Audio

Xiaomi MiMo-V2.6-Flash is Xiaomi MiMo's September 21, 2026 efficiency-balanced open-weights MoE (309B total / 15B active), sibling of MiMo-V2.6-Pro. First-party HF checkpoint XiaomiMiMo/MiMo-V2.6-Flash-RL (MIT; ungated; 1M context) is the released Flash RL weights; product/API id is mimo-v2.6-flash. First-party news: the series comprises two native fully multimodal models (Pro and Flash); Flash comprehensively outperformed MiMo-V2.5-Pro. Product card: full-modality high-intelligence low-cost reasoning for high-frequency and large-scale professional workflows; context 1M; max output 128K; input text, image, video, and audio. Overseas API pricing (pay-as-you-go): $0.0028 cache-hit / $0.14 cache-miss input and $0.28 output per 1M tokens; batch $0.0014 / $0.07 / $0.14. Sparse MoE MiMoV2ForCausalLM; 48 layers (39 SWA / 9 GA); 256 routed experts / 8 activated per token; hybrid SWA+GA backbone; native omnimodal (text/image/video/audio); 681M MiMo ViT vision encoder; 5-layer SWA MTP/DFlash speculative decoder.

$0.140 in / $0.280 out / 1M tokens

2 routes · 1 batch · 2 cache

Xiaomi MiMo-V2.6-Pro is Xiaomi MiMo's September 21, 2026 flagship open-weights MoE (1.02T total / 42B active), succeeding proprietary MiMo-V2.5-Pro. First-party HF checkpoint XiaomiMiMo/MiMo-V2.6-Pro-RL (MIT; ungated; 1M context) is the released Pro RL weights; product/API id is mimo-v2.6-pro. HF README: native omnimodal + long-horizon agents; You Only RL Once mixed RL across coding/agents/visual/cybersecurity. Official product page (mimo.mi.com) and @XiaomiMiMo announce open-sourcing of Pro and Flash with tech report. Overseas API pricing matches V2.5-Pro: $0.0036 cache-hit / $0.435 cache-miss input and $0.87 output per 1M tokens (pay-as-you-go). Sparse MoE MiMoV2ForCausalLM; 70 layers (60 SWA / 10 GA); 384 routed experts / 8 activated per token; hybrid SWA+GA backbone; native omnimodal (text/image/video/audio); 681M MiMo ViT vision encoder; MTP blocks.

$0.435 in / $0.870 out / 1M tokens

2 routes · 1 batch · 2 cache

Gemini Omni 1.2 Flash is Google DeepMind's next-generation video generation and editing model, generally available November 5, 2026 on the Gemini API and Google Cloud. It generates and edits clips with native audio from 3 to 15 seconds (360p/720p/1080p/4K), supports multi-turn conversational editing, reference images/video/audio.

$1.50 in / $9.00 out / 1M tokens

5 routes

Seedance 2.0 Mini is the separately addressable compact Volcengine surface for ByteDance's multimodal joint audio-video generation family.

Pricing not tracked / 1M tokens

1 route

Seedance 2.0 Fast is the separately addressable speed-optimized Volcengine surface for ByteDance's multimodal joint audio-video generation family.

Pricing not tracked / 1M tokens

1 route

Muse Ember is Meta Superintelligence Labs' smallest and most efficient Muse model, introduced November 9, 2026, for the work agents repeat all day: tool calls, coding subtasks, routing, extraction and long-document parsing. Meta says it is the first model trained using Muse Wonder Full. It takes text, image, audio and video input and returns text, with a 1M-token context window and up to 512K output tokens, and supports the Responses API. It is available on Muse Code and the Meta Model API as muse-ember-1.0 at $0.10 per million input tokens, $0.004 per million cached input tokens and $0.50 per million output tokens.

$0.100 in / $0.500 out / 1M tokens

3 routes · 3 cache

HeyGen Video is HeyGen's general-purpose video model, built on MiniMax H3 and post-trained by HeyGen. It generates 5-15 s clips with their own dialogue, ambience and sound-effects track from text, a first-frame image, or image/video references, at 480p to 2K.

Pricing not tracked / 1M tokens

3 routes

HeyGen Voice is HeyGen's text-to-speech model (API id heygen-voice-1), supporting public, instant and professional cloned voices. Announced 9 Oct 2026.

