Tag
Study shows GPT and Claude exhibit distinct, unreliable repair behaviors in multi-turn math dialogues, with some models resisting correction and others over-correcting.
DELEGATE-52 is a new benchmark revealing that current LLMs, including frontier models like GPT-5.4 and Claude 4.6 Opus, corrupt an average of 25% of document content during long delegated workflows across 52 professional domains. The research demonstrates that LLMs introduce sparse but severe errors that compound over interactions, raising concerns about their reliability for delegated work paradigms.
The article discusses the growing importance of reliability, security, and user protections as AI models become more capable and personalized.
Google introduces Flex and Priority inference tiers for the Gemini API, offering developers granular control over cost and reliability for synchronous requests. Flex provides 50% savings for latency-tolerant tasks, while Priority ensures high reliability for critical applications.
Anthropic released a postmortem detailing three infrastructure bugs between August and September that intermittently degraded Claude's response quality. The report explains the technical causes, including context window routing errors, and outlines measures to prevent future incidents.