• OpenAI announced on X on September 23, 2026 that ChatGPT Voice can now invoke plugins for email, calendar, and Slack, and runs on user-selectable GPT-6 Astra, Sol, and Luna models.
• Those connectors are not new: Gmail, Google Calendar, Slack, and GitHub have been available inside text-mode ChatGPT and ChatGPT Work since Work's July 10, 2026 launch, per VentureBeat's original coverage.
• Astra has run on GPT-6 for a few weeks; Sol and Luna were upgraded to GPT-6 the day before this announcement, per OpenAI and 9to5mac.
• VentureBeat flagged ChatGPT Work's connector access at launch as "a fundamentally different data surface area" warranting enterprise security review — no confirmation step for voice-triggered sends has since been documented.
• The "Her" framing is marketing shorthand Sam Altman has used since ChatGPT Voice's 2024 debut; the actual change here is a plumbing update, not a new assistant capability.

OpenAI said on September 23, 2026 that ChatGPT Voice can now call the same email, calendar, and Slack plugins available in text-mode ChatGPT, running on the company's newly completed GPT-6 lineup — Astra, Sol, and Luna. The announcement, posted by OpenAI's official account on X, frames the update as OpenAI listening to user complaints that Voice couldn't touch the tools text-based ChatGPT already used. That framing is accurate, and it is also the whole story: this is a plumbing fix that lets voice commands reach connectors that have existed for months, not a new set of capabilities.

What OpenAI Actually Shipped

Per OpenAI's own post, three things changed at once: Voice gained access to plugins including email, calendar, and Slack; it became selectable across GPT-6 Astra (OpenAI's largest model, running for a few weeks already), Sol, and Luna (both upgraded to GPT-6 the day before, September 22, per 9to5mac's reporting); and Voice now works inside ChatGPT Work on web and mobile, letting a spoken request produce documents, slide decks, or a website. None of the three is a standalone new feature — each extends an existing OpenAI product line into the voice modality.

The Connectors Aren't New

Gmail, Google Calendar, Slack, and GitHub connectors have been part of ChatGPT Work since it launched on July 10, 2026, running on GPT-5.6 and built on MCP-based plugins, according to VentureBeat's launch coverage. Those same connectors were also already reachable from text-mode ChatGPT. What changed on September 23 is that a spoken request can now trigger them, where before a user had to type. Judged against that baseline, "ChatGPT Voice gets closer to Her" overstates a backend wiring change as a leap in assistant capability.

The Governance Gap OpenAI Didn't Address

When ChatGPT Work launched in July, VentureBeat reported that OpenAI's only stated enterprise privacy measures were zero data retention and an opt-out from model training, and flagged the connector access itself as "a fundamentally different data surface area" that would need dedicated security review. Neither OpenAI's July documentation nor its September 23 Voice announcement, as reported, describes a confirmation step before a spoken command sends an email or posts to Slack. Voice interfaces are typically lower-friction to trigger than a typed request — a stray or misheard phrase can issue a command a text interface would have required a user to review first. That gap in documented safeguards, not the model upgrade, is the part of this launch worth enterprise security teams' attention.

Verdict

This is a legitimate, useful convenience upgrade for people already on a Plus, Pro, Team, or Enterprise plan who use ChatGPT Work's connectors — hands-free access to tools they already trusted in text form. It is not the "Her"-style leap the headline implies: the connectors, the agentic execution, and even the underlying GPT-6 Astra model predate this announcement by weeks. The one open question — how a voice-triggered send or post gets confirmed before it fires — has no public answer as of this reporting. Until OpenAI publishes one, treat the expanded voice-triggered action surface as running on the same governance OpenAI described for ChatGPT Work in July: functional, but not yet accompanied by a documented safeguard for the lower-friction interface now allowed to invoke it.