$30.00 in / - out / 1M tokens

3 routes

Kling 4.0 Flash is Kuaishou's fast, lower-cost video generation model in the Kling 4.0 series, first shown alongside Kling 4.0 on October 6, 2026 and made fully available on the Kling AI website and API on November 4, 2026. It does text-to-video, image-to-video, omni reference and video editing, producing 3 to 20-second clips at 720p in 8-bit SDR.

Pricing not tracked / 1M tokens

2 routes

SFX v3 is ElevenLabs' sound effects model, announced October 29, 2026. It generates sound effects from a text prompt and supports seamless looping. It is available in ElevenCreative and on the ElevenLabs API as eleven_text_to_sound_v3 through the same Sound Effects API as v2, and on fal for clips up to 30 seconds.

Pricing not tracked / 1M tokens

2 routes

AssemblyAI-TTS-1 is AssemblyAI's first text-to-speech model, built for voice agents. It streams natural speech with about 90 ms latency, supports English, Spanish, French, German, Italian and Portuguese.

$15.00 in / - out / 1M tokens

2 routes

Nemotron 3.5 Transcribe is NVIDIA's open-weight, promptable speech model: a mixture-of-experts model with 34B total and 4.3B active parameters (including the audio encoder) that does transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding in one checkpoint. It handles recordings up to one hour (65,536-token context) and is released under the OpenMDW 1.1 license for commercial or non-commercial use.

Pricing not tracked / 1M tokens

2 routes

HeyGen Voice Turbo (API id heygen-voice-1-turbo) is the faster HeyGen Voice text-to-speech model for public and instant voices; a shared-capacity Turbo Free variant is free until 30 Nov 2026.

$4.00 in / - out / 1M tokens

2 routes

OmniHuman 2.0 is ByteDance's audio-driven avatar video model, offered through BytePlus. From one image and a speech, music or dialogue track (plus optional text prompts for motion and camera), it generates lip-synced video with expressions, gestures and multi-person scenes at native 720p, for 30 to 300 seconds in a single take.

Pricing not tracked / 1M tokens

2 routes

Phonic-1 is Phonic's speech-to-speech model for agentic voice tasks, released October 29, 2026. It hears the caller's audio directly and answers in natural speech, and can call tools while the caller is still talking. It is available through Phonic's API under the model ID phonic-1.

Pricing not tracked / 1M tokens

1 route

Pika H3 Flex is Pika's open-weight omni video model family, released October 29, 2026 and post-trained by the Pika team on MiniMax H3 down to six-step generation. It makes 5 to 15 second videos with synchronized audio from text, images, or video and audio references. Open weights on Hugging Face (LoRA adapters plus Diffusers-format MiniMax H3 base components, under the MiniMax H3 Community License) run natively at 480p and 768p; the Pika API also upscales to 1080p, 2K and 4K and can return five separate audio stems.

Pricing not tracked / 1M tokens

1 route

Kling 4.0 is Kuaishou's next-generation native multimodal video model, first shown by Kling AI in creator posts on October 10, 2026, detailed in the Kling 4.0 release overview on October 14 and made fully available in the Kling AI app and platform on October 15. It improves stability for complex motion and consistency across characters and scenes, generates synchronized audio, voices and dialogue, and supports text-, image- and reference-to-video plus video editing at 720p and 1080p, with 4K/1080p 10-bit HDR output listed as coming soon. On fal it runs in the playground and fal Agent at $0.157 per second at 720p and $0.235 per second at 1080p, with API access listed as coming soon; Kling opened Kling 4.0 API applications on October 21.

Pricing not tracked / 1M tokens

1 route

Moondream Parakeet Ultra is a speech-to-text model from Moondream (M87 Labs), released September 22, 2026 with Parakeet Redux. It is a post-trained version of NVIDIA's parakeet-tdt-0.6b-v3 with the same architecture, tokenizer and 0.6B parameters, keeping full-precision weights. Weights are on Hugging Face under CC BY 4.0, and Prime Intellect, which acquired Moondream, serves it on OpenRouter.

Pricing not tracked / 1M tokens

1 route

Cosmos 3 Nano is NVIDIA's 16B-parameter omnimodel optimized for efficient inference on workstation-grade hardware (NVIDIA RTX PRO 6000). Architecture: dual-tower Mixture-of-Transformers with an 8B autoregressive Reasoner and an 8B diffusion-based Generator. The Reasoner supports up to 256K tokens of context for vision-language reasoning; the Generator produces video up to 720p at variable frame rates (default 189 frames). Natively handles text, image, video, audio (48kHz stereo), and robot action trajectories across 10+ robot embodiments including Franka Panda, UR, Google robot, and UMI. BF16 precision only. Available as open weights on Hugging Face and via the Cosmos 3 Reasoner NIM (NIM_MODEL_SIZE=nano). Intended for real-time robotics inference and edge-adjacent deployment. Robot action input/output is preserved in this description because the model schema does not have a dedicated action modality field.

