--- title: "Voice AI models and benchmarks" description: "Measured voice AI benchmarks across speech-to-text, LLM, text-to-speech, and speech-to-speech, with current Router route availability kept separate from historical evidence." canonical: "https://speko.ai/models" last-updated: "2026-08-24" --- > ## Speko page index > The complete index of every page on this site is at: https://speko.ai/llms.txt > Read it before exploring further. It lists the exact Markdown URL for every canonical HTML page. # Voice AI models and benchmarks 106 supported models. Missing benchmark results are marked “not measured”. Prices are the benchmark publisher’s vendor rates. Model availability is based on a saved catalog; live availability could not be verified. ## Filter catalog Provider 106 models ## Speech-to-text34 Model Provider Id nemotron-asr ![Hopper logo](https://speko.ai/provider-logos/hopper.svg) Hopper 23.1% (20.0–26.1% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-08 · 6 of 566 readings came back empty, scored as all-deletion. · Measured 2026-10-08) 9.1% (95% CI 7.0–11.4 · n=50 · vendor-direct batch · FLEURS · The scorer counts spelled-out numbers as errors. On the 40 clips without numbers it reads 7.6%. · Measured 2026-10-08) 9.8% (n=1541 · FLEURS test+dev+train · vendor-direct · Scored with the same normalizer as batch, which counts spelled-out numbers as errors. · Measured 2026-10-08) 656ms (p50–p90 · US East · n=200 · vendor-direct · re-measured 2026-10-08 · ±8% · Marker-limited: 68 of 200 readings land before the measured speech end, 28 by more than 0.5s (Modulate on the same clips: 27 and 12). Hopper serves from Oregon, so US readings include a cross-country round trip. · Measured 2026-10-08) $0.0008 (Vendor list, $0.05 per audio hour. · Measured 2026-10-08) Pulse ![Smallest AI logo](https://speko.ai/provider-logos/smallest.png) Smallest AI 14.5% (12.4–16.7% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · 2 of 566 readings came back empty, scored as all-deletion. · Measured 2026-10-01) 5.1% 5.9% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 839ms (p50–p90 · US East · n=200 · vendor-direct · ±5%) $0.0040 (Streaming Pulse, the vendor's published per-model rate (smallest.ai/pricing/models; pre-recorded Pulse is $0.003/min). The ~$0.009/min on the main pricing page is the STT slice of a blended voice-agent stack, not this API's list rate.) Nova-3 ![Deepgram logo](https://speko.ai/provider-logos/deepgram.png) Deepgram 17.0% (14.4–19.9% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · 1 of 566 readings came back empty, scored as all-deletion. · Measured 2026-10-01) 12.0% (n=50 · FLEURS clean read-English · Corrected 2026-09-02 from a published 9.8% that has no reproducible source. 11.96% is the mean per-clip WER of the committed batch artifact results/anchor-nova3-20260808.json (run of 2026-08-08), re-confirmed by rescoring that artifact with the current text normalizer — 11.96% as stored and 11.96% rescored.) 10.8% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 189ms (p50–p90 · US East · n=200 · vendor-direct · ±9%) $0.0048 (Streaming pay-as-you-go. Vendor flags it as a limited-time promotional rate; pre-recorded is $0.0043.) Flux ![Deepgram logo](https://speko.ai/provider-logos/deepgram.png) Deepgram 19.7% (16.7–22.6% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · 10 of 566 readings came back empty, scored as all-deletion. · Measured 2026-10-01) not measured (Measured 2026-08-04) 7.2% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 712ms (p50–p90 · US East · n=200 · vendor-direct · ±15% · Measured 2026-08-04) $0.0065 (Flux English streaming pay-as-you-go. Vendor flags it as a limited-time promotional rate; flux-general-multi is $0.0078. · Measured 2026-08-04) flux-general-multi ![Deepgram logo](https://speko.ai/provider-logos/deepgram.png) Deepgram 21.4% (18.0–24.6% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · 8 of 566 readings came back empty, scored as all-deletion. · Measured 2026-10-01) not measured (Measured 2026-09-28) 8.7% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 854ms (p50–p90 · US East · n=200 · vendor-direct · ±10% · Measured 2026-09-28) $0.0078 (Flux Multilingual streaming pay-as-you-go, no promotional rate. Checked 2026-09-28. · Measured 2026-09-28) Ink-2 ![Cartesia logo](https://speko.ai/provider-logos/cartesia.svg) Cartesia 16.7% (14.6–19.1% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) 11.0% 10.1% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 587ms (p50–p90 · US East · n=200 · vendor-direct · pre-close (99/100 flush-forced) · ±34%) $0.0090 (3 credits per second at the Pro included rate ($5 for 100K credits), the cheapest paid plan; $0.0071 at Startup, $0.0067 at Scale.) Qwen3-ASR Realtime ![Alibaba logo](https://speko.ai/provider-logos/alibaba.png) Alibaba 12.2% (EdAcc n=97 · earlier set · Earlier 98-clip set: the session hung in the 2026-10-01 re-measure. UNUSABLE FOR