Gladia Review, Pricing & Alternatives (2026)

Gladia is a Paris-based audio AI company founded in 2022 by Jean-Louis Quéguiner and Jonathan Soto. Its first asynchronous transcription and audio intelligence API launched in June 2023, based on a proprietary version of Whisper ASR. Gladia announced a $16 million Series A in October 2024, bringing total funding to $20.3 million. The company said its API had more than 600 customers and 70,000 users. Its target customers include meeting assistants, contact-center platforms, sales tools, media services, and other products that process recorded or live audio. This background matters for teams assessing continuity, scale, and intended use. Gladia is positioned as infrastructure for embedding audio processing, rather than as a consumer transcription application or standalone meeting workspace.

A typical workflow sends a file, audio URL, or live stream to Gladia’s API and receives a transcript through polling or callbacks. The service offers automatic language detection, code-switching, speaker diarization, timestamps, custom vocabulary, and name consistency. Gladia says its speech recognition supports more than 100 languages. Audio Intelligence add-ons cover summarization, translation, sentiment analysis, named entity recognition, PII redaction, and Audio to LLM. Solaria-3 is for pre-recorded audio, while real-time transcription uses a separate rate. In practice, developers can select asynchronous processing when delayed results are acceptable, or maintain a live session when applications need arriving transcript text. The available controls also support workflows that connect transcripts with downstream analysis, records, or user-facing features.

Gladia’s hosted models are proprietary. Its model documentation identifies Solaria-1 as the default model and describes Solaria-3 as its highest-accuracy model for European real-world audio, available for pre-recorded audio only. Recent releases include Solaria-3 in June 2026, PII Redaction and ISO 27001 and ISO 27701 certification updates in March 2026, and Zero Data Retention in February 2026. Gladia also released Audio to LLM generally in April 2026 and published an open speech-recognition benchmark in March 2026. Its SDKs and normalization library do not make the speech models open source. Consequently, teams wanting hosted model access can use the API, whereas teams requiring deployable speech weights should recognize that the documented models are proprietary.

Gladia uses prepaid credits rather than subscriptions. Self-serve asynchronous transcription starts at €0.16 per hour for Solaria-3 and costs €0.27 per hour for Solaria-1 at the lowest listed volume band. Real-time transcription costs €0.68 per hour. Monthly volume above 10,000 hours lowers Solaria-3 to €0.12 per hour, while add-ons cost extra. New accounts receive €50 in credits valid for one year. Self-serve allows 25 concurrent asynchronous jobs and 30 concurrent real-time sessions. Enterprise pricing is custom and includes unlimited concurrency, a 99.9% uptime SLA, custom hosting, custom models, and zero data retention. Buyers should therefore estimate both transcription and add-on usage, while checking whether concurrency limits fit expected workloads before selecting self-serve.

Gladia’s main limitation is product fit: Solaria-3 is pre-recorded only, and the service does not provide an offline speech engine. AssemblyAI, Deepgram, and OpenAI Speech-to-Text are direct alternatives, so buyers should compare supported languages, streaming behavior, data controls, concurrency, accuracy for target audio, and current rates before choosing. Gladia says self-serve model training is enabled by default, with free opt-out, while Enterprise is opted out by default. Its privacy notice also says uploaded data should not include specified sensitive information, including protected health information. Pick Gladia for a hosted API with multilingual coverage, European privacy controls, and usage-based billing, especially when those priorities outweigh offline processing or a pre-recorded-only model option.