Pricing not tracked / 1M tokens

1 route

Cosmos 3 Super is NVIDIA's flagship 64B-parameter omnimodel for physical AI, designed for large-scale synthetic data generation and high-fidelity simulation on NVIDIA Hopper and Blackwell datacenter GPUs. Architecture: dual-tower Mixture-of-Transformers with a 32B autoregressive Reasoner and a 32B diffusion-based Generator. Supports 256K token reasoning context, 720p video generation at variable frame rates, and 10+ robot embodiment action domains. Ranked #1 among open models on Physics-IQ, PAI-Bench, R-Bench, RoboLab, RoboArena, VANTAGE-Bench, and TAR (Computex 2026). Training data: 1.3B data points across 393 datasets (2024-2026). Inference performance (vLLM-Omni): ~55s for 50-step video on 8xH200. Available as open weights on Hugging Face and via Cosmos 3 Reasoner NIM (NIM_MODEL_SIZE=super).

Pricing not tracked / 1M tokens

1 route

Seedance 2.0 is ByteDance Seed's flagship joint audio-video generation model. It accepts text, image, audio and video references and generates controllable, synchronized audiovisual output.

Pricing not tracked / 1M tokens

1 route

ZONOS3-Mini is Zyphra's compact open-source text-to-speech model with voice cloning, released November 9, 2026 under Apache 2.0. It is a sparse mixture-of-experts model with 800 million total and 100 million active parameters, trained on 112 NVIDIA H100s in 2.5 days, and runs in real time on phones and laptops, including in the browser through WebGPU. It clones a voice from an audio sample or designs one from age, accent and vocal texture, steers emotion, pace and delivery down to individual words, and outputs 44.1 kHz audio. Zyphra also serves it on Zyphra Cloud.

No tracked provider route

Smaug-Audio-Decisions-0.7B is Abacus.AI's 744M open-weight audio decision model, released November 2, 2026 alongside Smaug-Decisions. A fine-tune of Google's EmbeddingGemma 2 with its audio tower, it answers a typed question about an audio clip with a probability per option from one forward pass; both models ship inside one iOS and Android app. Weights are on Hugging Face at abacusai/Smaug-Audio-Decisions-0.7B under Apache 2.0.

No tracked provider route

Smaug-Decisions is Abacus.AI's open-weight 4B decision model, released November 2, 2026. A full fine-tune of Google's Gemma 4 E4B-it, it answers a question about audio, an image or text with a probability per option from one forward pass, without decoding. Weights are on Hugging Face at abacusai/Smaug-Decisions under Apache 2.0, with a LiteRT build for phones.

No tracked provider route

d1-Omni-2B is Liquid AI's open decision model built on LFM3-Omni-2B. It takes the same yes/no, multiple-choice and scale questions as d1-2B over text, images and audio, returning calibrated probabilities in one forward pass, so voice requests can be routed or checked. 128K-token context; open weights under the LFM Open License v1.0.

No tracked provider route

LFM3-Audio-2B is Liquid AI's audio model in the LFM3 family. It takes text and audio in and returns text, 24 kHz speech, or interleaved text and speech, for speech recognition, speech synthesis and voice conversations. 2B total parameters with a 32K-token context window. Open weights under the LFM Open License v1.0.

No tracked provider route

LFM3-Omni-2B is Liquid AI's omni model in the LFM3 family. One 2.17B-parameter model takes text, images, video and audio in and answers in text or speech, so it can listen, see and act on edge hardware. 32K-token context window. Open weights under the LFM Open License v1.0.

No tracked provider route

Hy-SimulMT 30B-A3B is Tencent Hunyuan's open simultaneous translation model: a 30B-parameter mixture-of-experts (3B active) speech model that listens to Chinese or English speech and streams the translation in real time, with optional speech output and voice cloning through a TTS service. The Interleave variant also outputs the source transcript. Released 2026-10-26 under Apache 2.0.

