MIT-Founded Phonic Beats Grok and GPT With Fastest Voice AI Response

Phonic 1 (High) has debuted at number one on the Artificial Analysis Speech to Speech Index. Its score of 87.1 narrowly exceeds Grok Voice 3.0 at 86.9, while its 0.61-second average time to first audio is the fastest among the index’s top five models.

Phonic is a young company founded by MIT engineers . Its voice-agent platform handles audio end to end with proprietary foundation models. Conventional voice systems often chain speech recognition, a language model, and speech synthesis from separate vendors. Phonic’s design reduces those component handoffs and concentrates the stack within one platform.

Four tests behind the score

The index gives equal weight to four benchmarks, each covering a different part of a production voice agent’s job:

• Big Bench Audio measures speech reasoning with 1,000 audio questions adapted from Big Bench Hard. Tasks include formal fallacies, navigation, object counting, and Boolean word problems.
• τ-Voice measures policy compliance and tool use. A simulated customer calls with a service problem, and the model must follow written rules and use tools to resolve it.
• Arena Preference Elo reflects blind pairwise ratings from people who converse with competing voice models.
• Task Success Rate tracks the share of arena conversations that finish with the correct tool call.

Equal weighting prevents a model from leading through conversational preference or reasoning alone. The composite also rewards accurate tool use, policy adherence, and task completion.

Tool use creates the gap

Phonic records its largest lead on τ-Voice methodology , scoring 79.9%. That is 5.4 percentage points above GPT-Live-1 (Astra, medium) and 8.3 points above Grok Voice 3.0.

Those tasks resemble customer-service calls in which an agent must consult account data, apply a policy, invoke several tools, and produce a valid resolution. The result is particularly relevant to airline, retail, insurance, and telecom workflows.

Phonic also scores 98.9% on Big Bench Audio, placing third behind StepAudio 3 Realtime at 99.7% and Qwen Audio 3.0 Realtime Plus at 99.2%. In the live Speech Agent Arena, it debuts eighth of 30 models on preference Elo with a score of 1012. Its 92.3% Task Success Rate ranks fourth, behind two Grok Voice models and Gemini 3.8 Live.

The difference between its arena rankings provides useful context: users preferred several competing voices overall, while Phonic performed better at completing the requested task. Product teams still need to evaluate voice quality, conversational style, and task reliability separately.

Subsecond audio, mid-pack cost

Phonic’s average time to first audio on Big Bench Audio is 0.61 seconds. That is about 25% faster than Grok Voice 3.0 and less than half the measured delay of the listed OpenAI and Google configurations.

Only Phonic 1 and the two Grok Voice models in the top five begin responding within one second. Low first-audio latency reduces dead air after a caller finishes speaking, although full conversational responsiveness also depends on turn detection, network conditions, tool execution, and interruption handling.

Phonic charges $4.20 per hour of input audio, equivalent to $0.07 per minute. Grok Voice 3.0 costs $4.80, while GPT-Live variants range from $4.47 to $5.83. GPT-Realtime-2.1 High costs $10.75, and Gemini 3.8 Live Extended Thinking costs $3.50. These input-audio rates do not represent the complete cost of a call, which may also include output audio, tools, telephony, storage, and application infrastructure.

Inside the WebSocket integration

Phonic makes the model available through its API. A public third-party integration connects to wss://api.phonic.ai/v1/sts/ws and streams pcm_24000 audio in both directions. It uses 800 milliseconds of silence to determine when the caller has finished a turn.

The session configuration declares tools inline as strict, synchronous custom WebSocket tools. The application receives a tool request, validates and executes it, then returns the result through the active session. Phonic also exposes choose_not_to_respond , allowing an agent to remain silent when a response would be inappropriate.

Requests select an intelligence level, and Phonic recommends high for agentic workloads. Artificial Analysis evaluated that configuration, so developers comparing local tests with the leaderboard should use the same setting.

Because the available protocol example is a third-party implementation, production teams should verify authentication, session schemas, supported codecs, rate limits, concurrency, regional availability, and tool-call behavior against Phonic’s current documentation. Telephone networks commonly use lower-bandwidth codecs, so applications may also need to resample audio before sending 24 kHz PCM.

Production tests still matter

The index gives developers a compact comparison of reasoning, policy compliance, human preference, task completion, and response latency. Several deployment concerns remain outside the headline score, including noisy lines, accents, callers who interrupt, packet loss, tool timeouts, malformed tool results, privacy controls, data retention, and uptime commitments.

A 0.2-point overall lead is narrow, and leaderboard positions can change as models are updated and arena votes accumulate. Phonic’s more substantial result is the 5.4-point τ-Voice advantage combined with subsecond first-audio latency.

StepAudio 3 and Qwen Audio 3.0 lead on speech reasoning, while the larger commercial APIs offer broader platform ecosystems. Phonic’s results show a different optimization target: enterprise phone workflows that prioritize policy execution, correct tool calls, and short pauses.

Teams evaluating voice agents for appointment booking, claims handling, account service, or support triage can add Phonic 1 (High) to a replay test using real call audio and production tool schemas. Completion accuracy, interruption recovery, end-to-end latency, and total call cost will determine whether the leaderboard advantage survives in deployment.

A funding report describes a $27.5 million Series A led by First Round and says the model runs on Together AI infrastructure. Buyers with data-residency or subprocessor requirements should confirm how that hosting arrangement applies to their regions and workloads.