benchr
GPT-Live-1 mini arrived July 8 as part of the ChatGPT Voice rollout. It is easy to read the name as a cheaper developer model. OpenAI's announcement does not support that reading yet.
For now, mini is a ChatGPT tier
OpenAI places GPT-Live-1 mini in the Voice rollout for Free users, while GPT-Live-1 is the default for eligible paid plans. That describes the experience a ChatGPT user may receive. It does not give a developer an endpoint, throughput promise, or billing unit.
The name contains no price
The announcement lists no context window, maximum output, rate card, or API model string for GPT-Live-1 mini. Do not borrow a rate from GPT-Realtime or another OpenAI mini family. They are separate products with separate terms.
The Voice response may include a handoff
OpenAI says the Live models can delegate deeper work to other OpenAI models. What the user hears may be the result of a larger system, not one model acting alone. If you need a buildable voice agent, write down the endpoint, billing, tool, privacy, and fallback requirements first.
| Field | Verified record |
|---|---|
| Audience | ChatGPT Voice Free users |
| Interaction | Full duplex |
| API availability | Planned, not announced |
| Public rate / limits | Not published |
The honest comparison is access, not capability
The announcement distinguishes the two versions by ChatGPT audience: mini for Free users and GPT-Live-1 for Go, Plus, and Pro. It does not publish a matched capability table, latency figure, or developer price. Any claim that mini is a fixed percentage faster, weaker, or cheaper would go beyond the source.
| Question | GPT-Live-1 mini | GPT-Live-1 |
|---|---|---|
| ChatGPT audience | Free users | Go, Plus, and Pro users |
| Continuous voice | Yes, in the described rollout | Yes, in the described rollout |
| Public API contract | Not announced | Not announced |
| Public API rate | Not published | Not published |
Use the free rollout for user research
Ask participants to complete the same tasks across quiet and noisy environments, short and long pauses, interruptions, and multiple languages. Record task completion and conversational repair rather than asking whether the voice “sounds good.” This can reveal product needs now, but it still does not estimate an API service that has not been documented.
Design the fallback before the endpoint
A production voice agent needs a plan for dropped audio, tool failure, unsafe output, and transfer to a human. The launch also describes delegation to another model for deeper work, so the future integration may expose more than a single inference call. Keep the architecture provisional until the public API shows where those boundaries sit.
Compare cohorts without pretending they are model benchmarks
If the rollout gives a research team access to both tiers, use matched tasks and comparable devices, networks, languages, and environments. Record which ChatGPT plan and product version each participant used. A difference in the observed experience may come from the tier, a background model, a product feature, or rollout timing; do not publish it as an isolated mini-versus-standard model score without a controlled interface.
Focus the free-user study on accessibility and completion: can someone interrupt naturally, recover from a misunderstanding, and finish the task without looking at the screen? Add a quiet-room run and a realistic-noise run. Document where visual cards or text become necessary, because a voice-first workflow that repeatedly demands the display may not solve the intended problem.
Keep procurement separate. Free access inside ChatGPT says nothing about a future API's concurrency, service terms, data controls, or bill. The useful output of this study is a prioritized voice requirement list and a set of versioned conversations to replay when developer access arrives—not a guessed per-minute price.