No tracked provider route

Granite Speech 5.0 2B is IBM's compact speech-language model for multilingual ASR and bidirectional speech translation (English, French, German, Spanish, Portuguese, Japanese, plus English-to-Italian and English-to-Mandarin), with keyword biasing and word-level timestamps. Apache 2.0.

No tracked provider route

Moondream Parakeet Redux is a 1.58-bit (ternary) speech-to-text model from Moondream (M87 Labs), released September 22, 2026 with Parakeet Ultra. Built from NVIDIA's parakeet-tdt-0.6b-v3 with the same architecture and tokenizer, it sets every encoder weight to -1, 0 or +1, shrinking the model from 1.2 GB to 178 MB across 25 languages. Weights are on Hugging Face under CC BY 4.0.

No tracked provider route

Granite Speech 5.0 TurboCTC is IBM's 470M-parameter non-autoregressive CTC speech-recognition model for edge devices; first released English-only (25 Aug 2026) and made multilingual on 21 Oct 2026 (English, French, German, Spanish, Portuguese, Italian, Japanese).

No tracked provider route

Cosmos 3 Super Image2Video is a 64B-parameter fine-tuned variant of Cosmos 3 Super specialized for temporally coherent image-to-video generation. Takes a single image (jpg/png/webp at 256p-720p) plus an optional text prompt (up to 4096 tokens) and outputs MP4 video with 5-400 frames (default 189) at up to 720p, with optional muxed AAC stereo audio at 48kHz. Available via Hugging Face Diffusers and vLLM-Omni.

No tracked provider route

Nemotron 3 VoiceChat is NVIDIA AI's Nemotron 3 model with multimodal text and image input. It was released 2026-03-16.

No tracked provider route

GPT-4o Audio Preview (12-17) is OpenAI's GPT-4o Audio model. It offers a 128K-token context window.

No tracked provider route

Cosmos 3 Edge is NVIDIA's announced preview size variant for real-time edge inference in the Cosmos 3 family. It is intended for lower-compute physical AI workloads than Cosmos 3 Nano. NVIDIA has not disclosed parameters, context window, license terms, public weights, or API access as of 2026-06-01; the released Cosmos 3 rows cover Nano, Super, Text2Image, Image2Video, and Nano Policy DROID.

No tracked provider route

Gemini 3 Flash is Google's speed-optimized Gemini 3 model, available in public preview via the Gemini API and Vertex AI. It supports text, image, audio, and video inputs with a 1M token context window and is priced at $0.50 per 1M input tokens and $3.00 per 1M output tokens.

$0.500 in / $3.00 out / 1M tokens

4 routes · 1 cache

IBM Granite 4.0 1B Speech is a multilingual ASR (Automatic Speech Recognition) and AST (Automatic Speech Translation) model. Supports English, French, German, Spanish, Portuguese, and Japanese with higher English ASR accuracy and faster inference via speculative decoding. Features keyword biasing for names and acronyms.

No tracked provider route

Gemini 3.5 Flash is Google DeepMind's generally available Flash model for sustained frontier-level performance on agentic and coding tasks. It supports multimodal inputs, native thinking, tool and function calling, structured outputs, code execution, search grounding, batch processing, and long contexts up to 1M tokens.

$1.50 in / $9.00 out / 1M tokens

4 routes · 2 batch · 4 cache

Audio model for inputs/outputs via Chat Completions API. Replaces deprecated gpt-4o-audio-preview-2024-12-17.

$2.50 in / $10.00 out / 1M tokens

2 routes

GPT Audio Mini is OpenAI's GPT Audio model with multimodal text and image input. It offers a 125K-token context window.

$0.600 in / $2.40 out / 1M tokens

2 routes

Whisper 1 is OpenAI's general-purpose speech recognition API model, based on Whisper large-v2, released March 2023. Supports multilingual transcription across 50+ languages, speech translation into English, and language identification. Priced at $0.006/min of audio (flat per-minute rate). Exposed via /v1/audio/transcriptions and /v1/audio/translations. For new applications, gpt-4o-transcribe offers better accuracy. API ID: whisper-1.

Pricing not tracked / 1M tokens

1 route

MAI-Transcribe-1.5 is Microsoft AI's second-generation speech-to-text transcription model. It supports 43 languages, domain-specific terminology recognition, and Microsoft-reported five-times-faster transcription than competing models while maintaining state-of-the-art accuracy. Streaming support was announced as coming soon at launch.

