Introducing Resia Fullplex-1

We are incredibly excited to launch Resia Fullplex-1, a full-duplex voice engine pushing the pareto frontier of intelligence and cost efficiency. On the τ³-Voice public benchmark, Fullplex beats frontier models at a fraction of the cost. Fullplex is now available as a general purpose speech-to-speech API at $0.025 per minute, and in the Resia platform for use in fully managed call agents with a white-glove service.

Task performance

Resia Fullplex-1Resia Fullplex-1Tasks completed: 86.6%Price per hour: $1.50Plus the tokens of the task model you choose.
GPT-Live-1GPT-Live-1Tasks completed: 81.7%Price per hour: $3.00Plus backend model and tool usage, billed separately.
phonic-1phonic-1Tasks completed: 83.7%Price per hour: $4.20No separate LLM or per-token charges. Pro: $0.065 a minute with a $200 monthly minimum.
Grok Voice 3.0*Grok Voice 3.0 *Tasks completed: 79.6%Price per hour: $4.80Plus $0.004 per text input. Score from a run with a different simulated caller.
Gemini 3.8 Live*Gemini 3.8 Live *Tasks completed: 81.3%Price per hour: $2.32 (estimated)Price estimated from our 10-call test. Score: English part of a six-language run with a different simulated caller.

List price per hour of conversation (US dollars, lower is better)

Scores: τ³-Voice leaderboard. The score of Resia Fullplex-1 is from our own run of the same tasks. Our submission to the leaderboard is under review. Prices: makers' list prices, November 2026. Models without publicly available pricing are excluded. See notes below.

Backed by real-world experience

Resia's voice AI also powers Medfinder, which autonomously calls pharmacies to locate hard-to-find medications. It draws on our experience of more than 3 million real-world phone calls, helping more than 45,000 patients fill their prescriptions in spite of supply chain challenges. We learned from every agent failure, every unwarranted silence, and every frustrated callee, to build a voice experience that delivers results in the real world. These cumulative insights mean that Medfinder now has a task success rate of over 90%, and a 4.5 rating on Trustpilot. Today we make this technology available to power your business as well.

Full duplex

Traditional voice AI systems with cascaded architectures first convert speech to text, process the text through an LLM and re-convert the response text to speech. Each stage waits for the earlier ones, leading to awkward silences that ruin the experience.

Traditional cascade architecture

Recorded call · 0:26

Caller

Agent

3.1 s of silence2.9 s of silence
  1. CallerIs tomorrow anytime available?
  2. AgentI don't have access to the schedule, but I can send the request to the team. What time tomorrow would work best for you?
  3. CallerMaybe, like, ten AM…
  4. Callerif possible.
  5. AgentThanks. I'll send a request for John Smith tomorrow at ten AM for a possible root canal. The office will call you to confirm the time. Is there anything else?

Resia Fullplex

Recorded call · 0:28

Caller

Agent

Talks through “yeah” and “okay”Stops when you go onTalks through “uh-huh”
  1. CallerIs tomorrow anytime available?
  2. AgentI don't have access to the schedule, but I can send your request to the team. What time tomorrow would work best for you?
  3. CallerYeah.
  4. CallerOkay.
  5. CallerUh-huh.
  6. CallerMaybe, like…
  7. Callerten AM.
  8. Callerif possible.
  9. AgentThank…
  10. AgentThank you. To make sure I have everything right, you are John Smith, and you would like an appointment tomorrow at ten AM for a possible root canal. The office will call you to confirm the time. Is there anything else I can help with?
  11. CallerUh-huh.
Two recorded test calls with made-up details, cut at the same point of the booking. The agent that takes turns is set to answer after 0.8 seconds of silence. In its call a test agent plays the caller; in the Fullplex call a member of our team does.

Fullplex's architecture, however, is full duplex. It keeps thinking while you speak, producing blazing fast responses. It can tell apart a pause when you're thinking from when you are done speaking, and won't interrupt you mid-thought. Fullplex keeps talking during slow work. Your tools and longer tasks run while the conversation goes on. And yes, Fullplex is just a portmanteau of full duplex. Our marketing team insists it took more than five minutes.

Ignores speech not for it · higher is better

Resia Fullplex-1

72.5%

Ava Zero

47.4%

GPT-Live-1

45.7%

gemini-3.1-flash-live

43.1%

grok-voice-think-fast-1.0

28.4%

grok-voice-think-fast-2.0

27.7%

gpt-realtime-2

15.6%

qwen3.5-omni-plus-realtime

14.5%

Launa v1

11.6%

phonic-1

10.0%

Pine Voice 1.0

8.1%

The caller talks to someone else in the room.

Scores: τ³-Voice leaderboard submission files, read November 17, 2026, newest model of each maker. Benchmark: τ³-Voice, English (voice-v1): airline 50 tasks, retail 114, telecom 114, one trial per task, standard simulated caller. The score of Resia Fullplex-1 is from our own run of the same tasks. Our submission to the leaderboard is under review.

Industry leading cost efficiency

Frontier performance does not need to break the bank. Fullplex pairs it with one of the lowest costs per hour of conversation. At Medfinder, low cost was a necessity, not a luxury; our unit economics depended on it. Now we pass these savings on to you, so that formerly untenable use cases become feasible.

Price per hour

Resia Fullplex-1

$1.50

Gemini 3.8 Live

$2.32est.

