This brief covers AI news from 2026-09-11 UTC.
OpenAI opened the day with a managed agent runtime and closed it with a new frontier model, while chip and coding news kept pace.
OpenAI launches Agents API in public beta
What happened: OpenAI introduced the Agents API in public beta, a managed service that runs the Codex harness and infrastructure for long-running cloud agents. Developers specify a task, model, tools and environment, and can run agents in an OpenAI sandbox, on their own infrastructure, or with a sandbox partner.
Why it matters: Teams building agents can drop their own orchestration, context compaction and subagent plumbing and pay OpenAI to run the loop instead.
OpenAI unveils GPT-6 Astra, its most capable model yet
What happened: OpenAI introduced GPT-6 Astra, which it says scores 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. It is rolling out to a limited set of organizations today, then to ChatGPT Plus, Pro, Business and Enterprise users and through the OpenAI API, Microsoft Azure and AWS Bedrock.
Why it matters: Builders on ChatGPT plans, the API, Azure or Bedrock get a stronger default model for computer use, browsing and software engineering without changing providers.
Cognition releases SWE-2 coding model
What happened: Cognition introduced SWE-2, a coding model post-trained from Kimi K3 that scores 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while costing 64% less. Cognition says it scaled reinforcement learning to the multi-trillion-parameter regime for the first time.
Why it matters: Coding agents can get near-frontier scores at a fraction of the price, which changes the math on running large fleets of them.
DOJ investigates Nvidia’s $20B Groq licensing deal
What happened: The Justice Department is investigating Nvidia’s $20 billion non-exclusive licensing agreement with Groq to determine whether the company tried to skirt antitrust scrutiny, according to the New York Times. Axios reported that the structure was previously described as designed to avoid tripping antitrust wires.
Why it matters: If the arrangement is found illegal, the same licensing structure used across other AI industry deals could face the same scrutiny.
Cursor launches Projects for fleets of agents
What happened: Cursor launched Projects, which maintains context over months of work, delegates tasks to thousands of subagents and performs recurring work without prompting. Cursor says new users merge 30% more PRs and users who primarily use Projects merge six times as many.
Why it matters: Developers can hand off a migration or full feature to a coordinator agent that keeps working after the laptop closes.
OpenAI adds GPT-Live-1 full-duplex voice to the API
What happened: GPT-Live-1 is now in the API, handling full-duplex voice with interruption handling and delegating reasoning to a backend model the developer chooses. OpenAI reports 97.3% on Artificial Analysis’s Conversational Dynamics benchmark and 0.798-second response latency, at $0.05 per minute billed per second.
Why it matters: Voice apps can keep a conversation alive while a backend agent works, with WebRTC, WebSocket and telephony connections and separate billing for the backend.
Source: OpenAI Developer Community
Positron’s valuation jumps in new funding round
What happened: Reuters reported that AI chip startup Positron’s valuation skyrocketed in its latest funding round. The report did not disclose the round size or the new valuation.
Why it matters: Another well-funded chip challenger signals continued investor appetite for alternatives to incumbent AI accelerators.