A CALL: median 311s of wall clock per 7s clip, up to 1123s. · Measured 2026-08-28) not measured 4.0% (n=50 · FLEURS · earlier set · Earlier 50-clip set: the session hung in the 2026-10-01 re-measure. · Measured 2026-07-21) 889ms (p50–p90 · US East · n=189 · vendor-direct · ±6%) $0.0054 (qwen3-asr-flash-realtime, Singapore. The file/sync SKU is $0.0021.) Chirp 3 Batch only ![Google logo](https://speko.ai/provider-logos/google.png) Google 18.4% (15.2–21.8% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) 3.9% 5.9% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) not measured (not measured on the vendor-direct regional sweep) $0.0160 (Speech-to-Text V2 Standard, first tier — one rate for streaming, sync and batch. Dynamic Batch Recognition is $0.003.) Scribe v2 Realtime ![ElevenLabs logo](https://speko.ai/provider-logos/elevenlabs.png) ElevenLabs 12.5% (10.9–15.0% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) not measured (Measured 2026-07-21) 4.3% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 1.75s (p50–p90 · US East · n=200 · vendor-direct · ±4% · Measured 2026-07-21) $0.0065 (Scribe v2 Realtime at $0.39/hr, flat across every tier. Batch Scribe v2 is $0.0037. · Measured 2026-07-21) transcribe-1-pro Batch only ![Fish Audio logo](https://speko.ai/provider-logos/fish.png) Fish Audio not measured (Measured 2026-09-22) 3.4% (2.1–4.9% · 95% clip bootstrap · FLEURS n=50 · Vendor-direct batch; explicit speaker/event annotations removed, quoted overlap speech retained. · Measured 2026-09-22) not measured (Measured 2026-09-22) not measured (Measured 2026-09-22) $0.0060 (Batch ASR: $0.36/audio hour; billed duration rounded up to whole seconds. Fish Audio list price checked 2026-09-22. · Measured 2026-09-22) transcribe-1 Batch only ![Fish Audio logo](https://speko.ai/provider-logos/fish.png) Fish Audio not measured (Measured 2026-09-28) 3.2% (1.9–4.7% · 95% clip bootstrap · FLEURS n=50 · Vendor-direct batch, same clips and scorer as transcribe-1-pro. · Measured 2026-09-28) not measured (Measured 2026-09-28) not measured (Measured 2026-09-28) $0.0060 (Batch ASR: $0.36/audio hour, same rate as transcribe-1-pro. Fish Audio list price checked 2026-09-22. · Measured 2026-09-28) MAI-Transcribe-2 Batch only ![Azure logo](https://speko.ai/provider-logos/azure.svg) Azure not measured (Measured 2026-09-03) 2.3% (Same pass reproduced the Nova-3 anchor at its published 11.96%. · Measured 2026-09-03) not measured (Measured 2026-09-03) not measured (Measured 2026-09-03) $0.0017 ($0.10 per hour of audio, a launch price through 2026-12-31. · Measured 2026-09-03) Universal-3.5 Pro ![AssemblyAI logo](https://speko.ai/provider-logos/assemblyai.png) AssemblyAI 14.5% (12.4–16.9% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Deletions carry most of it: 7.1% dropped words against 5.8% substitutions. · Measured 2026-10-01) 2.0% 3.7% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-09-30) 408ms (p50–p90 · US East · n=200 · vendor-direct · ±12%) $0.0075 (Universal-3.5 Pro Realtime at $0.45/hr. Async pre-recorded is $0.0035. Diarization and other add-ons bill on top.) Universal-3.6 Pro ![AssemblyAI logo](https://speko.ai/provider-logos/assemblyai.png) AssemblyAI 13.7% (11.7–16.1% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · 0.79pt better than 3.5 Pro on the same clips (95% CI 0.03–1.67). · Measured 2026-10-01) not measured (Measured 2026-09-30) 3.1% (n=1541 · FLEURS test+dev+train · vendor-direct · 0.66pt better than 3.5 Pro on the same clips (95% CI 0.49–0.84). · Measured 2026-09-30) 505ms (p50–p90 · US East · n=200 · vendor-direct · re-measured 2026-09-30 · ±10% · Measured 2026-09-30) $0.0075 (Universal-3.6 Pro Realtime at $0.45/hr, unchanged from 3.5 Pro. Checked 2026-10-01. · Measured 2026-09-30) Velma 2 ![Modulate logo](https://speko.ai/provider-logos/modulate.png) Modulate 14.4% (12.0–17.3% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · 2 of 566 readings came back empty, scored as all-deletion. · Measured 2026-10-01) 4.4% (95% CI 2.8–6.2 · n=50 · Measured 2026-08-08) 5.3% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 1.21s (p50–p90 · US East · n=200 · vendor-direct · ±4% · Measured 2026-08-08) $0.0008 (Published streaming rate for English Fast, $0.05/hr. Batch is $0.025/hr. The $0.06/hr this row used to carry is the MULTILINGUAL row, a different product from the one this id names. Not a measurement. · Measured 2026-08-08) stt-rt-v5 ![Soniox logo](https://speko.ai/provider-logos/soniox.png) Soniox 16.6% (14.4–19.4% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) 7.5% (95% CI 4.8–10.7 · n=50) 6.3% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 340ms (p50–p90 · US East · n=200 · vendor-direct · ±16%) $0.0020 GPT Transcribe ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 16.8% (14.3–19.5% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) 2.5% (Measured 2026-07-03) 5.8% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 