How GPT-Live-1 voice billing actually splits
two invoices, not one bundle
The voice session is priced per minute. Backend reasoning is billed on tokens. Treat $0.05/min as an all-in phone-call price and the bill jumps the first time the session delegates to Astra or another model. QCode’s 30-day usage table has no gpt-live-1 row, so this page does not claim that voice layer is callable here.
Updated 2026-09-11
Split the bill into four facts
Per voice minute
The model page says: Voice sessions cost $0.05 per minute, billed per second. That is listen-and-speak, not the whole call.
Not rounded to the next minute
Same page: Session duration is not rounded up to the next whole minute. Cost in seconds, not “1m1s counts as two minutes”.
Reasoning and tools are a second invoice
Backend model and tool usage is billed separately. When Live-1 hands a hard turn to Astra / Luna / a third party, that Responses call uses the delegated model’s list price.
No gpt-live-1 usage row
QCode’s 30-day table has no gpt-live-1 line. gpt-realtime* rows are all 0. Category is default: we do not say this model is already on sale here.
How to split this class of bill
Full-duplex voice APIs fail people who treat “the layer that talks” and “the layer that thinks” as one price list. Order of operations stays the same: read the voice-layer rate and rounding rule on the official model page, list which turns will delegate, then multiply estimated tokens by the backend’s input/output rate. QCode can help with the backend invoice — one key for gpt-5.6-sol / gpt-6-astra. The voice id has no usage row, so it is not an in-sale claim.
What the official pages said on 2026-09-10
On 2026-09-10 the OpenAI model page for gpt-live-1 listed $0.05/min, billed per second, no whole-minute round-up, with backend model and tools billed separately. The developer forum post the same day described full-duplex ChatGPT-style voice in the API and delegation to a backend of your choice. The New Stack that day wrote the two-invoice version: pay the voice layer, then pay the Astra call it hands off.
Timeline
2026-09-10: gpt-live-1 appears on the developer model page. Pricing sentence: Voice sessions cost $0.05 per minute, billed per second. Backend model and tool usage is billed separately.
2026-09-10 forum announcement: full duplex plus backend delegation. Paired with GPT-6 Astra at medium reasoning, Tau3 first-attempt 83.6% versus 45.7% for GPT-Realtime-2.1. That is their eval sentence, not a QCode benchmark.
2026-09-11: the same model page still has the rate and the “backend billed separately” line. QCode CRS that day has no gpt-live-1 row. The openai.com product post is Cloudflare 403; facts are taken from developers.openai.com.
Confirmed versus easy to mis-add
On the model page (2026-09-10)
As of 2026-09-10 you can verify: model id gpt-live-1; voice sessions $0.05/min billed per second; duration not rounded up to the next whole minute; backend Responses calls use normal pricing for the configured model and tools.
What people say that the docs do not bundle
Around 2026-09-10 some posts treat $0.05/min as “three dollars an hour for a whole voice agent”. The model page does not say that. The New Stack writes about $3/hour for the voice layer, then Astra on top. Unconfirmed: when any reseller will list gpt-live-1, and how concurrent-session tiers map. QCode has not published that id.
Two columns, do not mix them
Voice layer only
Use this to estimate listen-and-speak cost: seconds × $0.05/60. Do not use it as a project budget — one delegation blows the column.
Voice layer plus backend
Use this for a tool-using voice agent. If the backend is gpt-5.6-sol or gpt-6-astra, add that model’s token price. Official GPT-5.6 Sol list is $4 / $20 (promo through at least 2026-11-21).
Three steps after a surprise invoice
First: open the official gpt-live-1 page and confirm the voice layer is $0.05/min per second, not a subscription. Second: list which turns delegate; those turns are token-priced on the backend model. Third: do not probe the voice id. Send a tiny text request, confirm the key and the echoed model. If your gateway has no gpt-live-1, do not hard-code it.
What to do on QCode, concretely
QCode’s 30-day table has no gpt-live-1, so this page does not claim the voice layer is callable. Wire the backend invoice first: copy the dashboard key after signup. Claude Code: ANTHROPIC_BASE_URL = https://api.qcode.cc and ANTHROPIC_AUTH_TOKEN set to that key. Codex CLI: point the OpenAI-compatible base URL at the same host. Probe with gpt-5.6-sol (official $4 / $20, promo through at least 2026-11-21), then switch to gpt-6-astra by task. Next step: one non-streaming tiny request, HTTP 200, matching model id in the body.
FAQ
Does $0.05/min include backend reasoning?
No. The model page splits voice sessions from backend model and tool usage. A delegated Astra (or other) call is billed at that model’s rate.
How is a partial minute billed?
Per second. The docs say duration is not rounded up to the next whole minute. 61 seconds is 61 seconds, not two minutes.
Can I call gpt-live-1 on QCode today?
We do not claim that. There is no usage row. A page that says “already available on QCode” would be false. Trust /models.
Why read this page then?
So the two-invoice math and the backend path are correct before any gateway lists the voice id. The math does not depend on QCode listing gpt-live-1.
Sol or Astra on the backend?
The docs let you choose. Sol’s official list is $4 / $20 (promo through at least 2026-11-21) for a pathfinder; heavier turns can switch to gpt-6-astra. Both have real QCode usage.
Is this the old Realtime price list?
No. gpt-live-1’s voice layer is per minute. Realtime ids are separate rows in the usage table and recently sit at 0. Do not paste old Realtime token rates onto Live-1 minutes.
Sources
Rate, per-second billing, backend separate, model id: developers.openai.com/api/docs/models/gpt-live-1 (fetched 2026-09-11). Full duplex and delegation: OpenAI developer forum 2026-09-10. Two-invoice write-up: The New Stack 2026-09-10. Sol promo window: official pricing page, “at least through November 21, 2026”. openai.com product post is 403 and is not the deciding source.
Voice layer not on sale here — wire the backend first
One QCode key for gpt-5.6-sol and gpt-6-astra. No gpt-live-1 claim. Plans from ¥60/month.
Related
Why GPT-6 Astra burns quota so fast
When you delegate to Astra, credits multipliers sit on top of the voice-layer invoice.
GPT-5.6 Sol versus Astra
Two backend tiers, official prices, so you do not merge voice and reasoning into one SKU.
OpenAI API outage checklist
On 5xx from voice or images, read the status page before you rewrite clients.
This page is a billing split, not an OpenAI statement, and not a claim that QCode lists gpt-live-1. The model page and /models win.