OpenAI puts its full-duplex voice model in the API at $0.05 a minute
OpenAI shipped twice in one day — a voice model developers can finally build on, and a finance product that has Wall Street asking what happens to the apprenticeship. Both push the same thesis: AI stops being an assistant you prompt and becomes infrastructure you assemble.
OpenAI put GPT-Live-1, its full-duplex voice model, into the API at $0.05 per minute for the voice layer. One model now handles listening and speaking at once — reasoning over incoming and outgoing audio together instead of chaining speech-to-text, a language model, and text-to-speech, where every handoff adds latency and drops conversational rhythm. The model can delegate deeper reasoning and tool calls to a backend model like GPT-6 Astra (or a third-party one), take tone and pacing direction through the system prompt, and run over phone lines with telephony support built in — restaurant reservations, customer support, the whole call-center stack. OpenAI reports a 30-point jump over GPT-Realtime-2.1 on Full Duplex Bench and the top Tau3 score when paired with Astra. Early customers: Yelp says its Host reservation agent handles background noise and side conversations better; Speak reports roughly 80% fewer interruptions during learners' thinking pauses; one healthcare customer cut 23,000 lines of code replacing its cascaded voice stack. Why it matters: voice agents stop being a demo and become a product category you assemble from parts — cheap conversation front end, reasoning backend of your choice. The cautious read: every benchmark here is the vendor's own, and "feels natural" is exactly the kind of claim those benchmarks measure only indirectly.
OpenAI also launched ChatGPT for Financial Services, built with Morgan Stanley and Evercore, targeting the research and pitchbook work of Wall Street's junior bankers. The product runs on GPT-6 Astra with native data access from LSEG, Daloopa, and PitchBook, citations that trace figures back to source filings, and administrative controls for sensitive deal materials. In the launch demo it analyzed an M&A target, pulled comparables into a spreadsheet, and generated a formatted deck against a bank's style guide. OpenAI framed it as teaching ChatGPT to research like an analyst — and when CNBC asked whether it reduces hiring, compared it to Excel: a tool that made analysts better, not redundant. The harder question is the apprenticeship model underneath. Goldman's AI lead warned last month that automating exactly these training tasks risks "cognitive atrophy" in the next generation of financiers — if the grunt work that teaches juniors to reason disappears, banks still need seniors, but the pipeline that produces them breaks. That is a slower, quieter disruption than layoffs, and harder to reverse.
What to watch: who ships the first generally available competitor voice stack — and whether banks staff analyst classes of 2027 like classes of 2025.
Would you trust a pitchbook no human ever touched — or an AI receptionist answering your dentist's phone? Tell us in the comments.