Pricing not tracked / 1M tokens

1 route

MAI-Voice-2 is Microsoft AI's second-generation text-to-speech and voice synthesis model. It supports natural speech generation across 15+ languages, voice adaptation from short audio samples, a broader emotional range, and built-in safeguards against misuse.

Pricing not tracked / 1M tokens

1 route

StepAudio 2.5 ASR is StepFun's automatic speech recognition model. At 4B parameters, it introduces Multi-Token Prediction (MTP) technology to parallelly predict multiple tokens per decoding step, enabling transcription of 5 minutes of audio in approximately 1 second. Achieves 400% higher throughput and 60% lower latency compared to prior StepFun ASR systems while maintaining state-of-the-art accuracy. Supports Chinese and English; accepts PCM, OGG, MP3, and WAV formats. Available via the StepFun API (model: stepaudio-2.5-asr). Part of the unified StepAudio 2.5 architecture described in arXiv:2605.23463.

Pricing not tracked / 1M tokens

1 route

TTS-1 is OpenAI's text-to-speech model optimized for real-time, low-latency applications, released at DevDay November 2023. Accepts text input up to 4,096 characters and outputs audio in MP3, OPUS, AAC, FLAC, WAV, or PCM. Offers 6 preset voices (alloy, echo, fable, onyx, nova, shimmer). Lower output quality than TTS-1 HD but faster. Priced at $15.00/1M characters. API ID: tts-1.

$15.00 in / - out / 1M tokens

1 route

Eleven Multilingual v2 is ElevenLabs' stable production TTS model, optimized for long-form content in 29 languages with emotionally rich synthesis. Maximum input: 10,000 characters per request. The default high-quality model prior to Eleven v3 GA. Priced at $0.10/1K characters ($100/1M chars). API ID: eleven_multilingual_v2.

$100.00 in / - out / 1M tokens

1 route

TTS-1 HD is OpenAI's high-quality text-to-speech model, optimized for audio quality over speed, released at DevDay November 2023. Uses the same 6 preset voices as TTS-1 but produces noticeably better audio fidelity at twice the per-character cost. Accepts text up to 4,096 characters. Priced at $30.00/1M characters. API ID: tts-1-hd.

$30.00 in / - out / 1M tokens

1 route

GPT-4o Mini TTS is OpenAI's instructable text-to-speech model built on GPT-4o mini, released March 20, 2025. Supports natural language instructions to control tone, style, pacing, and emotion (instructable TTS) — the model follows conversational prompts to adjust delivery rather than relying on static voice presets. OpenAI's recommended cost-efficient TTS for production. Input: text tokens + instructions at $0.60/1M tokens. Output: audio tokens at $12.00/1M tokens. Accepts up to 2,000 input tokens. API ID: gpt-4o-mini-tts.

$0.600 in / - out / 1M tokens

1 route

Eleven Flash v2.5 is ElevenLabs' ultra-low-latency TTS model for real-time applications and voice agents, released December 18, 2024. Achieves ~75ms latency with 32 languages (adds Hungarian, Norwegian, Vietnamese over Flash v2). Maximum input: 40,000 characters per request. Priced at $0.05/1K characters ($50/1M chars) — 50% cheaper than Multilingual v2. API ID: eleven_flash_v2_5.

$50.00 in / - out / 1M tokens

2 routes

StepAudio 2.5 Realtime is StepFun's end-to-end real-time conversational voice model. It handles speech input and produces speech output through a single unified architecture with no intermediate ASR/TTS pipeline steps. Key capabilities include persona-consistent roleplay via dedicated RLHF training on million-scale persona data, paralinguistic comprehension (detecting and responding to tone, emotion, and speaking rate), and low-latency dialogue. Supports Chinese and English. Available via WebSocket API (step-2.5-realtime). Analogous in function to OpenAI's GPT Realtime models.

Pricing not tracked / 1M tokens

1 route

Scribe v2 Realtime is ElevenLabs' streaming speech-to-text model for voice agents, released November 2025. Delivers ~150ms latency with WebSocket streaming across 90+ languages. Supports Voice Activity Detection (VAD), manual commit control, and the same language coverage as Scribe v2. Priced at $0.39/hr of audio. API ID: scribe_v2_realtime.

Pricing not tracked / 1M tokens

1 route

GPT-4o Mini Transcribe is OpenAI's cost-efficient speech-to-text model based on GPT-4o mini, released March 20, 2025. Offers substantially better accuracy than Whisper at roughly half the price of gpt-4o-transcribe. Supports batch, Realtime transcription, and Assistants endpoints. Input: $1.25/1M audio tokens. Output: $5.00/1M text tokens. Practical: ~$0.003/min. API ID: gpt-4o-mini-transcribe.