GPT-Live-1

$3.00

phonic-1 (Starter)

$4.20

grok-voice-3.0

$4.80

Eleven Realtime v1 Mini

$4.80

Eleven Realtime v1 Max

$4.80

gpt-realtime-2

$7.10est.

Prices: makers' list prices, November 2026. Models without publicly available pricing are excluded.

While healthcare and ed-tech domains have seen the highest demand for Fullplex so far, it generalizes well to other domains.

Tasks completed on τ³-Voice

Pine Voice 1.0

87.3%

Resia Fullplex-1

86.6%

phonic-1

83.7%

GPT-Live-1

81.7%

Ava Zero

80.4%

Launa v1

72.9%

grok-voice-think-fast-1.0

67.3%

grok-voice-think-fast-2.0

62.5%

qwen3.5-omni-plus-realtime

53.7%

gemini-3.1-flash-live

43.8%

gpt-realtime-2

42.4%

Pass^1: the share of tasks done right on the first try, averaged over the three domains. Higher is better. The score of Resia Fullplex-1 is from our own run of the same tasks. Our submission to the leaderboard is under review.

Tasks completed, by domain

Airline

50 tasks

Pine Voice 1.0Airline: 92.0%phonic-1Airline: 80.0%GPT-Live-1Airline: 82.0%Ava ZeroAirline: 86.0%Launa v1Airline: 66.0%grok-voice-think-fast-1.0Airline: 66.0%grok-voice-think-fast-2.0Airline: 56.0%qwen3.5-omni-plus-realtimeAirline: 54.0%gemini-3.1-flash-liveAirline: 64.0%gpt-realtime-2Airline: 58.0%Resia Fullplex-1Airline: 88.0%

Retail

114 tasks

Pine Voice 1.0Retail: 78.8%phonic-1Retail: 76.3%GPT-Live-1Retail: 78.9%Ava ZeroRetail: 75.4%Launa v1Retail: 71.9%grok-voice-think-fast-1.0Retail: 62.3%grok-voice-think-fast-2.0Retail: 59.6%qwen3.5-omni-plus-realtimeRetail: 46.5%gemini-3.1-flash-liveRetail: 45.6%gpt-realtime-2Retail: 47.4%Resia Fullplex-1Retail: 75.4%

Telecom

114 tasks

Pine Voice 1.0Telecom: 91.2%phonic-1Telecom: 94.7%GPT-Live-1Telecom: 84.2%Ava ZeroTelecom: 79.8%Launa v1Telecom: 80.7%grok-voice-think-fast-1.0Telecom: 73.7%grok-voice-think-fast-2.0Telecom: 71.9%qwen3.5-omni-plus-realtimeTelecom: 60.5%gemini-3.1-flash-liveTelecom: 21.9%gpt-realtime-2Telecom: 21.9%Resia Fullplex-1Telecom: 96.5%
Resia Fullplex-1Other systems on the leaderboard

Scores: τ³-Voice leaderboard submission files, read November 17, 2026, newest model of each maker. Benchmark: τ³-Voice, English (voice-v1): airline 50 tasks, retail 114, telecom 114, one trial per task, standard simulated caller. The score of Resia Fullplex-1 is from our own run of the same tasks. Our submission to the leaderboard is under review.

Conversational dynamics

Realism in voice conversations comes not only from the speed of response, but also from knowing when not to speak. As evident on the τ³-Voice benchmark, Fullplex is best at recognizing and ignoring a “mm-hmm” from the caller, a cough in the background, or words to someone else in the room.

Resia Fullplex-1Resia Fullplex-1Selectivity: 70.6%Price per hour: $1.50Plus the tokens of the task model you choose.
GPT-Live-1GPT-Live-1Selectivity: 67.8%Price per hour: $3.00Plus backend model and tool usage, billed separately.
phonic-1phonic-1Selectivity: 30.7%Price per hour: $4.20No separate LLM or per-token charges. Pro: $0.065 a minute with a $200 monthly minimum.
Grok Voice 3.0*Grok Voice 3.0 *Selectivity: 27.6%Price per hour: $4.80Plus $0.004 per text input. Score from a run with a different simulated caller.

List price per hour of conversation (US dollars, lower is better)

Scores: τ³-Voice leaderboard. The score of Resia Fullplex-1 is from our own run of the same tasks. Our submission to the leaderboard is under review. Prices: makers' list prices, November 2026. Models without publicly available pricing are excluded. See notes below.

HIPAA and Zero Data Retention

Resia supports HIPAA workloads. A Business Associate Agreement (BAA) is free for every customer, on request. Zero Data Retention is also available on request. Contact sales@resia.ai, and our team will get back to you promptly.

Test it in the Playground.

Talk to Fullplex in your browser. No code, no phone number, and no setup.

  1. 1

    Configure. Choose a voice. Write a prompt, or start from a template such as a front desk or a care line.

  2. 2

    Talk. Press Start and speak. Interrupt it, pause, and say “mm-hmm”. The words of both sides show as you talk.

  3. 3

    Ship. Save the setup as a Call Agent, or copy the code for your own app in Python or JavaScript.

A session uses your credit at the session price. A new account starts with $10 of free credit.

Put Fullplex on your first call today.

Sign up, describe the job to Butler, and test it on your own phone.

Notes

Get a call now

A Resia agent will call you right away, walk you through the platform, and put you through to our founder if you want to talk to him.

United States and Canada numbers only. Asking for a call means we may call this number. The call is recorded. See our Privacy Policy.

Book a demo

Tell us who you are and we will take you straight to the calendar with your details already filled in.