This Week in AI: Navier-Stokes Proof, OpenAI Agents API, and DeepSeek V4.1-Flash

This week pushed AI from frontier math to production agents: OpenAI claimed a Millennium Prize breakthrough and opened its agent harness to developers, while DeepSeek, Cohere, and voice APIs kept the model race moving on price and capability.

Models & Products

Introducing DeepSeek-V4.1-Flash — DeepSeek released a 552-billion-parameter MoE model with native multimodal support and up to one-million-token context, cutting API prices and routing legacy V4 Pro traffic to the new Flash tier starting September 14.

Build more natural voice experiences with GPT-Live-1 in the API — OpenAI brought ChatGPT's full-duplex voice model to the API at $0.05 per minute, letting developers pair real-time listening and speaking with backend reasoning models and tools.

Introducing ChatGPT Images 2.5 — OpenAI shipped a sharper image model with up to 50% lower latency, plus Sketch, templates, and on-image comments in ChatGPT, alongside GPT-Image-2.5 Flare and Sunburst variants in the API.

North Small Translate: sovereign open-weight machine translation — Cohere open-sourced a 218B-parameter MoE translation model covering 50+ languages, scoring 83.6 on WMT26 benchmarks and positioning sovereign translation as a first-class North family release.

Agents & Tools

Introducing the Agents API — OpenAI entered public beta with the same harness behind Codex: hosted sandboxes, MCP and programmatic tool calling, automatic context compaction, and multi-agent delegation in a single API call.

ADK for Kotlin 1.0 — Google reached GA for its Agent Development Kit on Kotlin Multiplatform, adding on-device LiteRT-LM support, Firebase AI Logic orchestration, and Room/AppSearch persistence for Android agents.

Introducing Projects — Cursor launched Projects in beta, giving developers a coordinator that retains long-term context, delegates to subagents, and can run on schedules or Slack triggers for migrations and multi-PR work.

Research

On the Navier–Stokes Millennium Prize Problem — OpenAI published an analytical proof and Lean formalization that smooth three-dimensional Navier–Stokes flow can develop a finite-time singularity, claiming resolution of a Clay Millennium Prize problem via a large multi-agent research system.

NASA-IBM Lunar Foundation Model — NASA and IBM open-sourced a foundation model trained on Lunar Reconnaissance Orbiter data to map craters, ice deposits, and volcanic features, extending the Prithvi geospatial model family to lunar science.

Policy & Health

ARPA-H ADVOCATE program awards — ARPA-H committed up to $62.7 million over four years to build an FDA-authorized agentic AI cardiovascular care system, working with the agency throughout to define regulation for patient-facing clinical agents.

Humanity AI $10 Million Open Call — MacArthur-backed Humanity AI opened a U.S. nonprofit grant round through October 21 to fund community-led projects shaping how AI affects education, labor, and arts and culture.

The through-line is infrastructure maturing in public: agent harnesses and long-running coordinators are shipping alongside models, while governments and health agencies are sketching the guardrails those systems will need in the wild.