Subscribe
AI
7 min read

What an Hour of Transcription Actually Costs in Late 2026

The rate cards say fifteen to forty cents an hour. They agree right up until the invoice, where the same thousand hours of audio can bill anywhere from forty dollars to nearly a thousand. The gap is not model quality. It is five incompatible billing meters, add-on line items, channel-counting rules, and a set of shutdown dates that land in February 2027. This is what an hour of machine transcription actually costs in late 2026, meter by meter. The accuracy menu moved, and it got cheap Artifi

What an Hour of Transcription Actually Costs in Late 2026

The rate cards say fifteen to forty cents an hour. They agree right up until the invoice, where the same thousand hours of audio can bill anywhere from forty dollars to nearly a thousand. The gap is not model quality. It is five incompatible billing meters, add-on line items, channel-counting rules, and a set of shutdown dates that land in February 2027.

This is what an hour of machine transcription actually costs in late 2026, meter by meter.

The accuracy menu moved, and it got cheap

Artificial Analysis runs the AA-WER v2 benchmark on roughly eight hours of audio drawn from three sets: its own AgentTalk conversation data (weighted 50 percent), cleaned VoxPopuli (25 percent) and cleaned Earnings22 (25 percent). The current non-streaming leaderboard, with the site's normalized cost per 1,000 minutes of audio, reads as follows:

  • Fun-Realtime-ASR-preview (Alibaba): 1.7% word error rate, $2.82 per 1,000 minutes
  • MAI-Transcribe-2 (Microsoft): 2.0%, $1.67
  • Voxtral Small Transcribe 3 (Mistral): 2.1%, $4.00
  • Scribe v2 (ElevenLabs): 2.2%, $3.67
  • Gemini 3.5 Transcribe (Google): 2.6%, $5.00
  • Universal-3.5 Pro (AssemblyAI): 3.0%, $3.50
  • Amazon Transcribe (AWS): 4.1%, $6.00
  • Chirp 3 (Google Cloud STT v2): 4.3%, $16.00
  • Whisper Large v3 Turbo (Groq): 4.6%, $0.67
  • Nova-3 (Deepgram): 5.2%, $4.30

Two things stand out. First, near-2% error now costs between ten and thirty cents an hour, a tier that would have cost dollars two years ago. Groq's Whisper Turbo at four cents an hour and Microsoft's model at ten cents have collapsed the budget end. Second, the incumbents no longer hold accuracy. Google's Chirp 3 is the most expensive entry in the top group by a factor of three and sits at 4.3% WER. Deepgram's Nova-3, still the developer default in a lot of stacks, is at 5.2%. The old trade (pay more to the established cloud, get fewer mistakes) no longer describes the market.

Vendors publish their own benchmarks too, always more favorable, and the streaming numbers compress differently. On Artificial Analysis's AA-WER Streaming index, Cartesia Ink-2 leads at 3.59% WER with 0.21 seconds to final transcription, ElevenLabs Scribe v2 Realtime is at 3.64% and 0.14 seconds, and Deepgram's Flux is the fastest endpoint at 0.020 seconds but transcribes at 7.36% WER. Speed and accuracy in live audio remain a trade-off. In batch they mostly are not.

Five meters for the same audio

Five vintage weighing scales measuring one sound wave

The list price per hour is the smallest part of the problem. The providers do not bill the same unit.

Per hour of audio. AssemblyAI charges $0.15 per hour for Universal-2 and $0.21 for Universal-3.5 Pro on pre-recorded files. ElevenLabs charges $0.22 per audio hour for Scribe v2 batch and $0.39 for the Realtime variant.

Per minute of audio. Deepgram's pay-as-you-go rate card lists Nova-3 Monolingual at $0.0077 per minute, which the company's own pricing FAQ converts to $0.46 an hour for streaming, with pre-recorded runs cheaper (price trackers reading the page in October and November record $0.0043 per minute, about $0.26 an hour, and note the streaming promotion at $0.0048 ended in late October). OpenAI sells gpt-transcribe at $0.0045 per minute and gpt-live-transcribe at $0.017. AWS bills standard batch at $0.006 per minute and streaming at $0.01 in US East. Google's Speech-to-Text v2 standard models run $0.016 per minute with data logging ($0.024 without), and $0.003 per minute on dynamic batch.

Per token. OpenAI's gpt-4o-transcribe family bills at $2.50 per million audio input tokens, which the company and price trackers estimate at roughly $0.006 per minute. The estimate is the point: with token meters, your cost depends on sample rate and encoding, not just duration.

Per session hour. Streaming products that bill connection time rather than speech time charge through silences. AssemblyAI's Realtime models run $0.45 per session hour, and independent pricing guides flag that idle connection minutes count on session-based meters.

