AssemblyAI Speech-to-Text Pricing: Pre-Recorded Hours and Streaming Sessions

As checked 2026-10-25: AssemblyAI’s first-party pricing page lists per-hour rates for speech-to-text products. This is a dated economics guide, not a permanent price guarantee or a personal hands-on benchmark.

Scope: global English reference; USD list prices where shown; local eligibility, currency, taxes, negotiated terms and model release dates require separate confirmation. Dynamic source-page prices have a check date, not a verified initial effective date.

A time meter is not a transcript meter

AssemblyAI’s first-party pricing page lists per-hour rates for speech-to-text products. In the October 25, 2026 observation, Universal-2 prerecorded transcription was listed at $0.15 per audio hour and Universal-3.5 Pro at $0.21 per hour. Real-time Universal-Streaming was listed at $0.15 per hour, while Universal-3.5 Pro Realtime was $0.45. These are distinct model SKUs with differing functionality; this article describes how their billing is metered, not their relative transcription quality.

The streaming disconnect hazard

The official billing FAQ says prerecorded files are charged by actual file duration in seconds, while streaming is based on session duration. Therefore a call that is connected for an hour but contains only a short spoken exchange might still accrue a full session-hour’s worth of time-based usage. The provider also explains session termination behavior. Engineering teams should explicitly close unused streams and log connection start and end times rather than estimating live costs from transcribed words.

Example with add-ons

A scenario of 40 hours of prerecorded audio at the quoted Universal-2 base rate produces 40 × $0.15 = $6 for recognition alone. If diarization or another separately priced feature is enabled, the relevant per-hour amount must be added for the actual eligible duration. Do not assume a feature is included merely because an API response contains related metadata; inspect the current plan and endpoint’s billed options.

Why rates need version and date

An API route using an earlier Universal generation may have a different rate or support window. The dynamic pricing page shows products that can change as the speech models evolve; preserve the observed date and mark a missing effective date null. Enterprise custom rate limits, volume commitments and tax or currency adjustments are not derivable from consumer-facing list prices.

Operational recommendation

Use prerecorded processing for completed recordings and streaming only where live captions or instant response are necessary. Measure accepted transcription accuracy against the required languages and domain-specific vocabulary under a separate evaluation protocol. Count retransmissions, session idle time and add-ons before deciding which architecture is cheaper in total.

A session-cleanup checklist

Close every streaming connection explicitly after the user ends a call, and persist the exact closure acknowledgement. Maintain alerting for sessions that remain open without speech, reconnect unexpectedly or approach a vendor documented maximum duration. A browser can lose connectivity without the application receiving a clean termination event. Reconcile billed usage to server-side session windows, not to the apparent time on the agent screen. This precaution is especially important if the app scales to thousands of unattended sessions.

How to compare output quality fairly

Select a representative sample of languages, accents, background-noise conditions and specialist terminology. Create a reference transcript reviewed by authorized annotators, then count actual correction work as well as billed processing hours. For a two-speaker interview, decide whether diarization is needed and include its posted cost. An inexpensive transcription that requires prolonged human correction can produce a higher total cost per accepted hour. The pricing article does not assert that any named model has a measured error rate on unseen audio.

Decision checklist

• Identify the specific plan or API endpoint, model version, region and currency before pricing a job.
• Measure input, output, caching, generated media and failed-job units separately.
• Check the official pricing page again before procurement, especially if the SKU has no effective-date notice.
• Keep source evidence and a timestamped price ID; do not treat unknowns as zero.

Sources & Further Reading

• AssemblyAI pre-recorded and streaming listed hours — official first-party source, accessed 2026-10-25; original effective date not specified in source.
• AssemblyAI recording duration versus live session duration accounting — official first-party source, accessed 2026-10-25; original effective date not specified in source.