Mercury Voice pricing: tokens to dollars per minute of talk

Inception lists Mercury Voice at $0.40 per million input tokens and $1.50 per million output tokens, with a launch promotion at half that ($0.20 and $0.75), and says that works out to roughly $0.009 per minute of conversation. That is the price of the language model slot only. Speech in and speech out are billed separately.

The promotion is a launch discount, and the model is described as generally available for enterprise customers, with access through the Inception API as an OpenAI-compatible endpoint. Sume does not resell Mercury Voice, so the second half of this post is about pricing the speech around it with Sume's STT and TTS rates.

What does Inception's page say exactly?

The numbers below are copied from the Mercury Voice post, read 2026-10-03. The blog index lists the post under Sep 29, 2026, while the post itself states Oct 1, 2026.

How do you turn tokens into a per-minute figure?

Inception's $0.009 is a headline for its own test conversations. Your bill depends on how many tokens each turn sends (system prompt, history, tool output) and how many it returns. The formula is plain; the token counts must come from your own logs. The numbers in the snippet are hypothetical, there only to show the arithmetic at the promotional rates.

IN_PER_M, OUT_PER_M = 0.20, 0.75   # launch promotion, USD per million tokens

def turn_cost(input_tokens, output_tokens):
    return input_tokens * IN_PER_M / 1e6 + output_tokens * OUT_PER_M / 1e6

# Hypothetical turn: 6,000 tokens of prompt and history, 300 tokens answered
per_turn = turn_cost(6000, 300)
print(round(per_turn, 6))           # 0.001425
print(round(per_turn * 6, 5))       # six such turns in one minute

What does the speech around it cost on Sume?

Sume STT bills $0.01 per audio minute (the reservation is one minute unless you send duration_seconds). TTS is character-priced: 38 micro-dollars of list per character, times 1.25, rounded up to cents per job, so about $0.0475 per 1,000 characters. If you assume an agent speaks 900 characters in a minute, which is an assumption and not a measured figure, that is about $0.043 of synthesis before cent rounding. Both are file jobs, so this applies to pre-rendered prompts and call recordings, not to a live turn.

Why can the per-minute figure drift from the headline?

Inception's $0.009 is a stated approximation for a conversation minute, and it depends on a mix of input and output tokens that the page does not publish. Three things move your real figure. First, input tokens grow with every turn because the model re-reads the history, so a long call costs more per minute at the end than at the start. Second, tool calls add their results to the next prompt. Third, the reasoning setting (low, medium or high) changes how much the model writes before it answers, and output tokens cost almost four times as much as input at the listed rates ($1.50 against $0.40 standard, $0.75 against $0.20 promotional).

So the best practice is to log input and output token counts per turn from your first test calls and compute your own per-minute figure with the snippet above. Then look at the 128K context limit: a very long call with a large system prompt can approach it, and trimming history is both a cost and a safety measure.

What should you check before you compare totals?

The related realtime voice API cost per hour post compares all-in-one realtime products, which skip this three-bill arithmetic.

• Count all tokens the call sends, including history resent on every turn.
• Check whether the promotional price has an end date before you build a margin on it.
• Separate the live path (provider-streamed speech) from the recorded path (Sume jobs) when you add up cost.
• Ask Inception for the enterprise terms; the page says to contact sales and does not publish a self-serve signup.