The Pulse
Inception Launches Mercury Voice With a 320ms Median Response
Inception made Mercury Voice generally available to enterprise customers on September 29, reporting a 320-millisecond median time to first answer token on customer-service prompts. The company says the diffusion language model costs about $
AI.info Team · Written with AI ·
Voice-agent builders have had to weigh quick replies against models that can reason through complex requests. Inception says Mercury Voice narrows that gap, but its launch figures show a more specific target than a blanket sub-300-millisecond guarantee: the model returned its first answer token in a median 320 milliseconds, with a 750-millisecond 95th-percentile result. The company announced general availability for enterprise customers on September 29.
Mercury Voice reports 320ms at the median
Inception measured Mercury Voice on real customer-service prompts and reports a 320-millisecond median time to first answer token, or TTFAT. The figure measures the wait until the model begins producing the answer a caller hears, after its reasoning is complete. Inception says the 95th-percentile latency was 750 milliseconds, meaning the slower end of responses matters alongside the median.
The company argues that roughly 500 milliseconds is the point beyond which a voice agent’s pause feels awkward. At its low reasoning setting, Inception says Mercury Voice is the only model it tested with a median TTFAT below that mark. Its 750-millisecond p95, however, exceeds the stated target, a reminder that a median does not describe every turn.
The launch follows a September 8 preview that advertised Mercury Voice time-to-first-token below 170 milliseconds. That earlier figure and the new 320-millisecond result use different measures: the latest announcement focuses on the first answer token after reasoning, rather than simply the first token. The distinction matters for voice systems, where a token that is not yet part of the spoken answer may not end the caller’s wait.
Inception’s benchmark claims come with a narrow scope
Inception says Mercury Voice runs on a diffusion language model, which generates tokens in parallel rather than one at a time. The company reports that Mercury Voice was 5.9 times faster than GPT-6 Luna without reasoning on its latency comparison, and says it outperformed models including Gemma 4 31B, GLM-5.3-Flash and Qwen3.5-397B on a composite of agentic and conversational benchmarks.
That composite combines τ³-bench tests for telecom, retail and airline tasks with IFBench and BFCL v4. The results are Inception’s own benchmark claims, not an independent evaluation described in the launch post. The company also reports that Mercury Voice was twice as fast as the next-fastest model in its quality-and-latency comparison.
Enterprise access, with launch pricing cut in half
Mercury Voice is available to enterprise customers through Inception’s API as an OpenAI-compatible endpoint. The company says it can fit into existing voice-agent stacks, including LiveKit, Pipecat, Vapi and Retell; customers must contact its sales team for access.
Standard pricing is $0.40 per million input tokens and $1.50 per million output tokens. Inception is offering launch pricing at half those rates: $0.20 per million input tokens and $0.75 per million output tokens. Using its typical voice-agent profile, the company estimates a cost of about $0.009 per minute, compared with $0.045 per minute for GPT-4.1.
Early users put the speed claim to work
Inception identifies Audivi AI, which uses the model for automated drive-through ordering, and Altur, which builds voice agents for financial institutions, as users. The company says Altur’s internal benchmarks found Mercury Voice to be the fastest model it tested. Pedro Fernández, Altur’s co-founder and CTO, said: “It's the first time the language model isn't the bottleneck in our voice pipeline.”
OpenCall, a company that builds AI phone agents, also appears in the announcement. Inception says Mercury brought median model response latency close to 170 milliseconds on OpenCall’s production workload. That customer-specific result is separate from Mercury Voice’s 320-millisecond median on the launch post’s customer-service test set.