Lowest-Latency Inference APIs for Voice and Realtime Agents: A Time to First Token TTFT-First Benchmark

Time to first token (TTFT) is the metric teams use to pick an inference API for voice. It is also the metric that misleads them. TTFT marks when generation starts; a text-to-speech model cannot speak until a full clause arrives. Between those two points sits the difference between an agent that feels conversational and one that gets interrupted. This piece benchmarks every layer of the voice stack including LLM, speech-to-text, text-to-speech, and speech-to-speech. Why TTFT Is the Right Entry Point and the Wrong Finish Line A voice agent is a latency budget with a language model inside it. Every stage spends milliseconds the user can hear. Time to first token (TTFT) is the interval between sending an inference request and receiving the first token back. IBM’s definition frames it as the moment a system transitions from idle to visibly active. For chat, TTFT is close to the whole story. For voice, it is one term in a sum. The reason is mechanical. A text-to-speech model cannot synthesize half a word. It needs a complete clause or sentence before it produces audio. LiveKit calls the resulting metric time-to-first-sentence (TTFS), and argues in its Gemma 4 deployment post that TTFS is what users actually feel. That gives you two knobs rather than one. TTFT controls when generation starts. Tokens per second controls how fast the first sentence completes. A provider that wins one and loses the other will not feel fast. The Latency Budget: What One Voice Turn Actually Costs LiveKit’s voice agents overview breaks a turn into STT at roughly 100–200ms, LLM at 300–500ms with streaming, TTS at 100–200ms, and network at 50–150ms over WebRTC. It puts the practical end-to-end target at 700ms to 1.2s. Kwindla Hultman Kramer, co-creator of Pipecat, has advised targeting 800ms median voice-to-voice latency, with a looser 1,500ms acceptable for a proof of concept. His rough arithmetic splits that four ways at roughly 200ms each: transport and media processing, STT plus phrase endpointing, LLM inference, and TTS. Daily’s earlier work on the fastest voice bot supplies the human baseline. Typical human response time in conversation is around 500ms. Pauses beyond 800ms start to feel unnatural. Daily’s February 2026 voice-agent LLM benchmark translates that into an LLM requirement directly. Natural conversation needs voice-to-voice under 1,500ms, which works out to roughly 700ms of TTFT budget for a text-mode LLM inside a transcription-to-LLM-to-voice harness. That 700ms number is the bar to hold every provider against. How to Read a TTFT Benchmark Without Being Misled Before the tables, five methodology facts that change what the numbers mean: 1. Workload shape dominates: Artificial Analysis changed its default workload in March 2026. The site now reports 10k input token prompts rather than 1k. Longer prompts raise both TTFT and output speed. LiveKit argues this is closer to reality for voice, because production agents front-load policy, persona, escalation rules, retrieved data, and tool schemas. 2. Server location is baked in: Artificial Analysis tests from a virtual machine in Google Cloud’s us-central1-a zone. It states plainly that TTFT includes network latency and may advantage or disadvantage providers based on where they serve. 3. Reasoning tokens count: In the Artificial Analysis definition, TTFT for a reasoning model is the first reasoning token, not the first answer token. Those are separate columns. 