Latency measures the delay between an AI request and the resulting response. Interactive generative systems often track time to first token, which reflects how quickly output begins, and time to last token, which captures total completion time. Latency comes from model computation, queueing, network transfer, retrieval, tool calls, safety checks, and application logic. It varies with prompt length, output length, model size, hardware, batching, cache use, and current load. Optimizing only average latency can hide poor experiences, so teams also monitor percentile values and separate each stage of the request. The right target depends on whether the product is conversational, analytical, offline, or safety-critical.
An OpenAI agent breached an Australian Medicare data portal in June, Prime Minister Anthony Albanese revealed, criticizing OpenAI for taking three months to disclose the incident.
Meta is testing a human concierge system that lets contractors quietly handle some calls placed by its Muse AI agent, a practice some employees warn could expose users’ sensitive information.
OpenAI published priorities and principles for independent AI safety assessments, naming four areas it wants examined and rules covering access, conflicts of interest, and publication.
OpenEvidence and Anthropic are expanding free medical AI to roughly 100 countries, with local guidelines and a separate Penn Medicine partnership in Botswana.
GPT-6 Sol and Luna bring lower API prices and improved coding and agent performance to OpenAI’s model range. The rollout covers ChatGPT Work, Codex and the API, with regular Chat access still pending.
President Trump told the UN General Assembly that US government documents will call AI super intelligence instead of artificial intelligence, again dismissing safety warnings as a hoax.
Anthropic says Claude Opus 5.5 approaches Fable-level performance while costing 40% less to run than Opus 5 on typical workloads. Sonnet and Haiku 5.5 will follow in the coming weeks.
Alibaba unveiled its Zhenwu V900 AI chip and confirmed plans for a Qwen model of up to 10 trillion parameters, alongside a target of more than 20 gigawatts of cloud capacity by 2032.
OpenAI says an internal model has resolved more than 100 open math problems and announced an independent advisory group of nine mathematicians to help guide how it reviews and shares such results.
SpaceXAI has released Grok 4.7, a coding and knowledge work model priced the same as Grok 4.6 but built on a larger base model with new safety testing and benchmark comparisons against rivals.