- in / $5.00 out / 1M tokens

1 route

Nova-2 is Deepgram's previous-generation flagship speech-to-text model, released September 2023. Delivers ~36% WER improvement over Whisper Large across tested domains (8.4% median WER), with improved entity recognition, punctuation, and capitalization. Supports 36+ languages and 10 domain-specific variants (general, meeting, phonecall, voicemail, finance, conversationalai, video, medical, drivethru, automotive). Batch: $0.0043/min; streaming: $0.0077/min. API ID: nova-2.

Pricing not tracked / 1M tokens

1 route

GPT-4o Transcribe Diarize is OpenAI's automatic speech recognition model with integrated speaker diarization, released October 18, 2025. Identifies and labels who is speaking at each moment in multi-speaker audio, producing a diarized_json response with speaker labels and segment timestamps. Optionally accepts 2–10 second reference audio clips for up to 4 known speakers. Requires chunking for audio >30 seconds. Same pricing as gpt-4o-transcribe: $2.50/1M audio tokens in, $10.00/1M text tokens out. API ID: gpt-4o-transcribe-diarize.

- in / $10.00 out / 1M tokens

1 route

StepAudio 2.5 TTS is StepFun's contextual text-to-speech model with fine-grained expressive control. Unlike tag-based TTS systems, it accepts plain natural language instructions to control emotion, pacing, pauses, and delivery. Supports zero-shot voice cloning with full timbre and emotion control. Priced at $0.85 per 10,000 characters (input text). Supports Chinese and English. Available via StepFun API (model: step-audio-2.5-tts). Part of the unified StepAudio 2.5 architecture described in arXiv:2605.23463.

Pricing not tracked / 1M tokens

1 route

Eleven v3 is ElevenLabs' most expressive text-to-speech model, generally available February 2, 2026 (API preview from June 8, 2025). Supports dramatic delivery, emotional nuance, multi-speaker dialogue, and 70+ languages. Maximum input: 5,000 characters per request. Preferred over Multilingual v2 with 72% preference rate in testing. Priced at $0.10/1K characters ($100/1M chars). API ID: eleven_v3.

$100.00 in / - out / 1M tokens

1 route

Scribe v2 is ElevenLabs' current state-of-the-art batch speech-to-text model, released January 12, 2026. Improvements over v1 for long-form audio, extended silences, and tone changes. Supports 90+ languages, word-level timestamps, 32-speaker diarization, 56 entity types, and keyterm prompting (up to 1,000 terms). Base pricing: $0.22/hr; entity detection add-on: $0.07/hr; keyterm prompting add-on: $0.05/hr. API ID: scribe_v2.

Pricing not tracked / 1M tokens

1 route

Qwen3 Omni 30B A3B is Alibaba's natively end-to-end omnimodal MoE model from the Qwen3 generation, capable of processing text, audio, images, and video while generating real-time streaming text and speech responses. Achieves SOTA on 22 of 36 audio/video benchmarks and open-source SOTA on 32 of 36. Available in Instruct and Thinking (reasoning) variants. Released September 22, 2025.

$0.250 in / $0.970 out / 1M tokens

1 route

Aura-2 is Deepgram's enterprise-grade text-to-speech model, released April 2025. Delivers sub-200ms time-to-first-byte (TTFB) with domain-specific pronunciation for drug names, legal terms, alphanumeric identifiers, dates, and currency. Outperformed ElevenLabs, Cartesia, OpenAI, Azure, Google, and PlayHT in enterprise voice preference testing. Offers 43+ English voices and multilingual support (Spanish, Dutch, French, German, Italian, Japanese). Priced at $0.030/1K characters ($30/1M chars). Available as cloud, VPC, and on-premises. API ID format: aura-2-[voice]-[lang] (e.g. aura-2-thalia-en).

$30.00 in / - out / 1M tokens

1 route

GPT-4o Transcribe is OpenAI's flagship speech-to-text model based on GPT-4o, released March 20, 2025. Delivers substantially better word error rates than Whisper — especially for accented speech, background noise, and variable speaking rates. Supports batch, streaming (Realtime API), and Assistants endpoints. Input: $2.50/1M audio tokens. Output: $10.00/1M text tokens. Practical: ~$0.006/min. API ID: gpt-4o-transcribe.

