GPT-Live-1 tops voice benchmark by splitting speech from reasoning

OpenAI’s GPT-Live-1 has reached first place on Artificial Analysis’s Speech-to-Speech Index, scoring 81.5 and edging Grok Voice Think Fast 2.0 High by 0.2 points. Its architecture separates real-time conversation from deeper reasoning: the voice model manages speech, pauses, and interruptions while a backend text model handles complex analysis and tool calls.

The hybrid design gives developers a configurable voice layer without requiring them to build a chained speech-to-text, language-model, and text-to-speech pipeline. Benchmark results show strong agentic performance and conversational timing, alongside weaker raw audio reasoning and slower response starts.

The voice layer owns the floor

GPT-Live-1 is a full-duplex model, which means it can receive audio while producing speech. That concurrent processing helps it recognize interruptions, backchannels such as “mhmm,” and pauses that should not end a turn. A developer-managed harness routes tasks requiring multi-step reasoning or tools to a separate text model.

1. The application streams a user’s audio to GPT-Live-1 through the API.
2. The voice layer tracks incoming and outgoing audio to manage turn-taking and interruptions.
3. The application delegates reasoning and tool calls to an OpenAI or third-party text model.
4. GPT-Live-1 delivers the resulting response while continuing to monitor the conversation.

This structure reduces the brittle handoffs common in chained voice systems. Developers can also select a backend according to their requirements for reasoning quality, latency, and token cost. OpenAI’s model documentation lists 12 available voices.

Artificial Analysis evaluated two backend configurations: Astra at medium reasoning effort and Sol at low reasoning effort. Astra produced stronger results on tool-oriented customer-service tasks, while Sol responded faster, cost less, and scored higher on conversational dynamics.

Agentic tasks create the lead

Artificial Analysis publishes a composite index alongside several component benchmarks described in its benchmark methodology. The tests cover customer-service workflows, conversational timing, general task completion, and reasoning from spoken questions.

Artificial Analysis benchmark results

Benchmark	Astra, medium	Sol, low	Leading comparison	What it measures
Speech-to-Speech Index	81.5, first	80.1, third	Grok Voice Think Fast 2.0 High: 81.3	Composite performance
Tau-Voice	67.9%	59.3%	Grok Voice Think Fast 2.0 High: 56.5%	Agentic customer-service tasks
Full Duplex Bench	94.9%	97.3%, second overall	Qwen Audio 3.0 Realtime Plus ranked first	Pauses, interruptions, and backchannels
Speech Agent Arena	87.4%	90.9%	Grok: 94.6%; GPT-Realtime-2.1 High: 91.5%	General task success
Big Bench Audio	90.1%	89.0%	Qwen: 99.2%; Grok: 97.2%	Reasoning from spoken questions

Astra’s largest advantage appears on Tau-Voice, which uses simulated airline, retail, and telecom support scenarios. Its 67.9% score leads Grok by 11.4 percentage points, indicating that the stronger backend improves workflows involving tools and multiple reasoning steps.

Sol’s 97.3% Full Duplex Bench result places it second behind Qwen Audio 3.0 Realtime Plus. That score reflects smoother handling of interruptions, pauses, and brief acknowledgements. Both GPT-Live-1 configurations trail Grok on Speech Agent Arena, showing that the composite lead does not extend across every task category.

Reasoning delay shows up

Big Bench Audio converts difficult text reasoning questions into spoken prompts. GPT-Live-1 scores 90.1% with Astra and 89.0% with Sol, behind Grok at 97.2% and Qwen at 99.2%. Applications centered on reasoning directly from audio may therefore receive stronger benchmark results from those alternatives.

Delegation also adds delay when a response depends on the backend model. Average time to first audio was 1.34 seconds for Astra and 1.24 seconds for Sol, compared with 0.70 seconds for Grok Voice Think Fast 2.0 High. That gap can affect assistants expected to respond immediately after each turn.

Backend choice sets the bill

Artificial Analysis estimated hourly input-audio costs using a fixed 40-question subset of Big Bench Audio. Its GPT-Live-1 figures include the delegated backend tokens, making them broader than the voice model’s per-minute API price.

Evaluation cost and response latency

Configuration	Cost per input-audio hour	Time to first audio
GPT-Live-1 with Astra, medium	$5.83	1.34 seconds
GPT-Live-1 with Sol, low	$4.47	1.24 seconds
Grok Voice Think Fast 2.0 High	$4.80	0.70 seconds
GPT-Realtime-2.1 High	$10.75	Not reported here

OpenAI lists the GPT-Live-1 voice layer at $0.05 per minute through the API. Backend model tokens, orchestration infrastructure, tool calls, and any telephony services add to the production bill. ChatGPT bundles access into Go, Plus, and Pro subscriptions without a separate per-minute meter.

Match the model to the workload

Tool-heavy support agents have the strongest benchmark case for GPT-Live-1 with Astra, particularly when workflows resemble Tau-Voice’s airline, retail, and telecom scenarios. Sol offers a lower evaluated cost, a faster response start, and stronger conversational timing, with lower performance on Tau-Voice and raw audio reasoning.

Language-learning platform Speak reported that early testing reduced false interruptions during learners’ thinking pauses by nearly 80% compared with traditional turn-based systems. The result covers one customer implementation, but it illustrates where full-duplex turn handling can improve tutoring and coaching applications.

Production evaluations should also cover conditions outside these published scores, including background noise, accents, multilingual speech, telephony compression, tool reliability, and sustained concurrent sessions. The leaderboard establishes GPT-Live-1 as the composite leader for the tested configurations; deployment results will depend on the selected backend, orchestration code, and traffic profile.