GPT-Live-1 API Launch — Voice Pricing & Cost Guide

OpenAI launched GPT-Live-1 in the API on September 10, 2026, bringing the full-duplex voice system behind ChatGPT Voice to developers. A September 11 Hacker News alert points to the same release, not a second launch or price change. The model listens and speaks simultaneously and can delegate deeper reasoning or actions to a backend agent.

The pricing change is structural. GPT-Live-1 charges for the front-end voice session by elapsed time, while the chosen backend model and tools create a second bill. The live tables above pull both layers from our maintained dataset instead of freezing rates in this article.

What changed with GPT-Live-1

GPT-Live-1 replaces the speech-to-text → language model → text-to-speech chain with one voice layer. OpenAI says this reduces brittle handoffs while a backend model works.

The model accepts audio and text and returns audio and text. It natively supplies ASR transcripts and response text, supports keyword biasing and function calling, and can still expose explicit turn boundaries. It does not accept images.

The API release also adds telephony support and a broader selection of preset voices spanning accents, dialects, and languages. Custom voice access requires a sales eligibility process.

The official model page lists streaming and function calling as supported, but not structured outputs, fine-tuning, predicted outputs, images, or video. GPT-Live-1 runs only through v1/live/sessions; it is not served through Responses, Chat Completions, or Realtime endpoints.

GPT-Live-1 pricing impact

The voice layer’s current price appears in the first live table. OpenAI bills session duration per second rather than rounding each session to a full minute. Continuous use therefore scales directly with connected time, but that is only the front-end cost.

Managed Responses delegation uses the selected model’s normal token rate, and tools are charged under their usual rules. Client delegation can instead connect a third-party model or a buyer-controlled agent harness. A routine reservation flow can delegate to GPT-5.6 Luna; a difficult support case can escalate to GPT-6 Astra. The second table and calculator show the maintained OpenAI backend rates.

Telephony, storage, observability, human escalation, taxes, and enterprise commitments can add costs OpenAI’s model table does not include. Budget per resolved call, not per connected minute alone. Compare the wider catalog on our OpenAI pricing page and model backend traffic with the token calculator.

Use this budget structure: connected session duration × the live voice rate above, plus backend tokens, tools, telephony, infrastructure, and human escalation. Because OpenAI bills actual duration per second, short calls are not rounded up to a full minute.

The Free usage tier is not supported. OpenAI documents concurrent-session limits from 25 at Tier 1 to 500 at Tier 5, so production capacity planning needs both a spend model and the applicable concurrency ceiling.

GPT-Live-1 versus GPT-Realtime-2.1

Decision	GPT-Live-1	GPT-Realtime-2.1
Voice behavior	Full-duplex: listens and speaks simultaneously	Realtime speech-to-speech with turn handling
API route	v1/live/sessions	v1/realtime
Billing shape	Connected voice time; backend model and tools separate	Text, audio, and image tokens inside the session
Backend reasoning	Delegates to an OpenAI or third-party agent	Reasoning and tools run through the Realtime model
Image input	Not supported	Supported

OpenAI reports a 30-percentage-point improvement on Full Duplex Bench over GPT-Realtime-2.1. With GPT-6 Astra at medium reasoning, it also reports GPT-Live-1 ranking first on Tau3 end-to-end voice-agent tasks. These are provider-run evaluations; they do not prove the same lift on a buyer’s accents, noise, telephony stack, tools, or policies.

Speak reported almost 80% fewer interruptions than its previous turn-based system. Treat customer evidence as directional, not a universal savings guarantee.

Who benefits—and who should wait

Phone support, reservations, language learning, commerce, and accessibility products benefit when interruption handling affects completion. Teams can also change backend reasoning depth without replacing the voice layer.

Wait if the product mainly needs transcription, one-way narration, image-aware voice, or deterministic batch work. A split pipeline may be cheaper and easier to audit. Compare xAI voice pricing and our earlier OpenAI Realtime launch analysis before choosing an architecture.

What developers should do now

1. Build against v1/live/sessions; do not treat this as a Realtime model-string migration.
2. Measure interruption success, first-response latency, background-noise failures, tool completion, escalation rate, and cost per resolved call.
3. Route routine backend work to a lower-cost model and escalate only the cases that need frontier reasoning.
4. Cap session duration and end abandoned or silent calls promptly.
5. Test real accents, languages, phone codecs, pauses, cross-talk, and noisy environments before increasing traffic.
6. Load-test against your account’s concurrent-session tier before promising production call volume.

For a separate premium voice baseline, benchmark ElevenLabs on the same resolved-call rubric.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored ElevenLabs link at no extra cost to you. It does not affect this analysis.

Labs coverage

GPT-Live-1 is not eligible for the current AI Pricing Guru Labs leaderboard. It has no Chat Completions or Responses inference route, and a text-only proxy would miss full-duplex audio, interruptions, voice quality, telephony latency and backend delegation.

A valid test needs fixed audio prompts, controlled noise and interruptions, identical backends, end-to-end latency, task success, and a complete cost ledger. See Labs coverage notes.

Bottom line

GPT-Live-1 is a new voice architecture, not merely a Realtime refresh. Its strongest proposition is simpler full-duplex interaction with backend model choice; its main budgeting risk is treating the visible session rate as the entire agent cost.

Start with a bounded call flow, keep the backend route explicit, and promote it only when fewer interruptions and better task completion reduce total cost per accepted outcome. The maintained OpenAI API pricing guide covers the rest of the model stack.

Sources: OpenAI’s official GPT-Live-1 launch, model documentation, API pricing, and Live API guide. Facts, endpoints, limits, and pricing verified September 11, 2026 at 10:32 UTC.