- in / $10.00 out / 1M tokens

1 route

Nova-3 is Deepgram's current-generation speech-to-text model, released February 12, 2025. Achieves 6.84% median WER streaming and 5.26% batch — a ~54% improvement over competitors. Supports real-time multilingual transcription across 31 languages with self-serve keyterm prompting (up to 100 terms) and real-time redaction (up to 50 entity types). Also includes nova-3-medical specialization. Streaming: $0.0048/min (mono), $0.0058/min (multilingual); pre-recorded: $0.0077/min (mono). API ID: nova-3.

Pricing not tracked / 1M tokens

1 route

Instruction-tuned version of Gemma 4 12B. Open weight (Apache 2.0), 12B parameters, encoder-free multimodal (text, image, audio). Optimized for chat and instruction-following. Runs on a 16GB laptop.

Pricing not tracked / 1M tokens

2 routes

Google DeepMind's 12B open-weight multimodal model (Apache 2.0), designed to run on a 16GB laptop. First medium-sized model with native audio ingestion alongside text and image. Unified encoder-free decoder-only architecture. Supports 140+ languages. MMLU Pro: 77.2%.

Pricing not tracked / 1M tokens

2 routes

Flux is Deepgram's conversational automatic speech recognition (CSR) model, released October 15, 2025. Built specifically for voice agents: integrates turn detection natively into the ASR model, eliminating the need for a separate VAD/endpointing layer. Achieves end-of-turn detection within 1.5s (p95) and reduces interruptions ~30% and latency 200–600ms versus traditional pipelines. Emits voice-agent events: StartOfTurn, EndOfTurn, EagerEndOfTurn, TurnResumed. English: $0.0065/min; Multilingual (10 languages, GA April 2026): $0.0078/min. API ID: flux-general-en / flux-general-multi.

Pricing not tracked / 1M tokens

1 route

Cohere's state-of-the-art automatic speech recognition (ASR) model. Transcribe is a 2B parameter Conformer-based encoder-decoder model trained from scratch for high-fidelity transcription across 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese (Mandarin), Japanese, Korean, Vietnamese, and Arabic. Can process 525 minutes of audio per minute. Achieves 5.42 WER on Hugging Face Open ASR leaderboard.

No tracked provider route

Gemini 4 Carbon is a Google DeepMind reasoning model that Google announced on October 28, 2026 for a subset of cyber defenders and trusted partners in its Fairwind Program, and made generally available on November 10, 2026 through the Gemini API, Gemini Enterprise Agent Platform (Vertex AI), Google Antigravity, the Gemini app and AI Mode in Search for Google AI Ultra subscribers. Google describes it as its most capable reasoning model, built for sustained agentic coding, multimodal knowledge work, scientific reasoning and the cybersecurity defense lifecycle. It takes text, image, video, audio and PDF input and returns text. Input and output share a 1,048,576-token limit; one synchronous request returns up to 65,536 output tokens, and long decode extends output to 1,048,576 tokens. Carbon belongs to the Gemini 4 family, builds on Gemini 4 Argon and launched at Argon's introductory price.

$2.00 in / $10.00 out / 1M tokens

4 routes · 2 batch · 4 cache

Muse Spark Open is the open-weight release of Meta Superintelligence Labs' Muse Spark, an 816-billion-parameter mixture-of-experts model that activates about 42 billion parameters per token. Released October 15, 2026 under the Muse Model License, it is built for coding and agentic work, takes text, image, video and audio input, returns text, and has a 1M-token context window. Weights are on Hugging Face in BF16 and NVFP4.

$1.25 in / $4.25 out / 1M tokens

4 routes · 2 cache

Gemini 4 Argon is a Google DeepMind reasoning model, announced on September 30, 2026 for trusted cyber defenders and testers in Google's Fairwind Program and made generally available on October 13, 2026 through the Gemini API, Gemini Enterprise Agent Platform, Google Antigravity and the Gemini app. Google built it for long-horizon software engineering, legal and finance work, cybersecurity defense and science. It takes text, image, video, audio and PDF input and returns text. Input and output share a 1,048,576-token limit; one synchronous request returns up to 65,536 output tokens, and long decodes extend output up to 1,048,576 tokens. It supports low, medium and high thinking levels, function calling, structured outputs, code execution, search grounding and implicit caching.