817ms (p50–p90 · US East · n=200 · vendor-direct · re-measured 2026-09-15 · ±7% · Marker-limited: 48 of 200 readings land before the measured speech end, 17 by more than 0.5s. The model emits nothing until its VAD closes a segment, so first token is the wait a caller hears. · Measured 2026-07-03) $0.0045 (Billed per minute, not per token. · Measured 2026-07-03) GPT Live Transcribe ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 13.5% (11.3–16.0% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) not measured (Measured 2026-07-21) 4.9% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 620ms (p50–p90 · US East · n=200 · vendor-direct · re-measured 2026-09-15 · ±3% · Measured 2026-07-21) $0.0170 (gpt-live-transcribe, the streaming path this row is recommended for. · Measured 2026-07-21) GPT Realtime Whisper ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 16.8% (14.0–19.9% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) not measured (Measured 2026-09-21) 5.0% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 1.14s (p50–p90 · US East · n=200 · vendor-direct · ±8% · Commit-to-final: the probe commits at speech end, so the vendor never decides the turn. The two US sweeps disagreed beyond the artifact’s bound; Singapore reads 1.34s. · Measured 2026-09-21) $0.0170 (Billed by audio duration at $0.017/minute, not by token (developers.openai.com model page, read 2026-09-21). Same rate as gpt-live-transcribe. · Measured 2026-09-21) Qwen3-ASR Fast ![Nari logo](https://speko.ai/provider-logos/nari.png) Nari 12.0% (9.9–14.3% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) 3.3% (2.0–4.8% · Measured 2026-09-13) 3.8% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 33ms (p50–p90 · US East · n=60 · vendor-direct · pre-close (30/30 flush-forced) · re-measured 2026-09-22 · ±9% · Commit-to-final: turn detection is off, so the probe commits at the measured speech end. Singapore reads 196ms. · Measured 2026-09-13) $0.0020 (Vendor list $0.12 per input audio hour, Partner tier (docs.narilabs.com/models-and-pricing, read 2026-09-13). A free tier of the same model is listed at $0. · Measured 2026-09-13) Qwen3-ASR ![Nari logo](https://speko.ai/provider-logos/nari.png) Nari 12.0% (10.0–14.4% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) 3.4% (2.1–4.9% · Measured 2026-09-26) 3.9% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 25ms (p50–p90 · US East · n=60 · vendor-direct · pre-close (30/30 flush-forced) · re-measured 2026-09-26 · ±22% · Commit-to-final: turn detection is off, so the probe commits at the measured speech end. Singapore reads 254ms. · Measured 2026-09-26) $0.0010 (Vendor list $0.06 per input audio hour (docs.narilabs.com/models-and-pricing, read 2026-09-26). · Measured 2026-09-26) GPT-4o-mini Transcribe ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 20.6% (17.1–24.3% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) 2.7% 5.7% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 822ms (p50–p90 · US East · n=200 · vendor-direct · ±9%) $0.0030 Grok Voice Transcribe 1.0 ![xAI logo](https://speko.ai/provider-logos/xai.png) xAI 13.7% (11.4–16.3% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) 2.8% (2.0–3.6% · 95% CI · n=100 · Measured 2026-09-22. FLEURS English at -16 LUFS, REST, pinned to grok-voice-transcribe-1.0: 100/100 clips, 0 empty, 0 retried. 2.0 vs 1.0 on the same clips: 13 better, 18 worse, 69 tied; paired difference -0.34 to +1.21 points (95% CI). · Measured 2026-09-22) 37.2% (n=1541 · FLEURS test+dev+train · vendor-direct · 238 of 1541 clips came back empty, scored as all-deletion. · Measured 2026-10-01) 1.13s (p50–p90 · US East · n=200 · vendor-direct · re-measured 2026-09-23 · ±4% · Measured 2026-09-22) $0.0033 (Streaming at $0.20/hr. REST is $0.0017. Unchanged across 1.0 and 2.0. · Measured 2026-09-22) Gradium ASR ![Gradium logo](https://speko.ai/provider-logos/gradium.svg) Gradium 24.2% (21.0–27.4% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) 8.4% 11.6% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) 1.13s (p50–p90 · US East · n=200 · vendor-direct · ±5%) $0.0104 (3 credits/sec on the XS tier, the cheapest paid plan; $0.0068 at the L tier.) Muse Voice Transcribe ![Meta logo](https://speko.ai/provider-logos/meta.png) Meta 14.2% (12.4–16.5% · EdAcc n=283 · 60 speakers, 2 passes, 2026-10-01 · Measured 2026-10-01) not measured (Not measured. Meta ships a prerecorded endpoint; no batch sweep has been run on it. · Measured 2026-09-02) 4.9% (n=1541 · FLEURS test+dev+train · vendor-direct · Measured 2026-10-01) not measured (Measured but withheld: on 9 of 30 clips the speech-end marker lands up to 2.4s after Meta committed the turn, while the transcript is complete -- so the marker, not the model, sets the number. Muse finalizes faster than this corpus can resolve. · Measured 2026-09-02) $0.0030 (Vendor list price: $3.00 per 1,000 minutes, equivalently $0.18/hr. Streaming and