Agent infrastructure, open weights, and frontier-model governance lead the last 24 hours. OpenAI packaged its Codex harness as a managed API, gave voice agents a full-duplex layer, and split image generation into speed and precision tracks, while Ant and IBM with NASA each opened large models builders can actually download. An open-source inbox for background agents and new Senate safety talks round out the day. Here are seven stories worth your time, each summarized in our own words with a link to the original reporting.
OpenAI opens the Agents API in public beta
OpenAI released the Agents API in public beta, exposing the harness behind Codex as a managed service for long-running cloud agents. The pitch is that session state, context compaction, crash recovery, and subagent coordination stop being code every team rewrites, with the agent running in an OpenAI-hosted sandbox, the team's own infrastructure, or a partner sandbox. Early testers cited in the announcement report lower latency and cost per task, and there is no added fee beyond tokens and tools used.
Read the full story at OpenAI (September 10, 2026).
GPT-Live-1 brings full-duplex voice to the API
OpenAI launched GPT-Live-1 in the API so voice agents can listen and speak at the same time instead of chaining separate speech-to-text, reasoning, and text-to-speech stages. The voice layer handles interruption, background noise, and turn-taking while delegating deeper reasoning and tool calls to a backend text model the developer chooses. It is priced as a per-minute front-end layer, with broader voice options across accents and languages promised in the coming months.
Read the full story at OpenAI (September 10, 2026).
Images 2.5 splits into Flare for speed and Sunburst for precision
OpenAI rolled out ChatGPT Images 2.5 with sharper detail, steadier multi-turn edits, and lower latency, plus two API models for different budgets. Flare is the default speed-first choice for everyday generation at much lower latency than its predecessor, while Sunburst trades longer generation time for tighter editing control on production creative. New creation helpers include Sketch drawing, reusable templates, and comments placed directly on an image for localized edits.
Read the full story at OpenAI (September 8, 2026).
Ant ships 124B-parameter Ling-3.0-flash-VL under MIT
Ant Group's inclusionAI lab released Ling-3.0-flash-VL, a natively multimodal open-weights model with 124 billion total parameters and only about 5.5 billion active per token. It reads images and video as well as text, carries a context window above 260,000 tokens, and ships under the MIT license with quantized variants for cheaper serving. Independent measurement has since put it well above its size-class median, making it a practical self-hostable vision option for builders.
Read the full story at DataNorth (September 11, 2026).
NASA and IBM open a lunar foundation model
NASA and IBM released the Lunar Foundation Model, an open vision-transformer system for mapping craters, volcanic features, and possible ice from mixed lunar observations. It was trained on roughly two million co-registered tile bundles across resolutions and lighting conditions, and the team reports gains over a SwinV2 baseline on ice prospectivity and crater detection with less training data. Weights, datasets, and benchmarks are public under Apache 2.0, aimed at research rather than certified mission decisions.
Read the full story at IBM Research (September 10, 2026).
AWS open-sources Pizza Bot, an inbox for background agents
AWS released Pizza Bot as an Apache 2.0 application that runs agents in the background and hands their work an inbox instead of a scrolling chat window. The agent returns only when it finishes or gets stuck, which fits meeting prep, follow-ups, summaries, and research that take longer than a live session tolerates. Builders install a desktop build, point it at their own model provider, and keep prompts and code going directly to that provider.
Read the full story at AWS Open Source Blog (September 10, 2026).
Senate negotiators weigh a duty of care for frontier AI
Bipartisan Senate negotiators are discussing legislation that would require leading AI developers to design against catastrophic risks, with the government able to block releases judged unsafe and companies able to appeal in federal court. Talks involve leadership and Commerce Committee figures, and part of the draft would preempt state laws on the same risks for a single national frame. Nothing is introduced yet, but the direction matters for release timelines, testing budgets, and compliance planning.
Read the full story at The Star (September 12, 2026).