Compare Speech-to-Text APIs and Transcription Pricing

Xiaomi: MiMo-V2.6-Flash

novita

Lowest price

MiMo-V2.6-Flash is Xiaomi's efficiency-balanced open-weight omnimodal Mixture-of-Experts model designed for coding, agentic workflows, computer use, visual understanding, cybersecurity, and long-horizon tasks. It has 309 billion total parameters with 15 billion activated parameters and natively processes text, images, video, and audio within a 1M-token context window. The model uses a sparse MoE architecture with a 681M-parameter MiMo vision encoder, dedicated audio encoders, and a 5-layer multi-token prediction speculative decoder. It is trained using mixed reinforcement learning across coding, general agents, visual tasks, and cybersecurity, targeting strong agentic capability with more efficient inference than the larger MiMo-V2.6-Pro.

Google: Gemini 2.5 Flash Lite

openrouter

Gemini 2.5 Flash-Lite is Google's lightweight and cost-efficient reasoning model in the Gemini 2.5 family, optimized for high-throughput, low-latency workloads. It supports multimodal inputs including text, images, audio, video, and PDFs, with optional thinking for applications that need a balance between speed, cost, and reasoning quality.

Qwen: Qwen3.8 Omni Flash

openrouter

Qwen3.8 Omni Flash is a multimodal Qwen3.8 Flash-class model designed for efficient reasoning, coding, agentic workflows, visual understanding, video understanding, and long-context processing. The underlying Qwen3.8 Flash architecture combines Gated DeltaNet with Qwen Sparse Attention and Mixture-of-Experts layers to improve inference efficiency for long-context and agentic workloads. It supports up to a 1M-token context window and multimodal understanding across text, images, and video.

Xiaomi: MiMo-V2.6-Pro

novita

MiMo-V2.6-Pro is Xiaomi's flagship open-weight omnimodal Mixture-of-Experts model designed for advanced coding, general agentic workflows, computer use, visual reasoning, cybersecurity, and long-horizon tasks. It has 1.02 trillion total parameters with 42 billion activated parameters and natively processes text, images, video, and audio within a 1M-token context window. The model uses a sparse MoE architecture with a 681M-parameter MiMo vision encoder, dedicated audio encoders, and a 5-layer multi-token prediction speculative decoder. MiMo-V2.6-Pro is trained with large-scale mixed reinforcement learning across coding, general agents, visual tasks, and cybersecurity to improve long-horizon reasoning, tool use, and autonomous task execution.

Google: Gemini 3.1 Flash-Lite

openrouter

Gemini 3.1 Flash-Lite is Google's low-latency, cost-effective multimodal model optimized for high-frequency and high-volume workloads. It supports text, image, video, audio, and PDF inputs and is designed for lightweight reasoning, data extraction, agentic workflows, tool use, and applications where latency and API cost are primary considerations.

Google: Gemini 2.5 Flash

openrouter

Gemini 2.5 Flash is Google's high-performance multimodal workhorse model designed for reasoning, coding, mathematics, science, document understanding, and agentic applications while maintaining Flash-tier speed and cost efficiency.

Google: Gemini 3.5 Flash Lite

google

Gemini 3.5 Flash Lite is Google's high-efficiency multimodal model with upgraded agentic capabilities, optimized for low-cost, high-volume workloads and focused subagents operating inside complex multi-agent systems.

gpt-realtime-mini is OpenAI's low-cost live speech-to-speech model. Connect to WS /v1/realtime, stream 24 kHz PCM microphone audio in and get spoken replies back with low latency, using the OpenAI Realtime event format. Supports function tools and input transcription with gpt-4o-mini-transcribe. Billed by the text and audio tokens each reply uses, so silence costs nothing.

Google: Gemini 3 Flash

google

Gemini 3 Flash is Google's fast multimodal reasoning model in the Gemini 3 family, designed for advanced multimodal understanding, coding, agentic workflows, and complex reasoning while maintaining Flash-class speed and efficiency. It supports a 1M-token input context window and up to 65,536 output tokens, with native support for text, images, audio, video, and PDFs. The model supports configurable thinking, function calling, computer use, code execution, Google Search and Maps grounding, file search, structured outputs, URL context, context caching, and combined built-in and custom tool use.

