GPT-6 Astra, GPT-Live-1, and Meta's sandboxed agent

GPT-6 Astra: The next generation in intelligence for work — is pitched as OpenAI's "most capable model for business, with advanced reasoning, computer use, and stronger writing and design judgment." Computer use plus better reasoning is the part that reshapes agent architecture and your eval suite.

Conspicuously absent from the announcement: benchmarks, pricing, and API details. Treat the capability claims as a roadmap until someone runs them. — openai.com

Build more natural voice experiences with GPT-Live-1 in the API — brings "natural, full-duplex voice conversations to the API, with stronger instruction following, custom voices, and telephony support." The practical win: prototyping a voice agent without duct-taping STT, an LLM, and TTS together — and the latency tax that stack charges. — openai.com

Meta Introduces Muse, a Personal AI Agent on Its Own Secure Cloud Computer — gives each user a dedicated cloud sandbox so the agent can "send emails, book travel, negotiate bills, and pursue long term goals." Per-user isolation plus approval checkpoints is the right shape for persistent agents — the open question is what happens the first time one negotiates the wrong bill. — Michal Sutter

POLICY

NYC mayor Mamdani imposes 1 year ban on AI for schools through 8th grade — pauses AI use through eighth grade across one of the largest school systems in the US. If you sell classroom copilots, that's a year of lost pipeline and a new bar for vendor safety requirements — and a template other districts will copy. — DeepLogin

DEBATE

AI handles incidents, engineers lose touch with their systems — argues that handing incident response to agents erodes the tacit knowledge engineers build by debugging live systems. The prescription is unglamorous: keep humans in observability reviews, keep escalation paths explicit, and don't let the runbook become a black box. — sylvainkalache

Funding grants for new research into AI and teen development — is a $5 million program for independent research on "how generative AI affects teen development, well-being, and safety." Applications are open, and the output is the kind of evidence safety teams currently have to guess at. — openai.com

AI 2027 (2025) — forecasts rapid capability gains, autonomous AI research, and geopolitical fallout through 2027. Read it as a planning artifact whose assumptions you stress-test, not as evidence about anything. — gafferongames

Artificial Analysis Intelligence Index v4.2 — bumps the composite model-ranking index another notch. The announcement doesn't spell out methodology changes, so check the details before letting v4.2 decide your model selection. — nojs