Three numbers from a forty-turn voice-agent benchmark are worth more than the rest of it: 17.3×, 2.6 seconds, and 1,216 milliseconds.
Presence ran the same scripted conversation — a codename on turn one, thirty-eight middle turns, a recall question at the end — against Gemini Live, OpenAI Realtime, and its own stack. Then it published everything, including the places it loses.
17.3×
That is how much more the fortieth turn cost than the first on Gemini. Turn one: $0.00215. Turn forty: $0.03714. Context grew from 559 to 13,231 tokens, and every turn was billed for all of it. Zero cached tokens across all forty turns. Seven minutes of audio came to $0.80.
The cost curve is roughly quadratic in turn count rather than linear in minutes, which punishes exactly the behaviour voice products encourage: rapid back-and-forth.
OpenAI has largely fixed this. By turn forty, 3,648 of 3,740 input tokens were cached and per-turn cost was flat, 1.25 times the first turn. But caching is best-effort, and the run shows what that means in practice: it did not engage at all for nine turns, cost climbed to $0.025, then collapsed to $0.0067 when 896 tokens finally cached. Two turns cost about five times the median. The problem is not growth. It is predictability.
Presence is flat by construction, because it bills occupied session time rather than tokens. That is a different unit of account, not a smarter optimization, and the authors say so.
2.6 seconds
Ask a voice agent something hard and it has to go away and think. The interesting question is what happens to the conversation while it does.
The test: while a tool call is pending, speak again — "how long will that take?" — and see whether the agent answers. Presence answers in 2.6 seconds. OpenAI answers in 5.8. Gemini does not answer until the tool returns.
On a twenty-second tool, that is twenty seconds of a session that appears to have died. Two seconds of silence is long; at four you say "hello?"; at ten you assume the line dropped.
This is a product difference, not a latency difference. An agent that can think without abandoning the conversation can look something up mid-sentence, start a slow job and keep talking, or say "I'm still working on that" and mean it. No single-turn benchmark can see it.
1,216 milliseconds
This is the number that nearly became the headline for the wrong reason.
Presence's first three runs said its stack was blocked, exactly like Gemini. The authors were preparing to publish that. The cause was the harness: the follow-up sent exactly 1,200 ms of trailing silence against a 1,200 ms endpointing threshold. Voice activity detection runs in 32 ms frames, so the first boundary at or above the threshold is 1,216 ms. They were one frame short. The turn was never committed, no transcript was published, and from outside it looked identical to a session that had stopped listening. Internally, everything had worked — transcription, model, speech synthesis — and a single missing frame withheld the audio.
A harness bug was indistinguishable from the finding. They published both, and added a rule: an inconclusive follow-up is reported as inconclusive, not as a block.
What survives
Latency numbers will age; Gemini was fastest to first audio, at 1,064 ms p50 against OpenAI's 1,253 and Presence's 1,395. Prices move monthly. The method is the durable part: measure the session, not the turn; test what the agent does while it thinks; and when your instrument disagrees with your expectation, suspect the instrument.
Especially when it is one frame off.