Running 30% faster and costing up to 30% less per task than Sonnet 5.
- Priced Competitively: $2 per million input tokens and $10 per million output tokens.
- Frontier-Style Cyber Safeguards: Comparable to Opus 5's capabilities, Sonnet 5.5 is the first Sonnet with restrictions previously reserved for top models.
- Reasoning Extraction Blocked: Classifiers prevent the extraction of a model's reasoning, a notable security enhancement.
- Performance on Diverse Tasks: Scoring 1,844 on GDPval-AA, a test across 44 occupations, against Opus 5.5’s 1,846 and Sonnet 5’s 1,449.
Context and Implications:
This release comes amid growing attention to AI safety and cybersecurity. Europe's AI Act requires providers of high-risk models to implement measures like those now incorporated into Sonnet 5.5.
While Anthropic claims Sonnet 5.5 doesn't represent a leap forward in capabilities, they highlight the significant advancements in its cyber defenses, justifying the "frontier-style" safeguards. However, the pricing structure requires further clarification, as the announcement seems to contradict previous pricing models.