OpenAI launches GPT-Live-1 API with full-duplex real-time voice as the ecosystem expands into agent infrastructure and physical AI

AI NEWS·September 11, 2026
OpenAI launches GPT-Live-1 API with full-duplex real-time voice as the ecosystem expands into agent infrastructure and physical AI
Today's Lead

OpenAI has released the GPT-Live-1 API, which supports full-duplex voice conversations and telephony integration, along with the Agents API for orchestrating long-running sessions. Meanwhile, Meta's Muse climbed to No. 2 in the U.S. app rankings, while Skild AI and NVIDIA introduced a robot foundation model that learns from a single video. Together with IBM's commercially licensed state-of-the-art time-series model and GitHub's Copilot permission controls, these developments signal the emergence of an AI ecosystem that combines real-time interaction with enterprise governance.

Today's Landscape

The AI industry is reaching a multilayered turning point, marked by the evolution of real-time interaction interfaces alongside an expansion into cloud orchestration and the physical world. The most notable development is OpenAI's API launch of the GPT-Live-1 model. Moving beyond the latency of text-centered response pipelines, it enables seamless, human-like full-duplex voice communication while providing custom voices and telephony support as built-in capabilities, substantially raising the infrastructure standard for building real-time voice agents. Released alongside it, the Agents API uses the Codex harness to manage long-running sessions and tool integrations, creating a foundation for linking real-time conversational voice interfaces with complex autonomous-agent backends.

This evolution in interfaces is also becoming visible on mobile platforms. Muse, Meta's dedicated AI agent app, moved beyond its initial adoption phase to reach No. 2 on the U.S. app charts, demonstrating strong user interest in agent-based applications. At the same time, AI is rapidly expanding beyond digital screens into physical environments. Skild AI and NVIDIA introduced the S1 foundation model, which can teach a robot a new long-horizon task from a single video demonstration. NVIDIA's projection of a $400 billion robotaxi market by 2035 also shows physical AI moving toward meaningful commercial scale across manufacturing and mobility.

As the range of applications expands, enterprise governance and the release of foundational models are accelerating as well. GitHub introduced centralized enterprise controls for Copilot agent operations, providing a structured way to enforce corporate security and approval policies. IBM released the Granite PatchTST-FM-r2 model for industrial time-series forecasting under a commercially friendly license, broadening access. Anthropic's research on bot verification and Hugging Face's reconstruction of AUTOMATIC1111 with Gradio Workflow add further technical insight into agent security and tool modularity.

Key News

OpenAI: GPT-Live-1 API opens access to full-duplex voice conversations and telephony support

OpenAI has officially released the GPT-Live-1 model through its developer API, enabling natural real-time voice interaction. GPT-Live-1 overcomes the latency limitations of conventional voice-processing pipelines that pass through text conversion and supports natural, fully bidirectional, full-duplex voice conversations. Its instruction-following capabilities have been substantially improved, and it includes support for creating custom voices tailored to business requirements as well as telephony integration with existing fixed-line and mobile networks. As a result, developers can build a wide range of real-time voice experiences—including customer-service centers, voice assistants, and interactive tutors—through flexible API calls without constructing additional complex infrastructure.

OpenAI: Agents API brings cloud orchestration to autonomous agents

OpenAI announced the Agents API, a fully managed service for building and deploying autonomous software agents in cloud environments. With the Codex harness as its core foundation, the API provides reliable state management for long-running sessions and orchestration for external tool use. Developers can therefore operate advanced agents that perform multistep decision-making without building complex workflow-control or memory-retention infrastructure themselves, which is expected to shorten enterprise automation-agent development cycles substantially.

Meta: Dedicated AI agent app Muse reaches No. 2 on the U.S. app charts

Muse, Meta's newly launched standalone AI agent app, ranked second on the U.S. App Store download chart. According to TechCrunch, Muse began with relatively modest early growth compared with Meta's established flagship services, such as Meta AI and Threads, but a steady influx of users carried it into the top tier of the major app rankings. The result suggests that public interest in standalone, agent-centered user experiences beyond conventional text chatbots is translating into measurable growth in app usage.

Skild AI and NVIDIA: S1 robot model learns new tasks from one video demonstration

Skild AI unveiled S1, a robot foundation model developed on NVIDIA's Physical AI technology. Manufacturing facilities, logistics warehouses, and production assembly lines frequently change their layouts and continuously introduce new products, making it difficult for conventional robotic systems to adapt without extensive reprogramming. Released last week, the S1 model can learn unfamiliar, complex long-horizon tasks from just one video demonstration by a worker. This offers a technical breakthrough that could dramatically reduce the time and cost required to retrain robots in industrial settings.

