GPT-Live 1 pricing per second vs Sume's per-job audio billing

OpenAI's changelog (Sep 10) says GPT-Live 1 is generally available, with voice sessions at $0.05 per minute billed per second, and backend model and tool usage charged separately. Sume has no live voice session; its audio is two jobs: text to speech billed per character, and speech transcription billed per audio minute, each reserved when you submit.

This page compares billing units only. Sume's rates come from the rate card (TTS $0.0475 per 1,000 characters; STT $0.01 per audio minute), read 2026-09-30.

What is a GPT-Live 1 session billed on?

Time: $0.05 per minute, metered per second while the session runs. The same entry says GPT-Live 1 builds full-duplex voice conversations that can continue while a backend model or agent handles reasoning and tools, and those backend costs are separate.

What are Sume's audio billing units?

TTS 1.0 is priced per transcript character after Sume margin. Spaces and punctuation count, and the maximum is 20000 characters per request. STT 1.0 is priced per audio minute, with 1 minute reserved if you send no duration and a maximum of 10 minutes. Paid generation reserves the estimated amount when the request is accepted, and successful completion captures it (generation admission).

Can I build a voice agent on Sume?

You can chain the jobs: transcribe a recorded turn, produce a reply with your own logic, then synthesize it. That is turn-based, with job latency between turns, not full-duplex audio. For narration use see TTS for video narration.

How should I estimate cost for each?

For a session, multiply minutes by the listed rate and add backend model usage. For Sume jobs, count the characters you will speak and the minutes you will transcribe; both are known before submit. If balance cannot cover the reservation, submit fails with 402 insufficient_credits before provider work starts.