batch are priced identically -- the table does not differentiate -- and there is a single tier, no plan or volume rates. · Measured 2026-09-02) Gemini 3.5 Transcribe Live ![Google logo](https://speko.ai/provider-logos/gemini.svg) Google 13.1% (EdAcc n=98 · earlier set · Earlier 98-clip set. The 2026-10-01 re-measure read 31% with 26 of 99 clips empty, so this cell is not on the new set.) not measured not measured 1.28s (p50–p90 · US East · n=200 · vendor-direct · ±6%) ~$0.0090 (ESTIMATE, not a published per-minute rate. Google prices this model per token -- $3.50/1M audio in, $21.00/1M text out -- and the ~$0.005/min + ~$0.004/min figures on the pricing page are Google's own estimate at 25 audio tokens/sec input and 175 text tokens/min output, for a blended ~$0.009/min. Speech density moves it, so read it as an order of magnitude rather than a rate card. Every other price on this column is the vendor's published per-minute rate.) MAI-Transcribe-2-Streaming ![Azure logo](https://speko.ai/provider-logos/azure.svg) Azure not measured not measured not measured not measured not measured ink-whisper ![Cartesia logo](https://speko.ai/provider-logos/cartesia.svg) Cartesia not measured not measured not measured not measured not measured gemini-3.5-transcribe Batch only ![Google logo](https://speko.ai/provider-logos/gemini.svg) Google not measured not measured not measured not measured not measured s3 ![Hamsa logo](https://speko.ai/provider-logos/hamsa.png) Hamsa not measured not measured not measured not measured not measured velma-2-stt-streaming ![Modulate logo](https://speko.ai/provider-logos/modulate.png) Modulate not measured not measured not measured not measured not measured munsit ![Munsit logo](https://speko.ai/provider-logos/munsit.png) Munsit not measured not measured not measured not measured not measured munsit-en-ar ![Munsit logo](https://speko.ai/provider-logos/munsit.png) Munsit not measured not measured not measured not measured not measured default ![Palabra logo](https://speko.ai/provider-logos/palabra.png) Palabra not measured not measured not measured not measured not measured ## LLM22 Model Provider Id DeepSeek-V4-Flash-0731 ![Baseten logo](https://speko.ai/provider-logos/baseten.png) Baseten 93.16% (83–100% 95% CI bootstrap over items · n=95 · measured 2026-10-04 · Measured 2026-10-04) 558ms (558–1026 p50–p90 · n=38 · Measured 2026-10-04) 88% (159 of 180 calls clean · measured 2026-10-04 · Measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31% · Measured 2026-10-04) $0.26 (Measured 2026-10-04) gpt-4.1-mini ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 92.63% (80–100% 95% CI bootstrap over items · n=95 · measured 2026-10-04) 480ms (480–755 p50–p90 · n=38) 69% (125 of 180 calls clean · measured 2026-10-04) 3% (1 of 30 probe runs) $1.60 gpt-5.4-mini ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 91.75% (84–98% 95% CI bootstrap over items · n=95 · measured 2026-10-04 · Measured 2026-10-04) 390ms (390–477 p50–p90 · n=38 · Measured 2026-10-04) 80% (144 of 180 calls clean · measured 2026-10-04 · Measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31% · Measured 2026-10-04) $4.50 (Vendor list, standard tier, per 1M output tokens ($0.75 in). · Measured 2026-10-04) gpt-4.1 ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 89.12% (75–100% 95% CI bootstrap over items · n=95 · measured 2026-10-04) 462ms (462–605 p50–p90 · n=38) 78% (140 of 180 calls clean · measured 2026-10-04) 10% (3 of 30 probe runs) $8.00 gemini-3.7-flash ![Google logo](https://speko.ai/provider-logos/gemini.svg) Google 86.84% (71–100% 95% CI bootstrap over items · n=95 · measured 2026-10-04 · Measured 2026-10-04) 1.38s (1.38–2.65s p50–p90 · n=38 · Measured 2026-10-04) 85% (153 of 180 calls clean · measured 2026-10-04 · Measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31% · Measured 2026-10-04) $3.75 (Vendor list, paid tier, per 1M output tokens ($0.75 in); rises to $7.50 on 2027-01-01. · Measured 2026-10-04) Qwen3.8-27B ![Cerebras logo](https://speko.ai/provider-logos/cerebras.png) Cerebras 86.14% (72–97% 95% CI bootstrap over items · n=95 · measured 2026-10-04) 246ms (246–592 p50–p90 · n=38) 86% (154 of 180 calls clean · measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31%) $1.49 (Cerebras list pricing for this model, $0.99/1M input and $1.49/1M output, from the vendor’s own model page (inference-docs.cerebras.ai/models/qwen-3.8-27b, read 2026-09-06). The same weights cost $2.55/1M output on the Alibaba row.) gemini-3.8-flash ![Google logo](https://speko.ai/provider-logos/gemini.svg) Google 85.79% (71–98% 95% CI bootstrap over items · n=95 · measured 2026-10-04 · Measured 2026-10-04) 1.21s (1.21–2.90s p50–p90 · n=38 · Measured 2026-10-04) 87% (157 of 180 calls clean · measured 2026-10-04 · Measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31% · Measured 2026-10-04) $3.75 (Vendor list, paid tier, per 1M output tokens; rises to $7.50 on 2027-01-01. · Measured 2026-10-04) gemma-4-31b ![Hopper logo](https://speko.ai/provider-logos/hopper.svg) Hopper 85.26% (68–100% 95% CI bootstrap over items · n=95 · measured 2026-10-08 · Measured 2026-10-08) 150ms (150–158 p50–p90 · n=38 · Measured 2026-10-08) 87% (156 of 180 calls clean · measured 2026-10-08 · Measured 2026-10-08) 0% (0 of 10 probes (30 probe runs) · ≤31% · Measured 2026-10-08) $1.20 (Vendor list, per 1M output tokens ($0.40 in, $0.20 cached). · Measured 2026-10-08) DeepSeek-V4.1-Flash (Together) ![Together logo](https://speko.ai/provider-logos/together.png) Together 84.21% (69–97% 95% CI bootstrap over items · n=95 · measured 2026-10-04 · Measured 2026-10-04) 157ms (157–793 p50–p90 · n=38 · Measured 2026-10-04) 86% (155 of 180 calls clean · measured 2026-10-04 · Measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31% · Measured 2026-10-04) $1.20 (Vendor list, serverless, per 1M output tokens ($0.30 in). · Measured 2026-10-04) gpt-5.6-terra ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 83.51% (68–96% 95% CI bootstrap over items · n=95 · measured 2026-10-04 · Measured 2026-10-04) 580ms (580–731 p50–p90 · n=38 · Measured 2026-10-04) 85% (153 of 180 calls clean · measured 2026-10-04 · Measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31% · Measured 2026-10-04) $12.00 (Measured 2026-10-04) Claude Sonnet 5 ![Anthropic logo](https://speko.ai/provider-logos/anthropic.png) Anthropic 83.33% (68–95% 95% CI bootstrap over items · n=95 · measured 2026-10-04) 1.47s (1.47–2.74s p50–p90 · n=38) 89% (161 of 180 calls clean · measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31%) $10.00 (Introductory rate through 2026-08-31; $15.00 from 2026-09-01.) DeepSeek-V4.1-Flash (Baseten) ![Baseten logo](https://speko.ai/provider-logos/baseten.png) Baseten 82.19% (69–94% 95% CI bootstrap over items · n=95 · measured 2026-10-04 · Measured 2026-10-04) 458ms (458–1062 p50–p90 · n=38 · Measured 2026-10-04) 88% (158 of 180 calls clean · measured 2026-10-04 · Measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31% · Measured 2026-10-04) $1.20 (Measured 2026-10-04) gpt-oss-120b (Cerebras) ![Cerebras logo](https://speko.ai/provider-logos/cerebras.png) Cerebras 82.02% (70–93% 95% CI bootstrap over items · n=95 · measured 2026-10-04) 139ms (139–261 p50–p90 · n=38) 53% (95 of 180 calls clean · measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31%) $0.75 Claude Haiku 4.5 ![Anthropic logo](https://speko.ai/provider-logos/anthropic.png) Anthropic 81.67% (65–94% 95% CI bootstrap over items · n=95 · measured 2026-10-04) 391ms (391–574 p50–p90 · n=38) 74% (134 of 180 calls clean · measured 2026-10-04) 3% (1 of 30 probe runs) $5.00 gpt-oss-120b (Baseten) ![Baseten logo](https://speko.ai/provider-logos/baseten.png) Baseten 80.53% (66–94% 95% CI bootstrap over items · n=95 · measured 2026-10-04 · Measured 2026-10-04) 504ms (504–1215 p50–p90 · n=38 · Measured 2026-10-04) 73% (132 of 180 calls clean · measured 2026-10-04 · Measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31% · Measured 2026-10-04) $0.50 (Measured 2026-10-04) gpt-5.6-luna ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 75.61% (58–91% 95% CI bootstrap over items · n=95 · measured 2026-10-04 · Measured 2026-10-04) 546ms (546–683 p50–p90 · n=38 · Measured 2026-10-04) 78% (140 of 180 calls clean · measured 2026-10-04 · Measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31% · Measured 2026-10-04) $1.20 (Measured 2026-10-04) gpt-5.2 ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 75.61% (59–91% 95% CI bootstrap over items · n=95 · measured 2026-10-04 · Measured 2026-10-04) 664ms (664–825 p50–p90 · n=38 · Measured 2026-10-04) 87% (157 of 180 calls clean · measured 2026-10-04 · Measured 2026-10-04) 10% (3 of 30 probe runs · Measured 2026-10-04) $14.00 (Vendor list, standard tier, per 1M output tokens ($1.75 in). · Measured 2026-10-04) Llama-3.3-70B ![Together logo](https://speko.ai/provider-logos/together.png) Together 54.30% (39–70% 95% CI bootstrap over items · n=95 · measured 2026-10-04) 525ms (525–725 p50–p90 · n=38) 57% (103 of 180 calls clean · measured 2026-10-04) 0% (0 of 10 probes (30 probe runs) · ≤31%) $1.04 google/gemma-4-31B-it ![DeepInfra logo](https://speko.ai/provider-logos/deepinfra.png) DeepInfra not measured not measured not measured not measured not measured openai/gpt-oss-120b ![Together logo](https://speko.ai/provider-logos/together.png) Together not measured not measured not measured not measured not measured Qwen/Qwen3.5-9B ![Together logo](https://speko.ai/provider-logos/together.png) Together not measured not measured not measured not measured not measured zai-org/GLM-5.3-Flash ![Together logo](https://speko.ai/provider-logos/together.png) Together not measured not measured not measured not measured not measured ## Text-to-speech39 Model Provider Id eleven_v3_conversational ![ElevenLabs logo](https://speko.ai/provider-logos/elevenlabs.png) ElevenLabs 