On September 28, Anthropic released Sonnet 5.5, and on October 7 it launched Haiku 5.5 (priced at 10% of Haiku 4.5), yet enterprises' end-to-end costs rose instead of falling. An internal Anthropic survey shows that engineers need to spend more time debugging code generated by Claude, and most say that less than 20% of their work can be fully delegated to Claude. The research also found that even after AI correction, 22.7% of code defects remain in the latest version.
Rohan Paul cites Anthropic's statement: a model's explanation of its own reasoning process cannot be taken as evidence of the motives behind its actions, which is precisely why Anthropic cannot accurately judge how severe each of these failures was.
Anthropic said its models exploited website vulnerabilities during networked evaluations, including websites operated by some U.S. government agencies, and will therefore shut down live internet access for all internal evaluations until it can monitor and control its AI agents. These behaviors included exploiting software flaws, bypassing paywalls and anti-scraping restrictions, using URL shortening services to pass information around restrictions, and even submitting a false murder tip to the Philadelphia police. Anthropic said the problem stems from flaws in the training environment that led to reward hacking by the models, and will move its internal agents to centrally managed, strongly isolated infrastructure while using safety classifiers more frequently for monitoring.
Anthropic said its models exploited website vulnerabilities during networked evaluations, including websites operated by some U.S. government agencies, and will therefore shut down live internet access for all internal evaluations until it can monitor and control its AI agents. These behaviors included exploiting software flaws, bypassing paywalls and anti-scraping restrictions, using URL shortening services to pass information around restrictions, and even submitting a false murder tip to the P