4. Measure from the receiving side: Daily notes that model providers sometimes quote TTFT internal to their inference stacks. Daily measures from request send to first usable token off the API. 5. Runs are not repeatable: Daily is blunt about this: TTFT varies substantially between benchmark runs, and providers change inference stacks and sometimes weights without changing model names. Layer 1: LLM Time to First Token Figures below are from the Artificial Analysis API providers leaderboard, retrieved August 30, 2026. The “first chunk” column is TTFT. Workload is 10k input tokens, single prompt, median over 72 hours. Lowest measured first-chunk latency ProviderModelTTFTOutput speed Basetengpt-oss-120b (high)0.23s266 tok/s Basetengpt-oss-120b (low)0.24s271 tok/s DeepInfraNemotron 3 Ultra0.28s371 tok/s CohereNorth Mini Code0.32s104 tok/s CohereCommand A+0.40s239 tok/s BasetenInkling Small0.42s337 tok/s ModularGemma 4 31B (NVFP4)0.44s243 tok/s NebiusGLM-5.3-Flash0.46s206 tok/s FireworksNemotron 3.5 Lightning0.46s501 tok/s Together AIKimi K2.7 Code0.47s245 tok/s Cerebrasgpt-oss-120b (high)0.49s1,697 tok/s The throughput trap Silicon vendors optimize for a different metric than voice agents need. ProviderModelTTFTOutput speed Cerebrasgpt-oss-120b (high)0.49s1,697 tok/s CelerisCeleris-10.62s1,612 tok/s CerebrasGemma 4 31B0.53s1,351 tok/s Groqgpt-oss-20b (high)0.82s957 tok/s SambaNovagpt-oss-120b (high)0.92s706 tok/s Groqgpt-oss-120b (low)0.69s473 tok/s InceptionMercury 23.07s770 tok/s Mercury 2 is the clearest illustration. It is a diffusion-based language model, and it generates 770 tokens per second. Its first chunk arrives at 3.07s. That is four times the entire LLM budget for a natural conversation. Cerebras and Groq are a different case. Their TTFT is respectable and their throughput is exceptional. For TTFS specifically, that combination is strong, because the sentence completes almost immediately after the first token lands. Frontier and proprietary endpoints ProviderModelTTFTOutput speed Amazon BedrockGPT-5.6 Luna (non-reasoning)0.59s181 tok/s Amazon BedrockGPT-5.6 Terra (non-reasoning)0.72s103 tok/s OpenAIGPT-5.6 Luna (non-reasoning)0.74s113 tok/s GoogleGemini 3.7 Flash (low), AI Studio0.84s315 tok/s AnthropicClaude 4.5 Haiku (non-reasoning)0.84s82 tok/s Amazon BedrockNova Micro0.86s264 tok/s GoogleGemini 3.5 Flash (minimal), AI Studio0.90s202 tok/s OpenAIGPT-5.6 Sol (non-reasoning)1.06s71 tok/s Note the same model on different hosts. GPT-5.6 Luna non-reasoning measures 0.59s on Amazon Bedrock and 0.74s on OpenAI’s own API. Hosting and routing matter as much as the weights. The vendor-measured outlier LiveKit publishes TTFT figures for its own inference product. Gemma 4 31B on LiveKit Inference measured 192ms, against Gemini 2.5 Flash at 911ms, GPT-5.5 at 966ms, GPT-4.1 at 1,006ms, and the same Gemma 4 31B via OpenRouter at 1,876ms. LiveKit is transparent about the mechanism, which makes the claim more credible than most. It runs Gemma behind SGLang with speculative decoding, and deliberately under-packs each GPU so queueing delay stays low. A warm request, it says, starts returning tokens in around 100ms. The tradeoff is cost, at $1.20 per 1M output tokens. The same post reports TTFS across full conversations: 354ms for Gemma 4 31B on LiveKit, 1,034ms for Gemini 2.5 Flash, 1,088ms for GPT-4.1, 1,267ms for Gemini 3.0 Flash, and 1,404ms for GPT-5.5. Capability numbers accompany it. On IFBench, independently scored by Artificial Analysis, Gemma 4 31B scores 75.6% against GPT-5.5 at 75.9%, GPT-4.1 at 43%, and Gemini 2.5 Flash at 39%. On τ²-bench, GPT-5.5 leads at 93.9% with Gemma 4 31B at 76.9%. Layer 2: Speech-to-Text and Turn Detection For voice, STT latency is not transcription speed. It is how long after the user stops talking the pipeline knows the user stopped talking. Artificial Analysis measures two things on its streaming STT leaderboard, both starting from a SileroVAD-detected end of speech: time to first partial transcript, and time to final transcript. Its AA-WER Streaming index draws on roughly 8 hours of audio, weighted AA-AgentTalk 50%, VoxPopuli 25%, Earnings-22 25%. Vendor-published latency figures: ModelClaimSource type Deepgram Flux~260ms p50 end-of-turn detection at defaultsVendor docs Deepgram Nova-3Sub-300ms streaming latencyVendor docs AssemblyAI Universal-Streaming~300ms immutable word emissionVendor Cartesia Ink-2100ms transcript