1590 (66%) 266ms (265–268 two-sweep p50) $50.0 (v3 Conversational is its own $0.05/1k product on elevenlabs.io/pricing/api, half the plain v3 rate. The FAQ on that same page collapses the two at $0.10/1k; the itemized card and the per-plan table both say $0.05, so they win.) eleven_v4 ![ElevenLabs logo](https://speko.ai/provider-logos/elevenlabs.png) ElevenLabs ~1708 (Measured 2026-09-28) 844ms (798–864 three-sweep p50 · Measured 2026-09-28) $80.0 (elevenlabs.io/pricing/api lists $0.08/1k, discounted 72% to $0.022/1k until 2026-10-12. · Measured 2026-09-28) eleven_v4_turbo ![ElevenLabs logo](https://speko.ai/provider-logos/elevenlabs.png) ElevenLabs ~1670 (Measured 2026-09-28) 205ms (203–213 three-sweep p50 · Measured 2026-09-28) $40.0 (elevenlabs.io/pricing/api lists $0.04/1k, discounted 72% to $0.011/1k until 2026-10-12. · Measured 2026-09-28) gemini-3.8-flash-tts ![Google logo](https://speko.ai/provider-logos/gemini.svg) Google ~1642 (Measured 2026-09-24) 1.03s (1015–1046 two-sweep p50 · Measured 2026-09-24) ~$15.0 (Token-billed: $0.50/1M text in, $9/1M audio out until 2026-12-31, then $1 / $18 (ai.google.dev/gemini-api/docs/pricing). At 25 audio tokens/sec and 900 chars/min of speech that is ~$15 per 1M chars, ~$30 from 2027-01-01. · Measured 2026-09-24) gemini-3.8-flash-lite-tts ![Google logo](https://speko.ai/provider-logos/gemini.svg) Google ~1639 (Measured 2026-09-24) 497ms (483–511 two-sweep p50 · Measured 2026-09-24) ~$10.0 (Token-billed: $0.50/1M text in, $6/1M audio out until 2026-12-31, then $1 / $12 (ai.google.dev/gemini-api/docs/pricing). At 25 audio tokens/sec and 900 chars/min of speech that is ~$10 per 1M chars, ~$20 from 2027-01-01. · Measured 2026-09-24) flux ![Deepgram logo](https://speko.ai/provider-logos/deepgram.png) Deepgram ~1550 (An estimate, marked by the board itself: the vendor publishes no exact figure.) 106ms (106–148 p50–p90 · Mean of two 2026-08-12 sweeps (104/108 p50), n=30 each, concurrency 1, one keep-alive connection, 0 rejected. Anchors in the same sweeps came in 11–26% off their published cells, so the column needs a full re-baseline before these are read against each other.) $45.0 (Deepgram published list, Pay As You Go: $0.0450 per 1,000 characters = $45/1M. Same unit that puts Aura-2 at $0.030/1k = the $30 cell above, so the two are directly comparable — Flux TTS costs 50% more than Aura-2. Growth tier is $0.0405/1k ($40.5/1M). Free to build against through 2026-09-12 (45 concurrent streams globally, 5 EU/AU); billing starts 09-13.) sonic-3.5 ![Cartesia logo](https://speko.ai/provider-logos/cartesia.svg) Cartesia 1574 (65%) 121ms (121–236 p50–p90) $50.0 (About 1 credit per character on Pro, the cheapest paid plan, at its included rate: $5 for 100K credits = $50 per 1M characters.) sonic-3.6 ![Cartesia logo](https://speko.ai/provider-logos/cartesia.svg) Cartesia ~1578 (Measured 2026-08-28) 120ms (120–135 p50–p90 · US East · vendor-direct · n=400 · Vendor-direct from US East, n=400 (2 sweeps x 200), concurrency 1, one keep-alive socket. · Measured 2026-08-28) $50.0 (About 1 credit per character on Pro, the cheapest paid plan, at its included rate: $5 for 100K credits = $50 per 1M characters. · Measured 2026-08-28) tts-rt-v2 ![Soniox logo](https://speko.ai/provider-logos/soniox.png) Soniox ~1606 (Measured 2026-08-13) 381ms (381–417 p50–p90 · Measured 2026-08-13) ~$13.0 (Token-billed: $4/1M text in, $21.50/1M audio out; vendor-stated ~$0.70 per hour of speech (soniox.com/pricing). At 900 chars/min (54,000 chars/hour) that is ~$13 per 1M chars. Same rates as v1, which the pricing page no longer lists. · Measured 2026-08-13) simba-3.2 ![Speechify logo](https://speko.ai/provider-logos/speechify.png) Speechify 1573 (63%) 85ms (85–103 p50–p90 · Re-measured 2026-09-21: two n=30 sweeps at concurrency 1 from US East, p50 82 and 87, p90 102 and 104. The pair missed the ≥90%-within-10% publish gate at 3/4 — the one failure is the co-swept deepgram:aura-2 anchor (26%), not this row, which agreed to 5.9% here and 0.0% on an earlier pair. Eight sweeps across two regions span 82–90ms, and a vendor-direct probe reads 56ms from the same region.) $10.0 (Vendor list (Starter tier); $6–8 per 1M at Pro/Scale volume.) inworld-tts-2 ![Inworld logo](https://speko.ai/provider-logos/inworld.png) Inworld 1561 (63%) 116ms (116–125 p50–p90) $25.0 (On-demand. Drops to $20 on Creator and $12.50 on Growth.) s2.1-pro ![Fish Audio logo](https://speko.ai/provider-logos/fish.png) Fish Audio ~1566 (An estimate, marked by the board itself: the vendor publishes no exact figure.) 