ChatGPT Voice can now use email, calendars, Slack, and other connected services to carry out spoken instructions. This matters if you want ChatGPT to do more than answer questions and instead assist with concrete workflows. At the same time, OpenAI is introducing GPT-6 Astra, Sol, and Luna, three models whose capabilities and prices affect how economically these workflows can run.
What can the new voice agent do?
According to the main report on the ChatGPT Voice update, the globally available voice feature is powered by GPT-6 Astra, Sol, and Luna. Plugins, meaning extensions that provide access to outside services, let the assistant do more than discuss a task: it can act on it. This moves ChatGPT toward an Artificial Intelligence (AI) agent, a system that completes several steps with a degree of autonomy.
One everyday example is calendar management. You can ask ChatGPT to find an available time and then draft or send an email. At work, the assistant can summarize Slack messages, prepare a document, or create a presentation. OpenAI’s demonstration also includes checking the weather, shopping online, and building a website with a checkout page.
The example involving financial services goes further. ChatGPT is supposed to detect duplicate charges in connected financial apps and arrange a cancellation. This also illustrates the difference between a useful response and an action with real consequences: an inaccurate summary is irritating, while an incorrect payment action or cancellation can be considerably worse.
Moving between voice and text should become smoother as well. The report on the mobile rollout says voice conversations will produce richer text output, and you will be able to start a conversation on the go before continuing it on a desktop. OpenAI previously launched GPT-Live as a conversational model and added voice commands to the desktop app’s work and coding areas; similar capabilities are now reaching phones.
Who gets access to the features?
The reports describe availability at different levels of detail. The main report says the voice feature is available worldwide in the latest app version and identifies ChatGPT Work on both the web and mobile devices. TechCrunch adds subscription distinctions: Plus and Pro users should be able to use the mobile Work tab to create documents, emails, websites, and presentations or summarize Slack content.
Free and Go users are listed as receiving access to plugins and connected apps. That does not mean every feature is identical across every plan. What you can actually do appears to depend on your account, the ChatGPT area you are using, and the services you have connected, but the sources do not provide a complete feature matrix.
For Switzerland, the key point is the stated global rollout: the update is not described as limited to the United States and should be available through the latest app version. However, the sources provide neither separate Swiss pricing nor specific information about performance with Swiss German. They also do not explain how voice recordings or data from connected workplace accounts are stored.
OpenAI continues to keep chat and workspaces separate. During the same period, Anthropic simplified handoffs between mobile and desktop and combined its Cowork and Chat interfaces. The comparison suggests that providers are pursuing similar assistant concepts while organizing their products differently.
What changes with GPT-6 Astra, Sol, and Luna?
The three models occupy different positions. Astra is the most expensive OpenAI model in the published pricing comparison and appears in company examples involving demanding, context-heavy work. Sol sits in the middle, while Luna targets particularly low costs for frequent or large-scale tasks.
The published prices apply to the Application Programming Interface (API), which lets companies and applications use a model automatically. They are not the same as ChatGPT subscription prices. The sources do not give new monthly fees for ChatGPT Plus, Pro, Go, or Work, so “50 percent cheaper” does not automatically mean your own ChatGPT bill will be cut in half.
The release report covering Sol and Luna lists GPT-6 Sol at $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 for input and $0.50 for output. A token is a small unit of text used to measure the amount of material a model processes.
A more detailed assessment of the new pricing contest includes cached input: $0.01 for Luna, $0.20 for Sol, and $1 for Astra per million tokens. Astra costs $10 for regular input and $50 for output. Prompt caching means reusing instructions and context the model has already processed, which can reduce the cost of agents that run for extended periods.
Sol and Luna are each described as costing roughly half as much as their GPT-5.6 predecessors, although the GPT-5.6 comparison rates are identified as promotional pricing. At the same time, the early assessment gives a reason not to draw firm quality conclusions: the models were still too new for a reliable evaluation. Lower prices can be measured immediately, but equally strong performance for every task cannot.
Pros and Cons of voice-controlled work
Pros:
- Fewer app switches – You can handle appointments, messages, and documents through one conversation instead of opening each service separately.
- Mobile continuity – You can start a task by voice while away from your desk and continue it later on a computer.
- Lower workflow costs – Sol and Luna’s lower API prices could make frequent automated tasks more economical.
- More than text – According to the reports, ChatGPT can also produce presentations, websites, and structured work documents.
Cons:
- Real-world mistakes – An agent can do more than give a wrong answer; it might send an unsuitable email, change an appointment, or trigger a financial action.
- Unclear plan boundaries – The reports describe different access rights but do not provide a complete overview of every account and feature.
- Sensitive connections – Email, calendars, Slack, and financial apps contain data whose permissions and storage are not explained in detail by the sources.
- Limited independent testing – The new models and the featured company results had not been extensively evaluated independently at the time of publication.
Voice control reduces the effort required to combine several steps into one instruction. It can also obscure which application is supplying information or carrying out an action. A clear confirmation before external actions would therefore be useful, but the sources do not say which ChatGPT tasks require that approval.
What does this mean in practice?
If you are a beginner, start with a task whose consequences are easy to review. For example, you could ask ChatGPT to summarize one day of connected calendar events or prepare an email draft without sending it immediately. This lets you check names, times, and context before the agent takes action.
If you are an advanced user, you can combine longer workflows, such as summarizing Slack discussions, turning the results into a document, and then preparing a presentation. Sol and Luna become relevant for recurring processes because of their lower API prices and improved prompt caching. The right model tier will depend on whether your priority is low cost or more demanding contextual work.
OpenAI’s company examples indicate the intended role of Astra. At legal software provider Harvey, the model is said to create more structured, context-aware legal drafts, leaving lawyers more time for strategy. This account of Harvey’s use of Astra comes from the provider and is not independently verified in the supplied sources.
According to OpenAI, Airbnb is widening access to Astra and other frontier models for engineering teams working on bug fixes, system design, and faster product delivery. This is an example from professional software work, not evidence that every routine task requires the most expensive model. Cheaper models may be a more economical choice for basic summaries or frequent standardized processes.
At video company invideo, Astra is said to plan edits more precisely, improve color correction and grading threefold, and produce 50 custom effects in one day. Those figures also come from an OpenAI case study about invideo and have not been independently confirmed. Together with the Airbnb example for GPT-6 Astra, it mainly illustrates the intended range across text, engineering, and media work.
Voice access is turning ChatGPT from a conversation tool into an interface for connected services. Sol and Luna can make these workflows substantially cheaper at the API level, while Astra is positioned for more demanding work. The unresolved issues are how reliably the agent performs outside selected demonstrations and how clearly permissions, confirmations, and data use will be presented in everyday use.
Sources
- ChatGPT Voice greift jetzt auf E-Mail, Kalender und Slack zu und erledigt Aufgaben per Sprache – The Decoder, 2026-09-23
- ChatGPT mobile app gets voice-based agentic features – TechCrunch, 2026-09-23
- OpenAI Releases GPT-6 Sol and Luna: 50% Cheaper API Pricing and Benchmarks – MarkTechPost, 2026-09-23
- Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war – Simon Willison’s Weblog, 2026-09-22
- Harvey turns legal context into stronger drafts with GPT-6 Astra – OpenAI, 2026-09-23
- Airbnb widens access to GPT-6 Astra and OpenAI frontier models – OpenAI, 2026-09-23
- How invideo improves color grading 3x with GPT-6 Astra – OpenAI, 2026-09-23
Image: Theo Decker via Pexels