Google: Gemini 3.6 Flash

google

Gemini 3.6 Flash is a frontier-class Google multimodal Flash model designed for fast reasoning, coding, multimodal understanding, tool use, and scalable agentic applications while retaining Flash-tier latency and efficiency.

Google: Gemini 3.7 Flash

google

Gemini 3.7 Flash is a frontier-class Google Flash model focused on high-performance reasoning, coding, multimodal understanding, tool use, and agentic execution with the latency and cost characteristics of the Flash family.

Google: Gemini 3.8 Flash

google

Gemini 3.8 Flash is Google's most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, complex enterprise workflows, advanced reasoning, and multimodal understanding. Building on Gemini 3.7 Flash, it delivers substantial improvements across coding, agentic tasks, and critical multi-step reasoning while retaining Flash-level speed and cost efficiency. It supports a 1M-token context window, configurable thinking levels, function calling, computer use, code execution, search grounding, structured outputs, file search, and URL context.

Meta: Muse Spark 1.1

meta

Muse Spark 1.1 is Meta's flagship proprietary multimodal frontier model developed by Meta Superintelligence Labs (MSL). Designed for complex reasoning and agentic tasks, it can actively manage a 1-million-token context window, orchestrate multi-agent systems, and natively process text, image, video, and audio inputs. It excels at long-horizon software development, tool use, and computer-use workflows.

Meta: Muse Spark 1.2

meta

Muse Spark 1.2 is a proprietary coding-focused frontier model developed by Meta Superintelligence Labs. Co-trained specifically with the Muse Code terminal agent, it features significant improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows. It leverages planning, goal conditioning, and context compaction to sustain progress on long-horizon tasks like whole-repository generation and GPU kernel optimization.

Google: Gemini 3.5 Flash

google

Gemini 3.5 Flash is Google's high-efficiency multimodal model offering near-Pro-level coding and reasoning at Flash-tier speed and cost. It is optimized for coding, parallel agentic execution, multimodal understanding, and scalable production workloads.

Google: Gemini 2.5 Pro

openrouter

Gemini 2.5 Pro is Google's advanced Gemini 2.5 reasoning model for complex problem solving, coding, mathematics, scientific analysis, long-context understanding, multimodal reasoning, and sophisticated agentic workflows.

Gemini 4 Argon

google

Gemini 4 Argon API on OneInfer: Google's flagship reasoning model with a 1M-token context. $2/1M input, $10/1M output tokens until Nov 30, 2026.

Google: Gemini 3.1 Pro

google

Gemini 3.1 Pro is Google's advanced intelligence model for complex reasoning, problem solving, software engineering, multimodal analysis, long-context tasks, and powerful agentic and vibe-coding workflows.

Meta: Muse Wonder 1.0

meta

Muse Wonder 1.0 is Meta's frontier multimodal reasoning and agentic AI model developed by Meta Superintelligence Labs. It is designed for complex software engineering, long-running autonomous workflows, professional knowledge work, and advanced computer use. The model supports a 1,048,576-token context window, text, image, audio, video and PDF understanding, tool calling, structured outputs, search grounding, and configurable reasoning effort from minimal to max. It introduces Active Memory for retaining relevant information across conversations and Active Focus for sustaining multi-step tasks, recovering from errors, delegating to sub-agents and verifying results. Muse Wonder is optimized for coding, financial analysis, legal research, healthcare reasoning, browser automation, desktop interaction and agent orchestration.

ElevenLabs: Scribe v2

elevenlabs

Scribe v2 is ElevenLabs' state-of-the-art speech-to-text model designed for highly accurate transcription of long-form and complex audio across more than 90 languages. It provides precise word-level timestamps, smart speaker diarization for up to 32 speakers, dynamic audio-event tagging, automatic multilingual language detection, keyterm prompting, entity detection, and natural-language transcript editing. The model is optimized for batch transcription, subtitles, captions, meetings, interviews, podcasts, media archives, research, and enterprise transcription workflows, with improved stability across long recordings, pauses, changing speakers, accents, delivery styles, and extended silences.

