GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

OpenAI Puts GPT-Live-1 in the API at $0.05/Min: 30-Point Benchmark Jump, #1 on Tau3

OpenAI launched GPT-Live-1 in the developer API on Wednesday, priced at $0.05 per minute for the voice layer. The model was introduced in ChatGPT in July but is now available for third-party developers to pair with any backend model or tool harness.

GPT-Live-1 is a full-duplex system: it listens and speaks at the same time rather than operating on a turn-based model. The design change targets a class of conversational bugs — interruption handling, backchannelling, response timing — that prior voice APIs addressed imperfectly.

Benchmark Numbers

OpenAI’s release measured GPT-Live-1 against GPT-Realtime-2.1 on Full Duplex Bench, the standard evaluation for bidirectional voice systems. GPT-Live-1 improves the score by 30 percentage points, with the gains concentrated in two areas: turn-taking latency and interactive behaviour. Those are the dimensions Full Duplex Bench specifically penalises prior architectures on.

On Tau3 — which measures end-to-end voice-agent performance on real tasks rather than individual turn quality — GPT-Live-1 paired with GPT-6 Astra at medium reasoning effort ranks first across tested systems. Tau3 is distinct from Tau2-bench; it focuses on the full voice-to-result loop rather than isolated task completion by a language model.

Architecture and Pricing Model

The model is a front-end layer, not a complete agent. The $0.05/min rate covers the voice input/output handling; backend reasoning and action execution cost separately at whatever model the developer connects. OpenAI’s example pairing — GPT-Live-1 for voice, GPT-6 Astra for reasoning, Codex for code execution — makes the full-stack cost a function of task type.

That split pricing is the practical difference from GPT-Realtime-2.1, which bundled voice and reasoning into one token-based rate. Separating the voice layer lets OpenAI price the front end like infrastructure and the intelligence like an API call.

Context: The Voice API Progression

OpenAI shipped GPT-Realtime-2 in May 2026, GPT-Realtime-2.1 in July 2026, and GPT-Live in the same month. GPT-Live-1 in the API on September 10 is the third distinct voice-layer product in four months. The cadence reflects pressure from ElevenLabs, Hume, and native speech-to-speech evaluations that ranked GPT-Realtime behind newer entrants on naturalness and latency.

GPT-Live-1 is available immediately through the OpenAI API. OpenAI says it will continue releasing newer models into the front-end voice layer as they become available.