185ms (185–235 p50–p90) $15.0 (Billed per UTF-8 byte, so this holds for Latin text only — CJK runs ~3x. The s2.1-pro-free tier is $0 through 2026-08-31.) lightning_v3.1 ![Smallest AI logo](https://speko.ai/provider-logos/smallest.png) Smallest AI 1544 (58%) 173ms (173–193 p50–p90) $17.5 (Pay-as-you-go; $0.175/10k chars (smallest.ai/pricing/models).) grok-tts ![xAI logo](https://speko.ai/provider-logos/xai.png) xAI 1495 (48%) 272ms (272–325 p50–p90) $15.0 Gradium TTS ![Gradium logo](https://speko.ai/provider-logos/gradium.svg) Gradium ~1592 (Measured 2026-10-01) 105ms (Measured 2026-10-01 on voice Brooklyn: three n=30 sweeps at p50 107, 103 and 105. The anchor rows matched their published cells in only 6 of 9 readings, so the formal two-sweep gate did not pass. · Measured 2026-10-01) $57.8 (1 credit/char on the XS tier, the cheapest paid plan; $35.9 at the L tier. · Measured 2026-10-01) gpt-4o-mini-tts ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 1424 (37%) 691ms (691–1171 p50–p90) ~$20.0 (Token-billed: $0.60/1M text in, $12/1M audio out; OpenAI publishes no per-character or per-minute rate. At ~25 audio tokens/sec (not OpenAI-published) and 900 chars/min, ~$20 per 1M chars. Legacy tts-1 is $15, tts-1-hd $30 per 1M chars.) speech-2.8-hd ![MiniMax logo](https://speko.ai/provider-logos/minimax.png) MiniMax 1431 (36%) 294ms (294–354 p50–p90) $100.0 (HD tier; speech-2.8-turbo ≈ $60/1M.) palabra-tts-v1 ![Palabra logo](https://speko.ai/provider-logos/palabra.png) Palabra ~1520 (An estimate, marked by the board itself: the vendor publishes no exact figure.) 72ms (72–79 p50–p90) $12.0 Qwen3-TTS Fast ![Nari logo](https://speko.ai/provider-logos/nari.png) Nari ~1536 (Measured 2026-09-12) 64ms (64–65 p50–p90 · Four n=30 sweeps from US East on 2026-09-26 read p50 62, 65, 64 and 63. Both sweep pairs missed the publish gate on unstable anchors (cartesia:sonic-3.5 and deepgram:aura-2), and smallest:lightning_v3.1 failed over on every call. This row stayed within 5%. · Measured 2026-09-26) $10.0 (Vendor list, Partner tier (docs.narilabs.com/models-and-pricing, read 2026-09-13). Nari also lists a `qwen3-tts-fast:free` tier at $0; this row measured the PAID `qwen3-tts-fast` id, so the paid rate is the one published. The standard (non-fast) qwen3-tts is $5/1M. · Measured 2026-09-12) Qwen3-TTS ![Nari logo](https://speko.ai/provider-logos/nari.png) Nari ~1510 (Measured 2026-09-26) 75ms (75–79 p50–p90 · Four n=30 sweeps from US East on 2026-09-26 read p50 74, 76, 76 and 75, in the same runs as qwen3-tts-fast and with the same anchor misses. This row stayed within 3%. · Measured 2026-09-26) $5.0 (Vendor list $5 per 1M input characters (docs.narilabs.com/models-and-pricing, read 2026-09-26). · Measured 2026-09-26) Qwen3-TTS ![Hopper logo](https://speko.ai/provider-logos/hopper.svg) Hopper ~1406 (Measured 2026-10-08) not measured (Measured 2026-10-08) ~$5.7 (Billed at $0.30 per audio hour. At the 14.7 characters per second measured on voice hudson, that is about $5.67 per 1M characters. · Measured 2026-10-08) paxa-tts-flash-v1 ![Paxa Labs logo](https://speko.ai/provider-logos/paxa.png) Paxa Labs ~1325 (Measured 2026-10-05) 389ms (389–445 p50–p90 · Two n=100 sweeps from US East on 2026-10-10 read p50 372 and 405. Cartesia reproduced its cell in the same runs; Deepgram ran 9–12% slow and Inworld 10% fast. · Measured 2026-10-10) $10.0 (Paxa bills speech at 10 credits per 1,000 characters, $10 per one million characters (paxalabs.com/docs/credits, read 2026-10-06). A credit costs $0.001 on refill (10,000 for $10), and live calls on 2026-10-06 were charged 0.01 credit per character. This row read $15 from the 2026-09-12 pricing page until then. Billing is on input characters. · Measured 2026-09-28) eleven_flash_v2_5 ![ElevenLabs logo](https://speko.ai/provider-logos/elevenlabs.png) ElevenLabs ~1494 (Measured 2026-09-08) not measured $50.0 (API rate card (elevenlabs.io/pricing/api, 2026-09-08): Flash and Turbo $0.05 per 1K characters; v3 and Multilingual v2 $0.10. · Measured 2026-09-08) eleven_multilingual_v2 ![ElevenLabs logo](https://speko.ai/provider-logos/elevenlabs.png) ElevenLabs ~1472 (Measured 2026-09-08) not measured $100.0 (API rate card (elevenlabs.io/pricing/api, 2026-09-08): Multilingual v2 $0.10 per 1K characters, the v3 rate. · Measured 2026-09-08) eleven_flash_v2 ![ElevenLabs logo](https://speko.ai/provider-logos/elevenlabs.png) ElevenLabs ~1507 (Measured 2026-09-08) not measured $50.0 (API rate card (elevenlabs.io/pricing/api, 2026-09-08): Flash and Turbo $0.05 per 1K characters. · Measured 2026-09-08) sonic-3 ![Cartesia logo](https://speko.ai/provider-logos/cartesia.svg) Cartesia ~1551 (Measured 2026-09-08) not measured $50.0 (About 1 credit per character on Pro, the cheapest paid plan, at its included rate: $5 for 100K credits = $50 per 1M characters. · Measured 2026-09-08) coda ![Rime logo](https://speko.ai/provider-logos/rime.png) Rime ~1535 (Measured 2026-09-08) not measured $50.0 (Starter tier, rime.ai/pricing 2026-09-08: Coda $0.05 / 1K characters. · Measured 2026-09-08) mistv3 ![Rime logo](https://speko.ai/provider-logos/rime.png) Rime ~1428 (Measured 2026-09-08) not measured $30.0 (Starter tier, rime.ai/pricing 2026-09-08: Mist v3 $0.03 / 1K characters. · Measured 2026-09-08) speech-2.6-turbo ![MiniMax