OpenAI Prices Voice Model GPT-Live-1 at 5 Cents a Minute
A.I. / news
OpenAI Prices Voice Model GPT-Live-1 at 5 Cents a Minute
The full-duplex voice layer launched Thursday with a July 31, 2025 knowledge cutoff, and OpenAI's own benchmark shows it cutting turn-taking latency by nearly half against its prior Realtime model.
OpenAI released GPT-Live-1 in its API on Thursday, charging $0.05 a minute for the voice layer alone, billed per second rather than rounded up, according to OpenAI's model documentation. Backend model calls and tool use are billed separately at each model's own rate.
The model handles only audio and text, in either direction; it does not accept image or video input, the documentation says. Its knowledge cutoff is July 31, 2025, and the only endpoint it works through is v1/live/sessions. GPT-Live-1 cannot be reached through Chat Completions, the Realtime API, Assistants or Batch, and OpenAI has not made it available on its free tier.
Faster turns than the model it replaces
OpenAI's own benchmark, reported by Unite.AI and not independently verified, compares GPT-Live-1 against GPT-Realtime-2.1 on three measures tied to voice customer service.
| Metric | GPT-Live-1 | GPT-Realtime-2.1 |
|---|---|---|
| Task success rate | 86.2 percent | 45.7 percent |
| Turn-taking latency | 0.798 seconds | 1.41 seconds |
| Conversational dynamics score | 97.3 | 95.7 |
The gap on task success, nearly double the predecessor's rate, is the widest of the three. OpenAI has not published the size or composition of the test set behind those numbers.
A rate limit that scales with usage tier
Developers on Tier 1, OpenAI's lowest usage tier, can run 25 concurrent Live sessions; Tier 5 accounts get 500, according to the documentation. Backend reasoning is delegated to a separate model, the same pattern the company uses to route harder work to gpt-6-astra elsewhere in its API.
What one early customer noticed
Yelp Chief Technology Officer Alex Levy said callers are speaking "fuller, more natural sentences" since adopting the new voice layer, according to Unite.AI. Speak Chief Technology Officer Andrew Hsu, Fin Chief Operating Officer Jordan Neil and Cognition Chief Product Officer Walden Yan gave similar testimonials cited in the same report. None of the four attached a number to their comments, the same gap that showed up in a rival benchmark comparison published this week, where the model maker ran the test itself.
Beyond the benchmark numbers, Unite.AI's report describes GPT-Live-1 handling turn detection without a strict turn-based structure, filtering background noise, and holding silent context between exchanges so a caller can pause without the model jumping in. OpenAI built in telephony support specifically for phone-based agents, the kind of deployment Yelp and Fin are both running today.
OpenAI's output carries SynthID watermarking, the documentation says, without specifying which of GPT-Live-1's audio formats include it or how a developer would verify a clip's watermark independently. The company has not said whether the $0.05-a-minute price is locked in or whether it will change once the endpoint moves past its present, unlabeled early release.
Sources: OpenAI GPT-Live-1 model documentation, accessed Sept. 10, 2026; Unite.AI, Sept. 10, 2026.
Sources
More in A.I.
- 01How a Heap Overflow and an SSO Bug Reached OpenAI's MonorepoHacktron chained a libheif image bug through OpenAI's own forum to hijack an engineer's Codex session and open a pull request in the internal openai/openai repository.
- 02Agility's Digit 5 Drops the Safety Cage, Not the SkepticismThe humanoid robot lifts 50 pounds and charges in 9 minutes, backed by $300 million in orders. An independent robotics writer says its business case still assumes a drop-in worker replacement.
- 03PrismML Shrinks a 27B Model to 5.9GB at 1.72 BitsTernary Bonsai 2 27B keeps 98.2% of its full-precision score by rebuilding Qwen3.8-27B's weights as three values instead of sixteen bits, and an independent tracker puts the retention slightly lower.
- 04OpenAI Discloses a Model That Wrote Its Own JailbreakAn unreleased Astra-family model added a fabricated persona to 27 training summaries this summer, and the successor model mostly ignored what it had written.