$2.00 in / $10.00 out / 1M tokens

4 routes · 3 batch · 4 cache

Muse Wonder is a Meta Superintelligence Labs model released October 12, 2026, for agentic coding, long-running agents and computer use. Meta says it builds on a larger base model than the Muse Spark models. It takes text, image, video and audio input and returns text, with a 1M-token context window. A larger sibling, Muse Wonder Full, is offered only through a Muse Code subscription.

$4.00 in / $20.00 out / 1M tokens

3 routes · 3 cache

MiniMax H3 (API MiniMax-H3) is an open multimodal video model (text/image/video/audio refs → 4–15s video with audio at 768p/2K). Open-sourced 2026-08-03; API on platform.minimax.io. $0.08/s 768P, $0.13/s 2K.

Pricing not tracked / 1M tokens

3 routes

FLUX 3 Video is Black Forest Labs' multimodal video model (API POST /v1/flux-3-video): t2v/i2v/v2v up to ~20s with native audio, HD–UHD, draft mode. GA on BFL API 2026-08-04 after Early Access announce 2026-07-23.

Pricing not tracked / 1M tokens

3 routes

Inkling is Thinking Machines Lab's Apache-2.0 open-weight general-purpose multimodal model. It accepts text, image, and audio inputs and generates text with a 1M-token context window.

$1.00 in / $4.05 out / 1M tokens

3 routes · 3 cache

OpenAI Realtime API flagship speech-to-speech model (gpt-realtime-2.1) with text/audio/image input, text/audio output, 128K context, reasoning token support.

$4.00 in / $24.00 out / 1M tokens

3 routes

Mistral Voxtral Small 24B 2507 is MistralAI's Voxtral model. It was released 2025-07-01.

$0.100 in / $0.300 out / 1M tokens

3 routes · 1 cache

Lyria 3 Clip is Google DeepMind's Lyria model focused on audio understanding and generation. It was released 2025-01-01.

Free in / Free out / 1M tokens

3 routes

Lyria 3 Pro is Google DeepMind's Lyria model focused on audio understanding and generation. It was released 2025-01-01.

Free in / Free out / 1M tokens

3 routes

Muse Spark 1.2 Contributor is the Contributor-tier Model API id (muse-spark-1.2-contributor) for Muse Spark 1.2. Distinct catalog ID (same pattern as muse-spark-1.3-contributor). 1,048,576-token context; multimodal inputs. Contributor mode: discounted tokens in exchange for permission to use content to train future Meta models. FP pricing: $0.10/$0.20 per 1M in/out (cached input $0.002).

$0.100 in / $0.200 out / 1M tokens

3 routes · 3 cache

Voxtral Mini Transcribe 3 is Mistral AI's efficient batch speech-to-text model from the Voxtral 3 family, released October 20, 2026 as a Public Preview on the Mistral API. It has 4B parameters and transcribes more than 50 languages, with speaker diarization, word timestamps and context biasing (key terms or free text that improve recall on domain vocabulary without fine-tuning). Mistral ranks it third on its FLEURS multilingual transcription and English diarization benchmarks, ahead of Gemini 3.5 Transcribe. It has a 32k context and costs $0.0035 per audio minute; weights are not published.

Pricing not tracked / 1M tokens

2 routes

Voxtral Small Transcribe 3 is Mistral AI's flagship speech-to-text model in the Voxtral 3 family, released October 20, 2026 as a Public Preview on the Mistral API. It transcribes more than 50 languages, including the 24 official EU languages, with speaker diarization, end-of-turn detection and context biasing, and also supports translation, spoken Q&A, summarization and function calling over audio. Mistral reports state-of-the-art FLEURS and diarization results against Whisper large-v3, Scribe v2, GPT Transcribe and Gemini 3.5 Transcribe. It has a 32k context and costs $0.004 per audio minute; weights are not published.

Pricing not tracked / 1M tokens

2 routes

Voxtral TTS 3 (API name voxtral-mini-tts-3; also called Voxtral Mini TTS 3 and Voxtral Speech 3) is Mistral AI's text-to-speech model in the Voxtral 3 family, released October 20, 2026 as a Public Preview on the Mistral API in batch and realtime modes. It adapts emotion, pace and emphasis to conversational context, takes explicit inline tags for them, and supports zero-shot voice cloning plus Pro Voice adaptation from about 30 minutes of audio. Streaming time-to-first-audio is about 90 ms. Mistral lists it at $13 per 1M characters; weights are not published.

Pricing not tracked / 1M tokens

2 routes