Top spot on voice conversation leaderboard changes hands: Phonic 1 wins by 0.2 points, ranks only 8th in human blind listening

Artificial Analysis has released a new edition of its Speech to Speech leaderboard. Phonic 1 (High) ranks first with 87.1 points, 0.2 points ahead of Grok Voice 3.0. The gap is so small that the comments first ask about the margin of error. More interesting is another set of numbers: in Speech Agent Arena, which measures human blind-listening preference, it ranks 8th out of 30 models. Being first overall and being what people like to hear are not the same thing.

Phonic 1 is a Speech to Speech model from @phonic_co, available through the Phonic API, priced at $0.07 per minute. Artificial Analysis evaluated the High intelligence tier. Speech to Speech means audio input goes directly to audio output, without going through the serial pipeline of 'speech recognition to text, text model generation, then speech synthesis.' For voice agents, latency, interruptions, tool calling, and policy adherence are all affected.

First overall, but wins by only 0.2 points

The Speech to Speech Index averages four items: Speech Reasoning (Big Bench Audio), Agentic Performance (Tau Voice), Arena Preference, and Task Success Rate. Phonic 1 (High) ranks first among 22 models with 87.1 points; Grok Voice 3.0 is at 86.9. A 0.2-point gap is suitable for spotting trends, not for treating as domination.

8th in human preference, 4th in task success rate

Speech Agent Arena has real people blind-listen to live voice conversations and then choose the model they prefer. Phonic 1 (High) debuts at 8th out of 30 models with an Elo of 1012. It ranks behind Cartesia Managed Agents, Gemini 3.1 Flash Live Minimal, and Gemini 3.8 Live, and ahead of Eleven Realtime v1 Max. Its task success rate is 92.3%, ranking 4th, behind Grok Voice 3.0, Grok Voice 2.0, and Gemini 3.8 Live.

Some commenters, upon seeing this, said that it is first overall and first on Tau Voice, yet only 8th in human blind listening, and they want to know what exactly the top 7 are better at. This question has more model-selection value than the total score.

Leads Tau Voice by 5.4 points

Tau Voice measures agentic performance, strict policy adherence, and multi-step tool use, with scenarios leaning toward long customer-service tasks. Phonic 1 (High) scores 79.9%, ranking first, 5.4 points ahead of second-place GPT-Live-1 (Astra, medium). Behind it are Grok Voice 3.0 at 71.6%, Gemini 3.8 Live Extended Thinking (High) at 68.6%, and GPT-Live-1 (Sol, low) at 59.3%.

Arena's task success rate reflects short conversations, while Tau Voice tests strategy and tool calling in long tasks. Both metrics are called tasks, but the pressure is different.

Third in speech reasoning

On Big Bench Audio, Phonic 1 (High) scores 98.9%, ranking third. Ahead are StepAudio 3 Realtime at 99.7% and Qwen Audio 3.0 Realtime Plus at 99.2%. It leads Grok Voice 3.0 at 98.5% and Gemini 3.8 Live Extended Thinking (High) at 97.7%.

First-packet latency 0.61 seconds, fastest in the top five

Phonic 1 (High)'s average first-packet latency on Big Bench Audio is 0.61 seconds, about 25% faster than Grok Voice 3.0's 0.81 seconds, and less than half of GPT-Live-1 (Astra, medium)'s 1.34 seconds and Gemini 3.8 Live Extended Thinking (High)'s 1.35 seconds. It is the fastest model in the top five of the Speech to Speech Index; among the top five, only Phonic 1 (High), Grok Voice 3.0, and Grok Voice 2.0 are under 1 second.

$4.20 per hour

Phonic 1 (High)'s input audio cost is $4.20 per hour, equivalent to $0.07 per minute. That is lower than Grok Voice 3.0 at $4.80, GPT-Live-1 (Sol, low) at $4.47, and GPT-Live-1 (Astra, medium) at $5.83; higher than Gemini 3.8 Live Extended Thinking (High) at $3.50; and about 2.6 times cheaper than GPT-Realtime-2.1 High at $10.75.

Real-world scenarios beyond the leaderboard

X user Lincoln Powers mentioned that benchmarks use studio audio, while real voice agents face a pickup truck driver with the windows down at 70 mph. This complaint is not against evaluation; it just reminds us of one thing: 0.61 seconds in the lab and 0.61 seconds under real noise are not the same experience.

Other commenters asked whether the 0.2-point gap between 87.1 and 86.9 has a margin of error. Some also separate task success rate and human preference and choose models by scenario. This approach is closer to actual procurement.

This split is more useful than the first-place ranking. The overall score averages reasoning, agency, preference, and task success rate, but when selecting a model you cannot just average them. For long-workflow customer service, look first at Tau Voice and task success rate; for a branded voice entry point, look first at Arena preference and latency; for cost-sensitive scenarios, calculate the hourly audio cost first. Phonic 1 puts agentic capability and price in a competitive position, but human blind-listening preference has not caught up yet.

One commenter's sarcasm is quite accurate: customer service can finally misunderstand me with lower latency. The top-ranked voice model is not the same as the top-ranked phone-answering model.

Comparison page: https://artificialanalysis.ai/models/speech-to-speech

Speech Agent Arena leaderboard: https://artificialanalysis.ai/speech-to-speech/arena

Methodology: https://artificialanalysis.ai/methodology/speech-to-speech-benchmarking