AI Daily — September 11, 2026
2026-09-11
openai.com
GPT-Live-1 brings full-duplex voice to the API with telephony support
OpenAI released GPT-Live-1 via API, enabling real-time, full-duplex voice conversations with improved instruction following, custom voice configuration, and native telephony integration. This extends the Realtime API capabilities meaningfully for developers building call-center, IVR, and voice-agent applications. Custom voices and stronger turn-taking make it a more production-ready offering than earlier Realtime API versions.
openai.com
ChatGPT for Financial Services launches with GPT-6 Astra and built-in financial data
OpenAI is launching a vertical-specific ChatGPT product for financial services that bundles GPT-6 Astra with curated financial data feeds, targeting research, financial modeling, and client deliverable generation. This marks OpenAI's clearest move yet into domain-specific enterprise verticals with pre-integrated data rather than relying on operator-side RAG setups. The product competes directly with Bloomberg Terminal AI and similar finance-focused LLM offerings.
openai.com
ChatGPT Work 'Data agent' connects company data to interactive dashboards via natural language
OpenAI's new Data agent within ChatGPT Work allows enterprise users to connect internal data sources and generate interactive dashboards and analyses using natural language queries. The feature is aimed at replacing or augmenting traditional BI tooling without requiring SQL or visualization expertise from end users. It represents a significant expansion of ChatGPT's agentic data workflow capabilities for enterprise.
openai.com
OpenAI and GSA offer $0 licensing and 50% usage discounts to US governments
OpenAI and the General Services Administration are partnering to provide federal, state, local, and tribal governments with zero-cost licensing and 50% off usage fees, alongside expanded cyber defense support. This is a substantial procurement deal that positions OpenAI as a primary AI vendor across US government at multiple levels. The cyber defense component suggests OpenAI models will be used for threat detection and security operations workflows.
rss.arxiv.org
Subagents vs. Agent Skills: invoking skill packages as subagents improves long-horizon task robustness
This paper investigates whether agent skill packages—multi-file bundles of instructions, scripts, and resources—are better executed by loading them into context (current norm) or by invoking them as independent subagents. The key finding is that context-loading degrades reasoning quality as task horizons grow due to context accumulation, whereas the subagent invocation pattern maintains performance. The work has practical implications for how multi-agent orchestration frameworks should handle reusable capability modules.