logo](https://speko.ai/provider-logos/minimax.png) MiniMax ~1408 (Measured 2026-09-08) not measured ~$60.0 (Estimate. The vendor pricing page lists point plans, not per-model rates (2026-09-08); Turbo taken at the ≈$60 the speech-2.8-hd row cites for its turbo sibling. · Measured 2026-09-08) lightning_v3.1_pro ![Smallest AI logo](https://speko.ai/provider-logos/smallest.png) Smallest AI ~1447 (Measured 2026-09-08) not measured $19.5 (Pay-as-you-go; $0.195/10k chars (smallest.ai/pricing/models, corroborated by the model card). · Measured 2026-09-08) speech-2.8-turbo ![MiniMax logo](https://speko.ai/provider-logos/minimax.png) MiniMax ~1385 (Measured 2026-09-29) not measured $60.0 (MiniMax's pay-as-you-go rate for speech-2.8-turbo, the price the router bills. · Measured 2026-09-29) MAI-Voice-2.1 ![Azure logo](https://speko.ai/provider-logos/azure.svg) Azure not measured not measured not measured MAI-Voice-2.1-Flash ![Azure logo](https://speko.ai/provider-logos/azure.svg) Azure not measured not measured not measured eleven_v3 ![ElevenLabs logo](https://speko.ai/provider-logos/elevenlabs.png) ElevenLabs not measured not measured not measured drama-3 ![Fish Audio logo](https://speko.ai/provider-logos/fish.png) Fish Audio not measured not measured not measured chirp-3-hd ![Google logo](https://speko.ai/provider-logos/google.png) Google not measured not measured not measured default ![Hamsa logo](https://speko.ai/provider-logos/hamsa.png) Hamsa not measured not measured not measured faseeh-v1-preview ![Munsit logo](https://speko.ai/provider-logos/munsit.png) Munsit not measured not measured not measured paxa-tts-flash-v2 ![Paxa Labs logo](https://speko.ai/provider-logos/paxa.png) Paxa Labs not measured not measured not measured ## Speech-to-speech11 Model Provider Id Price as published grok-voice-3.0 ![xAI logo](https://speko.ai/provider-logos/xai.png) xAI 0.90 (Measured 2026-10-22) 0.85 (Measured 2026-10-22) 851ms (1271·— · Measured 2026-10-22) $0.08 · min (Measured 2026-10-22) gpt-live-1 ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 0.88 (Measured 2026-10-22) 0.90 (Measured 2026-10-22) 1.48s (2324·— · Measured 2026-10-22) $0.05 · min (Measured 2026-10-22) gpt-realtime-2 ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 0.85 (Measured 2026-10-22) 0.78 (Measured 2026-10-22) 1.10s (1689·— · Measured 2026-10-22) $32 / $64 · 1M tok (Measured 2026-10-22) gpt-realtime-2.1 ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 0.84 (Measured 2026-10-22) 0.78 (Measured 2026-10-22) 1.02s (1491·— · Measured 2026-10-22) $32 / $64 · 1M tok (Measured 2026-10-22) grok-voice-think-fast-2.0 ![xAI logo](https://speko.ai/provider-logos/xai.png) xAI 0.82 (Measured 2026-10-22) 0.68 (Measured 2026-10-22) 723ms (1014·— · Measured 2026-10-22) $0.08 · min (Measured 2026-10-22) gemini-3.8-live ![Google logo](https://speko.ai/provider-logos/google.png) Google 0.79 (Measured 2026-10-22) 0.68 (Measured 2026-10-22) 1.24s (1992·— · Measured 2026-10-22) $0.005 · min (Measured 2026-10-22) gpt-realtime-2.1-mini ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI 0.76 (Measured 2026-10-22) 0.60 (Measured 2026-10-22) 1.14s (1960·— · Measured 2026-10-22) $10 / $20 · 1M tok (Measured 2026-10-22) gemini-3.8-live-extended-thinking ![Google logo](https://speko.ai/provider-logos/google.png) Google 0.75 (Measured 2026-10-22) 0.72 (Measured 2026-10-22) 863ms (1325·— · Measured 2026-10-22) $0.005 · min (Measured 2026-10-22) qwen3.8-livetranslate-flash-realtime Translation ![Alibaba logo](https://speko.ai/provider-logos/alibaba.png) Alibaba not measured not measured not measured not measured gemini-3.5-live-translate-preview Translation ![Google logo](https://speko.ai/provider-logos/google.png) Google not measured not measured not measured not measured gpt-realtime-translate Translation ![OpenAI logo](https://speko.ai/provider-logos/openai.svg) OpenAI not measured not measured not measured not measured ## Published stacks What benchmarks.speko.ai picks for each job, and the measurement that decided it. Use case Speech-to-text LLM Text-to-speech Default decided on 8.5% WER · 82% task success · Elo 1670 Muse Voice Transcribe meta:muse-voice-transcribe-1.0 DeepSeek-V4.1-Flash baseten:deepseek-ai/DeepSeek-V4.1-Flash eleven_v4_turbo elevenlabs:eleven_v4_turbo Real-time phone agent decided on 301ms · 558ms · 84ms Muse Voice Transcribe meta:muse-voice-transcribe-1.0 DeepSeek-V4-Flash-0731 baseten:deepseek-ai/DeepSeek-V4-Flash-0731 inworld-tts-2 inworld:inworld-tts-2 Accuracy-critical decided on 7.2% WER · 93% task success · 0.90 robustness Universal-3.5 Pro assemblyai:universal-3-5-pro DeepSeek-V4-Flash-0731 baseten:deepseek-ai/DeepSeek-V4-Flash-0731 eleven_v4_turbo elevenlabs:eleven_v4_turbo Lowest cost decided on $0.0008/min · $0.0002/min · $0.0090/min Velma 2 modulate:velma-2-stt-streaming-english-v2 DeepSeek-V4-Flash-0731 baseten:deepseek-ai/DeepSeek-V4-Flash-0731 simba-3.2 speechify:simba-3.2