voice ai · what a conversation actually costs

They bill you by the minute.
We don't.

Every frontier voice API meters audio in and audio out, per token, on rented GPU time. The longer your users talk, the more you pay. Same conversation, published rates, as of August 2026.

Audio, per million tokens

ProviderAudio inAudio outPer minutePer hourvs HawkTalk
HawkTalk WebSockets$1.00$1.00$0.0015$0.09
Gemini 3.1 Flash Live$3.00$12.00$0.0113$0.687.5×
Gemini 2.5 Flash Native Audio$3.00$12.00$0.0113$0.687.5×
Gemini 3.5 Live Translate$3.50$21.00$0.0184$1.1012.2×
GPT Realtime 2.1 mini$10.00$20.00$0.0225$1.3515.0×
GPT Realtime 2.1$32.00$64.00$0.0720$4.3248.0×

Per-minute figures use Google's own conversion of 25 audio tokens per second, at a 50/50 split between the user talking and the model answering.

What a year of talking costs

Provider1 hr/day, 1 user1 hr/day, 10 users1 hr/day, 100 users
HawkTalk — flat tier$120 – $3,000$120 – $3,000$120 – $3,000
Gemini 3.1 Flash Live$246$2,464$24,637
GPT Realtime 2.1 mini$493$4,927$49,275
GPT Realtime 2.1$1,577$15,768$157,680
The row that matters is the first one. Ours does not move. Theirs multiplies by every user you add and every minute they talk — which means the more successful your product is, the worse their pricing gets for you.

The tiers

$10/mo

Your AI

One person, one assistant. Voice, tools and text on one key, on our silicon.

$50/mo

You and the people you deal with

Better models, real concurrency. For an app with users, a practice, a small team.

$250/mo

Build on it

You are shipping your own AI products and companies, and HawkTalk is what they run on. The full family, no caps.

not in the table above

HawkTalkLive

The comparison stops at WebSockets because that is where their products stop. GPT Realtime and Gemini Live give you one model, one stream, one thing at a time. HawkTalkLive runs speaking, thinking, acting and logging as independent lanes at the same time, and lets you re-point a live call at a different brain without dropping it.

It costs more than the WebSocket tier. There is nothing on the other side of the table to compare it to.

Rates checked August 2026 from OpenAI's and Google's published pricing pages. Providers change prices; the structural difference does not — theirs scales with usage, ours does not.