ElevenLabs: Scribe v2 Medical

elevenlabs

Scribe v2 Medical is ElevenLabs' specialized speech-to-text model fine-tuned for medical and clinical audio. Built on Scribe v2, it improves recognition of medication names, dosages, anatomy, pathology, clinical terminology, and dense medical dictation while maintaining the base model's accuracy on everyday speech. It supports more than 90 languages and includes speaker diarization, precise word-level timestamps, dynamic audio tagging, keyterm prompting, entity detection, PHI-related entity handling, and no-verbatim transcription. The model is designed for clinical documentation, clinician-patient conversations, medical dictation, patient intake, care-coordination calls, and other healthcare transcription workflows where the generated transcript is reviewed by an authorized healthcare professional.

Fish Audio: Transcribe 1 Pro

openrouter

Transcribe 1 Pro is Fish Audio's advanced speech-to-text model optimized for interviews, meetings, podcasts, and multi-speaker conversations. It converts audio into text while automatically detecting the spoken language, identifying different speakers with inline speaker markers, and preserving expressive information such as emotions and vocal events including laughter. The model can optionally return timestamped word-level segments, making it suitable for transcription, subtitle generation, meeting intelligence, podcast processing, speaker-aware transcripts, and other structured speech recognition workflows.

Fishaudio: S2 Pro

openrouter

Fish Audio S2 Pro is an advanced 4B parameter open-weight text-to-speech model built on a dual-autoregressive architecture. It supports approximately 80 languages and introduces fine-grained, word-level inline tag control, allowing users to embed open-ended natural language instructions inside square brackets directly in the script (e.g., [whispering], [pitch up]). It outputs 44.1 kHz audio and natively supports zero-shot voice cloning from a reference clip.

Microsoft: MAI Transcribe 1.5

openrouter

MAI-Transcribe-1.5 is Microsoft's high-accuracy multilingual speech-to-text model designed for fast production-scale transcription across 43 languages. It combines low word error rates with very high transcription throughput and is optimized for meetings, calls, media, enterprise transcription, subtitles, voice applications, and noisy real-world audio. The model supports automatic multilingual transcription and keyword biasing for domain-specific terminology such as names, products, medical terms, and internal acronyms. Microsoft reports that keyword biasing can reduce word error rate by up to 30% on FLEURS. The model is optimized for batch transcription and can process an hour of audio in under 15 seconds under Microsoft's high-throughput measurements.

Microsoft: MAI Transcribe 2

openrouter

MAI-Transcribe-2 is Microsoft's high-accuracy multilingual speech-to-text model designed for fast, production-scale transcription across 60 languages and challenging real-world audio conditions. It provides automatic language identification, speaker diarization, word-level and segment-level timestamps, keyword biasing for domain-specific terminology, code-switching support, and configurable clean or verbatim transcription styles. The model is optimized for meetings, call-center conversations, clinical notes, legal documentation, captions, accessibility, media transcription, content workflows, and voice-agent applications. It combines high transcription accuracy with extremely high batch throughput, processing approximately one hour of audio in around 10 seconds under Microsoft's benchmark conditions.

Sarvam AI: Saaras V3

sarvam

Saaras v3 is a state-of-the-art speech recognition and translation model. It natively supports streaming for real-time applications and features specialized modes for transcription, direct-to-English translation, and transliteration. It is specifically tuned for noisy environments and complex code-mixed (e.g., Hindi-English) speech.

Sarvam AI: Saarika V2.5

sarvam

Saarika v2.5 is Sarvam AI's legacy speech recognition model designed for Indian languages and accents. It transcribes audio in the same language spoken, excelling in multi-speaker conversations, telephony audio (8kHz), and code-mixed speech. Supports 11 languages (Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu, English) with automatic language detection and speaker diarization. Achieves 4.96% CER and 18.32% WER on VISTAAR benchmark.