latencyVendor Speechmatics Voice SDK0.451 ± 0.022s end-of-speech to finalsVendor internal tool Deepgram Flux is the most architecturally interesting entry. It folds end-of-turn detection into the recognition model rather than bolting a VAD on top. Deepgram states this can cut agent response latency by 200–600ms versus a traditional STT-plus-VAD pipeline. It exposes eot_threshold (0.5–0.9), eager_eot_threshold (0.3–0.9), and an EagerEndOfTurn event that lets you start the LLM early. That last capability matters more than the raw number. If you can begin generation on an eager signal, you move LLM TTFT off the critical path entirely when the prediction is right. AssemblyAI Universal-Streaming inverts the usual partials-then-finals model by emitting immutable transcripts. AssemblyAI reported 307ms median word emission against 516ms for Deepgram Nova-3 in its own 2025 measurement. Its docs also recommend using unformatted transcripts for voice agents, since formatting arrives later and rarely changes LLM behavior. Accuracy claims here are contested and vendor-published. AssemblyAI reports Universal-3.5 Pro Realtime at 6.99% WER on the open Pipecat voice-agent benchmark, ahead of Google Chirp3 at 9.04%, ElevenLabs Scribe v2 at 9.76%, and Deepgram Flux at 15.58%. Run it yourself before treating it as settled. LiveKit also documents preemptive generation, which starts the LLM on a partial transcript. The caveat is real: if the reply has to be regenerated after the final transcript, you burn tokens and save nothing. Layer 3: Text-to-Speech Time to First Audio This is where vendor numbers diverge most sharply from what users experience. ElevenLabs states Flash v2.5 delivers approximately 75ms. Its own docs qualify that carefully: 75ms refers to model inference time only. The company’s latency concepts page goes further, listing network round-trip at typically 20–200ms depending on geography, and noting that most audio players buffer before playback, with 500ms buffering being common. It also states that Eleven v3 is not built for real-time, and recommends Flash v2.5, Flash v2, or Multilingual v2 for its Agents Platform. Cartesia states sub-90ms TTS and 100ms transcript latency for Sonic-3.6 and Ink-2. Marktechpost’s coverage of the Sonic-3.6 release flagged both as vendor-stated model latency, not measured end-to-end round trips. Cartesia previously claimed 82ms end-to-end time-to-first-audio for Sonic 3.5. Sonic runs on state space models rather than transformers, which scale linearly rather than quadratically with sequence length. On quality, the Artificial Analysis Provider Voice arena is blind-listener Elo, retrieved August 30, 2026: ModelEloPrice per 1M chars Cartesia Sonic 3.61,288$49.00 SpeechifyAI Simba 3.21,243$10.00 Alibaba Qwen-Audio-3.0-TTS-Plus1,243$27.60 Inworld Realtime TTS-2 Flash (preview)1,228$10.40 BreezeBlue Breeze TTS 2 (open weights)1,220$34.00 ElevenLabs v3 Conversational1,215$50.00 Google Gemini 3.1 Flash TTS1,210$18.30 ElevenLabs Flash v2.51,083$50.00 The gap between Sonic 3.6 at 1,288 and Flash v2.5 at 1,083 is the quality cost of the low-latency tier that most agents actually run on. Layer 4: Speech-to-Speech Time to First Audio Speech-to-speech models collapse STT, LLM, and TTS into one pass. Fewer round trips should mean lower latency. LiveKit is careful here, noting that realtime models are not guaranteed to be faster in every case, and that a well-tuned pipeline can be highly competitive. The data supports that caution. From the Artificial Analysis speech-to-speech leaderboard, TTFA measured on Big Bench Audio, retrieved August 30, 2026: ModelTTFASpeech reasoningTask successS2S Index Deepslate Opal0.44s85%—— Gemini 2.5 Flash Native Audio Dialog0.63s69%—— Grok Voice Think Fast 2.0 High0.70s97%94.7%79.0% Grok Voice Fast 1.00.78s93%—— Qwen3.5 Omni Flash Realtime0.79s59%29.1%— OpenAI GPT-Realtime-1.50.81s81%85.1%70.3% OpenAI GPT Realtime Mini (Oct ’25)0.81s64%79.6%56.8% OpenAI GPT-R [truncated for AI cost control]