Latest AI intelligence

The September 10 beta provides durable sessions, progress streaming and sandbox choices while OpenAI manages orchestration, context compaction and recovery.

Authorized administrators can review changes made through Codex Policies & Configurations in the UI or retrieve them through the Compliance API.

Administrators can cap newly created keys' lifetimes at organization or project level, making rotation a concrete production task.

The UAI paper lets agents share only with peers that have similar local causal effects, but its gains remain theoretical and synthetic.

A two-week training pause has ended only in part: OpenAI says its biggest planned run remains blocked by stricter security and monitoring gates.

The company is linking machine-led triage to bounded responses while keeping people responsible for its highest-impact security decisions.

The new MultiVectorEncoder brings ColBERT-style late interaction into the main library, widening support for token-level text and visual search while increasing index and migration complexity.

The research prototype approaches a DXA-based classifier in one 132-person validation cohort, but it is not a diagnostic product or evidence of clinical readiness.

The new experience automatically routes identified 13-to-17-year-olds into teen-specific learning tools and protections, while independent evidence about outcomes remains absent.

The August 18 cutoff removes Deep Research from the consumer Copilot app while preserving saved reports and steering Microsoft 365 Premium users to Researcher.

The August 17 cutoff affects saved legacy prompts, evals and three experimental prompt endpoints, while Anthropic's updated Workbench uses a reduced persistence model.

A controlled vendor benchmark combines hybrid retrieval with compressed Qwen inference, raising successful task throughput while exposing important workload and cost boundaries.

Customers can choose European or US processing, request a preview service tier and run GLM-5.2 on Mistral infrastructure, while residency and capacity claims remain vendor-defined.

The agent-focused model keeps its existing API name while adding native Responses API support, new reasoning controls and a revised time-based pricing schedule.

Unity AI Gateway can choose a model for each coding task, but the beta relies on first-turn classification, restricted integrations and vendor-reported savings.

Stund links conversation, edits and media to a shared canvas so AI can query and reshape meeting history, but the prototype has no user study and only minimal sandboxing.

Eligible Claude Enterprise reviewers can now retrieve Claude Code and Cowork conversations, but coverage excludes ZDR, HIPAA-ready tenants, some cloud surfaces and device activity that never reached the API.

Consumer and Edu users can answer quiz questions inside a chat, but OpenAI has not published assessment-quality or learning-outcome evidence.

Eligible unshared projects can now move between default and project-only memory, while shared projects stay isolated and ChatGPT Work remains unavailable in project-only mode.

Eligible web users can browse Drive content, add files or folders to Chat and Work, and keep Google documents beside a conversation, while Shared Drives and mobile remain outside the first rollout.

The timeline records selected interaction events rather than screenshots or audio, while access, administrator approval and regional exclusions limit the initial rollout.

The new desktop build supports five current Ubuntu, Debian and Fedora releases worldwide, with browser actions enabled but control of other Linux applications still unavailable.

The controlled red-team exercise shows how a modified open-weight model can preserve ordinary task performance while hiding a trigger-bound objective, but it does not establish how common such tampering is.

The public evaluation pairs 2,165 expert-transcribed pages with a reproducible harness, while its six-volume natural-history sample limits how broadly the scores can travel.