September 15, 2026
GPT-Live-1 takes first place among voice models: 81.5 in the Speech to Speech Index
On September 15, 2026, GPT-Live-1 with Astra took first place in Artificial Analysis' Speech to Speech Index with a score of 81.5. The configuration with Sol scored 80.1 and placed third, while Grok Voice Think Fast 2.0 High came second with 81.3. The OpenAI API lets the model hold a conversation while a backend model reasons separately and calls tools.
Artificial Analysis
@artificialanlys
GPT-Live-1 from @OpenAI, a new full-duplex Speech-to-Speech model, took first place in the Artificial Analysis Speech to Speech Index with a score of 81.5, using Astra as its delegated backend model and surpassing Grok Voice Think Fast 2.0. GPT-Live-1 is a new full-duplex Speech-to-Speech model from @OpenAI. It can delegate reasoning and tool work to a backend text model while continuing the conversation. Developers send audio through the API and receive speech in return. The backend text model is configured separately. We evaluated two backend-model configurations: Astra with medium reasoning effort and Sol with low reasoning effort. Key findings: ➤ Speech to Speech Index: GPT-Live-1 with Astra and medium reasoning effort scored 81.5 and took first place, while GPT-Live-1 with Sol and low reasoning effort scored 80.1 and ranked third. Grok Voice Think Fast 2.0 High sits between them with a score of 81.3. ➤ Speech Agent Arena: GPT-Live-1 with Sol and low reasoning effort ranked third on preference with 1,053 Elo and a 90.9% task-success rate. GPT-Live-1 with Astra and medium reasoning effort ranked fourth with 1,048 Elo and 87.4% task success. Gemini 3.1 Flash Live Minimal leads on preference with 1,096 Elo, while Grok Voice Think Fast 2.0 High leads on task success with 94.6%. ➤ Tau Voice: GPT-Live-1 with Astra and GPT-Live-1 with Sol took the top two places in our agent-performance benchmark with scores of 67.9% and 59.3% respectively, ahead of Grok Voice Think Fast 2.0 High at 56.5%. ➤ Big Bench Audio: GPT-Live-1 with Astra scored 90.1%, while GPT-Live-1 with Sol scored 89.0% on audio-reasoning tasks. They trailed Grok Voice Think Fast 2.0 High at 97.2% and Qwen Audio 3.0 Realtime Plus at 99.2%. ➤ Speed: average time to first audio on Big Bench Audio was 1.34 seconds for GPT-Live-1 with Astra and 1.24 seconds for GPT-Live-1 with Sol. Grok Voice Think Fast 2.0 High recorded 0.70 seconds. ➤ Price: GPT-Live-1 with Astra costs $5.83 per hour of input audio, while GPT-Live-1 with Sol costs $4.47 including delegated backend-model usage. Grok Voice Think Fast 2.0 High cost $4.80 in our fixed pricing set for Big Bench Audio. More below ⬇️
· 78.7K views
GPT-Live-1 became generally available in the OpenAI API on September 10, 2026. According to OpenAI, it improved its Full Duplex Bench result by 30 percentage points over GPT-Realtime-2.1. The model carries on a two-way conversation while a backend model or agent reasons and calls tools in parallel.
More than one ranking. In Speech Agent Arena, Gemini 3.1 Flash Live Minimal led on preference with 1,096 Elo. GPT-Live-1 with Sol placed third with 1,053 Elo and 90.9% task success, while the Astra configuration ranked fourth with 1,048 Elo and 87.4%.
On Tau Voice, Astra and Sol took the top two spots with 67.9% and 59.3%. On Big Bench Audio, Astra scored 90.1%, while Sol scored 89.0%.
The cost of conversation. A voice session costs $0.05 per minute and is billed by the second without rounding. The backend model and tools are charged separately, there is no free tier, and the concurrent-session limit is 25 for Tier 1, 50 for Tier 2, and 500 for Tier 5.
Going forward, GPT-Live-1's place in the voice stack will depend on the backend model selected: Astra delivers the highest index, while Sol lowers cost and latency.