AI voice agent pricing splits between flat rates and metered stacks

AI voice agent pricing is diverging into two broad camps: bundled rates that are easy to forecast and layered billing that can change with every design choice. In material published on Oct. 20, 2026, AssemblyAI said its Voice Agent API charges a flat $4.50 an hour, or $0.075 a minute, billed per second, while public competitor price cards showed lower entry points on hosting alone and higher totals once model, telephony and other usage charges were added.

Evidence and context

AssemblyAI's published guide said AI voice agents "usually lands between a few cents and a few tens of cents per minute," depending on whether the buyer chooses a per-minute platform with pass-through model fees, a per-layer stack, a token-metered speech model or a flat all-in hourly rate. The company said public platform hosting starts at $0.05 a minute before models are added, while all-in platform rates run from about $0.07 to more than $0.30 a minute depending on the setup, with all public prices checked on Oct. 20, 2026.

The same guide listed AssemblyAI's Voice Agent API at $4.50 an hour, billed per second, and said the rate includes speech-to-text, LLM reasoning, voice generation, turn detection, interruption handling, recordings, transcripts, hosting and orchestration. It also said telephony remains a separate cost: on the Voice Agent API, a customer's Twilio account bills carrier minutes and AssemblyAI bills the session, with no markup in between.

AssemblyAI compared that bundled model with other public pricing structures. It cited Vapi at $0.05 a minute for hosting with models billed at cost, Retell at $0.07 to $0.31 a minute depending on components, ElevenLabs Agents with included minutes on plan and $0.08 per additional minute, and OpenAI's gpt-realtime-2.1 at $32 per 1 million audio input tokens and $64 per 1 million audio output tokens. The company also noted that OpenAI's GPT-Live-1 uses a $0.05-a-minute session fee plus model and tool usage.

Those figures are not directly comparable because the inclusions differ. In background material verified on Oct. 20, 2026, Vapi was described as charging $0.05 a minute for hosting with model costs passed through, while Retell's advertised range did not include every possible add-on. Bland was described as bundling language model, speech recognition and speech synthesis into its minute rate but billing telephony separately. ElevenAgents used subscription plans with included minutes and separate charges for extra usage.

AssemblyAI illustrated the difference with a 5-minute call. Under its flat rate, the call costs $0.375 before carrier minutes. Under a per-layer stack, the guide said speech-to-text using Universal-3.6 Pro Realtime at $0.45 an hour would cost $0.0375 for five minutes, and an example LLM calculation came to about $0.009, before voice generation, hosting and telephony. The company said a token-metered model such as OpenAI's realtime offering depends on how the conversation unfolds, making measurement of actual calls more reliable than assumptions.

The broader point in the guide is that the bill moves with agent design. AssemblyAI said longer prompts, more history re-sent each turn, chattier agents, tool calls, caching, concurrency limits, support minimums and silence or hold time can all change the total. It also warned that some platforms add monthly minimums or extra concurrent-call fees, and that connected time can remain billable if a session is left open.

The article's conclusion is operational rather than promotional: flat pricing is easier to forecast, while stacked billing can be cheaper on short or scripted calls but is harder to predict. AssemblyAI said teams should price their own call distribution, then calculate cost per completed task rather than relying on a single meter or headline rate.