Pipeline or speech-to-speech? Run the numbers.

It depends on the model. At September 2026 list prices, a pipeline (Deepgram Nova-3 at $0.0058 per minute, a cached cheap LLM, Cartesia Sonic at $0.03 per minute of speech) costs about $0.02 per minute of call, whatever the prompt size. GPT-Realtime-2.1-mini lands in the same range. Gemini 3.1 Flash Live starts higher and grows with the prompt size and the call length because it re-bills the whole context on every turn. GPT-Live-1 is a flat $0.05 per minute plus backend tokens.

The Gemini Live API has no prompt caching: on every dialogue turn it bills the system prompt, the tool definitions and all the audio exchanged so far as input tokens. A 3,000-token prompt on a 2-minute call with 11 turns is billed 11 times, 33,000 tokens, before any audio. A pipeline sends the same prompt to its LLM on every turn too, but prompt caching bills the repeated prefix at 10% of the input price.

Speech-to-speech models keep the whole conversation in context. Each new turn re-reads the audio history, so the cost of a turn grows with the minutes already spoken: a long call costs more per minute than a short one. Gemini Live's context window compression caps that history with a sliding window. A pipeline is flat: STT is billed on the call duration, TTS on the words the agent says, and the transcript history is small text that caching makes almost free.

A voice agent sends a new request to its language model on every dialogue turn, and each request carries the whole context again: the system prompt, the tool definitions, the transcript so far and the new utterance. On an 11-turn call a 3,000-token prompt is sent 11 times. Prompt caching means the provider recognises that the beginning of the request is identical to the previous one and bills that repeated part at a fraction of the price: 10% on OpenAI and Gemini text models, about 3% on the Realtime audio history. It works automatically when the stable content (prompt, tools) comes first and the changing content (date, caller data, new turn) last. The Gemini Live API has no cache, which is why its cost climbs with the prompt size. The simulator applies each provider's native caching as is: on for the pipeline LLMs, GPT-Realtime-mini and the GPT-Live-1 backend, never for Gemini Live.

Paste it into the simulator: it counts about 4 characters per token, in your browser, without sending anything. Gemini and OpenAI tokenizers are within about 10% of each other on prose. Each tool definition adds roughly 180 tokens (name, description, parameter schema) that ride along on every turn.

The speech-to-speech one, for the feel: full-duplex listening, natural prosody, interruptions and backchannels handled by the model, and no STT to LLM to TTS hand-offs. The trade-off is control. A speech-to-speech model behaves like a fixed temperature of 1 and offers little structured output, while a pipeline gives you temperature 0, deterministic tool calls and the freedom to swap any component or voice.

GPT-Live-1 is OpenAI's full-duplex speech-to-speech model released in September 2026. The voice layer costs a flat $0.05 per minute billed per second, silences included, and it delegates reasoning and tool calls to a backend text model (GPT-5.6 Luna, Terra, Sol or GPT-6 Astra) billed per token with prompt caching. It ranks first on the Artificial Analysis Speech to Speech Index with Astra as backend, with about 1.3 seconds to first audio.

Gemini 3.8 Live is billed like Gemini 3.1 Flash Live: no prompt caching, the prompt, tools and audio history re-billed on every turn, at about 44% less per token. The ratio comes from Google DeepMind's model evaluation, which reports the Artificial Analysis cost per hour of input audio on the Big Bench Audio subset: $0.84 per hour for Gemini 3.8 Live against $1.50 for Gemini 3.1 Flash Live (Minimal) and $1.75 for Gemini 3.1 Flash Live (High). The simulator applies that 0.56 ratio to every Gemini 3.1 Flash Live token rate. The same document gives Gemini 3.8 Live a 76.0 score on the Artificial Analysis Speech to Speech Index (82.6 with extended thinking) against 71.5 for Gemini 3.1 Flash Live (High), and the top experience score on ServiceNow's EVA-Bench. In our listening it is slightly disappointing on voice and feel next to Gemini 3.1 Flash Live, whose voice family leads the Artificial Analysis Pronunciation Robustness benchmark, so the simulator rates its feel and voice below 3.1's. No separate latency figure is published, so the response-time chain of Gemini 3.1 Flash Live is reused. Source: Gemini 3.8 Audio (Live, Live Extended Thinking) model evaluation, Google DeepMind, September 2026.

Grok Voice Think Fast 2.0, xAI's speech-to-speech model launched on 29 July 2026, bills $0.08 per minute of session audio (up from $0.05 for version 1.0), plus $0.004 per text item injected into the conversation; function call outputs are not billed. Nothing depends on the prompt size, the history or the number of turns, so the simulator prices it as a flat line: it is never the cheapest on a short call with a compact prompt, and it wins once minutes or tokens grow. Artificial Analysis measures 0.70 s average time to first audio, the only top-5 model under a second, and an 82.9 Speech to Speech Index for the High reasoning variant, which sets its control score. No EU data residency option is published. Source: Artificial Analysis, Grok Voice Think Fast 2.0 on the Speech to Speech Index, July 2026.

Each model and each pipeline component carries a voice score, 1 to 5, for English, French, Spanish and Dutch: reputation from blind listening benchmarks, vendor language lists and what we hear in production, not a measurement, and editable in the code. Gemini 3.1 Flash Live has the best voice and feel of the field: its voice family, Gemini 3.1 Flash TTS, leads the Artificial Analysis Pronunciation Robustness benchmark at 88.1% (ahead of SpaceXAI TTS at 87.6% and ElevenLabs Eleven v3 at 85.6%), the test of whether names, amounts, dates, codes and emails are said correctly. It scores 5 in every language and the top feel, and every other option, Gemini 3.8 Live included, sits half a point below its reputation on both. In French, Gradium, ElevenLabs and Cartesia are the strong pipeline voices; OpenAI's voices keep an accent; Grok has no French reputation yet. In Spanish, Gradium and ElevenLabs lead the pipeline, Gemini 3.8 Live and Grok's regional voices follow. In Dutch and Flemish, ElevenLabs has native Flemish voices, while Cartesia and Deepgram support it with an accent and Gradium does not ship it. The score moves the feel (a 5 earns half a dot, a 3 loses a full dot, a 2 loses two: a bad voice in the caller's language is a bad experience whatever the model does elsewhere), an option scoring under 3 is not proposed as best value while a better-sounding one fits the budget, the best-value pick follows, and the pipeline selects show the score of every STT and TTS in that language. Belgium is covered by French and Dutch. Source: Artificial Analysis Pronunciation Robustness benchmark, September 2026.

Yes, with all of them except Grok Voice 2.0, which has no published EU option, at different prices. Gemini 3.1 Flash Live and the Gemini text models run in an EU region through Vertex AI at regional pricing, 10% above the global endpoint. OpenAI offers EU data residency projects: GPT-Realtime-mini and the text models keep their list price, GPT-Live-1's voice layer costs 10% more. A pipeline is the easiest to keep in the EU because residency is decided component by component: STT and TTS on their EU endpoints, the LLM on an EU project or an EU region, and any vendor that cannot comply is swapped without touching the rest. Tick EU data residency in the simulator to apply these prices and adapt the recommendation.

Public list prices checked on 15 September 2026 from the Gemini API and Vertex AI pricing pages, the OpenAI API pricing page, xAI's pricing for Grok Voice, the Deepgram, Cartesia and Gradium pricing pages (Gradium's per-character credits converted at about 900 characters per spoken minute), and artificialanalysis.ai for quality and latency. Volume discounts, committed-use contracts and enterprise rates are not modelled. Actual bills differ with the conversation shape: adjust the turns per minute, speech share and tool calls in the advanced settings.