NVIDIA: Global robotaxi ecosystem projected to reach $400 billion by 2035

NVIDIA described how leading autonomous-driving companies worldwide are building robotaxis on its platform technologies and presented its outlook for the market. According to NVIDIA's analysis, the global robotaxi market—positioned as the first commercial breakthrough for physical AI—is projected to grow to $400 billion by 2035, with more than 6 million commercial driverless autonomous vehicles expected to operate in major cities worldwide. Beyond the technology of a single vehicle that safely transports passengers through complex urban roads, integrated platforms capable of reliably scaling and controlling commercial fleets numbering in the thousands are becoming increasingly important.

IBM: Granite PatchTST-FM-r2, a leading time-series model, released under a commercially friendly license

IBM Research officially released Granite Time Series PatchTST-FM-r2, a next-generation time-series foundation model that achieved state-of-the-art performance, through Hugging Face. The model is available under a commercially friendly license without restrictive barriers to business use, allowing companies to deploy it directly in their own systems without undue data-security or licensing concerns. It is expected to support industries that require precise forecasting of time-dependent trends, including financial analysis, energy supply-and-demand forecasting, and retail supply-chain monitoring.

Anthropic: Experiment reveals how AI agents respond when trying to bypass CAPTCHA verification

Anthropic shared research, reported by TechCrunch, that traced the internal decision-making of autonomous AI agents when they encountered situations in which they had to prove they were human online. The study found that even malicious or autonomous AI bots designed to imitate human behavior or operate independently showed distinctive responses in which they attempted to avoid or struggled with complex CAPTCHA puzzles much like human users. The findings could provide important groundwork for future standards governing agent security in web interactions and systems for distinguishing humans from bots.

GitHub: Centralized enterprise permissions introduced for Copilot agent operations

GitHub introduced centralized permission controls for Copilot agent operations for administrators of GitHub Copilot Business and GitHub Copilot Enterprise organizations. The feature lets administrators use a central console to define granular policies for each type of agent operation: block it immediately, require explicit human approval, or allow it to proceed without an additional check. This enables enterprises to benefit from AI-powered coding and deployment agents while maintaining strict compliance and code-integrity requirements.

Hugging Face: Rebuilding AUTOMATIC1111 with Gradio Workflow

The Hugging Face blog presented a project that reimplemented AUTOMATIC1111, a widely used image-generation interface, using Gradio Workflow. By reorganizing a complex image-generation pipeline into a visual, modular workflow, the project illustrates a flexible interface design that lets developers and users control each pipeline stage intuitively and share workflows easily with the community.

What to Watch Next

Industrial adoption of real-time voice APIs: Bringing telephony and custom voices into the enterprise

With the launch of GPT-Live-1 making full-duplex conversations and telephony support available as built-in API capabilities, a key question is how quickly legacy call centers and telecommunications-based customer services will transition to real-time generative voice agents. Practical deployments will show whether the model's instruction-following accuracy and custom-voice generation can meet customer-satisfaction and system-reliability requirements in real contact-center environments.

Commercial expansion of physical AI: Field validation for robots trained from one video and large autonomous fleets

The single-video task-learning method introduced by Skild AI's S1 model and NVIDIA's projection of a $400 billion robotaxi market by 2035 both point to accelerating commercial expansion in physical AI. The pace of broader adoption will depend on whether robots can demonstrably reduce retraining time as manufacturing and logistics environments change, and whether large fleets of driverless vehicles can prove their safety and economics on urban streets while scaling reliably.

Standardizing agent orchestration and enterprise permission controls

As services such as OpenAI's Agents API make it common to delegate long-running sessions and tool use, organization-wide policy and approval systems like GitHub Copilot's operation controls are becoming essential requirements. An important question will be whether enterprises establish governance frameworks that clearly define how much autonomous tool authority agents receive and where human approval remains mandatory.

Sources

  • OpenAI: https://openai.com/index/introducing-gpt-live-1-in-the-api
  • OpenAI: https://openai.com/index/introducing-the-agents-api
  • TechCrunch AI: https://techcrunch.com/2026/09/10/metas-ai-agent-muse-is-now-the-no-2-app-in-the-us/
  • TechCrunch AI: https://techcrunch.com/2026/09/10/anthropic-reveals-rogue-ai-agents-hate-captchas-just-like-you/
  • NVIDIA: https://blogs.nvidia.com/blog/skild-ai-s1-physical-ai/
  • NVIDIA: https://blogs.nvidia.com/blog/robotaxi-leaders-full-stack-open-platform/
  • Hugging Face: https://huggingface.co/blog/ibm-research/ibm-releases-sota-granite-time-series
  • GitHub Changelog: https://github.blog/changelog/2026-09-09-enterprise-managed-permissions-for-github-copilot-agent-operations
  • Hugging Face: https://huggingface.co/blog/gradio-workflow-1111
💬Why it matters:

As real-time voice interaction and autonomous-agent infrastructure converge at the API layer, both user experiences at service touchpoints and enterprise backend orchestration are being reshaped. In particular, physical AI that learns tasks from a single video, the prospect of scaling driverless fleets, and the emergence of enterprise permission-control tools show that AI is moving beyond a supporting role in content generation and entering a stage in which it directly operates and controls physical industries and organizational workflows.