OpenAI has revised the pricing for the ChatGPT Voice in Desktop voice mode, through which users control ChatGPT Work and Codex agents in the desktop client: voice time now costs $0.05 per minute, and for Business, Edu, and Enterprise plans — 1.25 credits per minute. For the same amount of credits, users get approximately 2.4 times more voice minutes, which the company calls an approximate 60% price reduction.

image
image
image

What happened

According to the official Codex pricing page, the rate for the Voice in Desktop voice mode is set at $0.05 per minute, and for Business, Edu, and Enterprise with credit billing — 1.25 credits per minute. The update was announced on September 14, 2026, via an approved repost from the OpenAI Developers account: for the same amount of credits, approximately 2.4 times more voice minutes are available, while the previous rate was not officially published. Billing is divided into two parts: GPT-Live voice dialogue minutes are billed separately, and the compute resources for the agent to perform the task itself are paid for at the standard token rates for Work and Codex from the general budget.

Context

This is a pricing decision, not a technical result: no new model, architecture changes, or benchmarks are mentioned in the sources. Previously, voice for agentic scenarios remained an expensive input channel, and the new scheme fixes a cost structure in which voice input is assessed as a separate and inherently cheaper resource than the compute for task execution. The limitations are significant: the change only affects the desktop application on macOS and Windows, does not affect the regular voice mode, and voice control is not available via the API.

Why this matters for the industry

This is the first notable price reduction for a voice interface specifically for agentic scenarios: OpenAI is lowering the barrier to controlling multiple parallel Work and Codex agents by voice, and dictating tasks begins to compete with typing long prompts. The split scheme of “voice channel separately, task tokens separately” sets a measurable expense accounting model that competitors like Anthropic and Google may copy for their own agentic products if the trend persists. The rate itself is a weak but readable market indicator: either the inference of the voice stack has become cheaper, or OpenAI is consciously subsidizing long voice sessions, stimulating voice as an orchestration channel.

Why this matters for users

For those who use ChatGPT Work or Codex on macOS or Windows, the effect is immediate: without migrations or new integrations, the same credits cover approximately 2.4 times more voice minutes. A separate predictable line item for voice appears in the budget, so a voice briefing becomes a convenient default input for long tasks that were previously easier to describe in text. It is important to remember: a task launched by voice still consumes the general token budget of Work and Codex at standard rates, so the actual cost of a session is limited by it.

What is still unknown / limitations

The figures “approximately 60%” and “2.4 times more minutes” are OpenAI’s own characterization from the OpenAI Developers repost, and the previous rate was not officially published, so the calculation cannot be independently verified. The sources contain no data on recognition quality, latency, interruption handling, or the success rate of agentic tasks launched by voice, so there is currently no basis for comparing the price of voice with the price of long prompts in terms of practical effectiveness. A possible competitor reaction with similar rates is an interpretation, not a confirmed plan, and conclusions about a long-term trend cannot be drawn from a single pricing change.

Sources

Author

Look at AI, editorial team