Per GPU hour. Microsoft's MAI-Transcribe models ship through Azure AI Foundry on GPU-hour pricing, which Artificial Analysis normalizes to $1.67 per 1,000 minutes for MAI-Transcribe-2. The effective rate depends on how hard you run the hardware.

The streaming tax is consistent

Across every provider that sells both batch and live transcription, real-time costs a multiple of the same model batched: roughly 1.8x at AWS ($0.006 to $0.01 per minute), 2.1x at Deepgram ($0.26 to $0.46 per hour), 1.8x at ElevenLabs ($0.22 to $0.39), 2.1x at AssemblyAI ($0.21 to $0.45), and 3.8x at OpenAI ($0.27 to $1.02 per hour). If your product only needs captions after the call ends, the batch meter is always the answer, and several providers (Google dynamic batch at $0.003 per minute, AssemblyAI's cheaper Universal-2) price the flexible-latency tier below the standard one.

Where the add-ons bite

A long receipt turning into a descending staircase of price blocks

Base rates assume you need nothing else. Most production workloads need speaker labels, and many need redaction or domain vocabulary.

  • AssemblyAI prices each feature per hour on top of transcription: diarization +$0.02, keyterms prompting +$0.05, PII text redaction +$0.08, Medical Mode +$0.15. Universal-2 includes keyterms prompting in the base rate.
  • Deepgram charges add-ons per minute: diarization, redaction and keyterm prompting each run about $0.0020 per minute, roughly $0.12 an hour apiece.
  • AWS and Google bundle speaker diarization, custom vocabularies and language identification into the standard rate. What AWS bills extra for is content redaction (tiered from $0.0024 per minute) and domain verticals: Transcribe Medical runs $0.075 per minute, about twelve times its standard batch rate.
  • ElevenLabs includes speaker labels for up to 32 speakers in Scribe v2's $0.22 and sells keyterm prompting at +$0.05 per hour.

Worked example, 1,000 hours of recorded call audio per month with speaker labels required:

  • Groq Whisper Large v3 Turbo: $40, no diarization, 4.6% WER
  • Microsoft MAI-Transcribe-2 (Azure): about $100 as normalized by Artificial Analysis, GPU-hour billing
  • Google dynamic batch: $180, diarization included
  • AssemblyAI Universal-2 plus diarization: $170
  • AssemblyAI Universal-3.5 Pro plus diarization: $230
  • ElevenLabs Scribe v2: $220, labels included
  • OpenAI gpt-transcribe: $270, no speaker labels (the diarizing model retires in February)
  • AWS Transcribe batch: $360, diarization included
  • Deepgram Nova-3 pre-recorded plus diarization: about $378
  • Google STT v2 standard, with logging: $960

Same month of audio: a twenty-four-fold spread, most of it explained by units and add-ons rather than accuracy.

One channel-count gotcha flips rankings on stereo call recordings. AWS states that a two-channel conversation is billed on total duration, not per channel. Deepgram's pricing page bills multichannel audio on total processed duration across all channels, which can double the meter on two-track recordings before any feature flags.

Deprecations are the hidden price change

OpenAI will shut down whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize on February 26, 2027, per its API deprecation schedule. For teams on gpt-4o-transcribe, moving to gpt-transcribe is a price cut ($0.006 to $0.0045 per minute). For teams on gpt-4o-mini-transcribe it is a 50% increase, and for anyone using diarization through OpenAI the replacement does not offer it at any price. Deepgram's promotional streaming rates, which held Nova-3 at $0.0048 per minute through October, have already expired. The lesson from the last year of speech APIs is the same as for text models: the rate on the pricing page is a snapshot, and the migration window is where budgets break.

What to check before you commit

  1. Benchmark on your own audio. The AA-WER gap between a 2% model and a 5% model is invisible on clean studio files and very visible on your call recordings, accents and jargon. Vendors say this themselves. Deepgram's own materials concede that measuring word error rate on your audio is the only benchmark that matters.
  2. Divide by the right unit. Convert everything to dollars per 1,000 minutes of your typical file length, at your channel count, with every feature you actually call.
  3. Ask what silence costs. Session-hour meters, 15-second rounding (historical on Google) and per-second billing (AWS, Deepgram, AssemblyAI) diverge most on short clips and long idle connections.
  4. Read the retirement schedule. Three of OpenAI's current transcription endpoints are dead in fourteen months.
  5. Price the free tier honestly. Deepgram gives $200 in non-expiring credits, AssemblyAI $50, Google 60 free minutes per month indefinitely and AWS 60 per month for twelve months. Useful for pilots, irrelevant at volume.

The uncomfortable conclusion for the incumbent clouds is that transcription stopped being a differentiator. On batch files, the accuracy leader costs seventeen cents an hour and the cheap-and-good options cluster between ten and twenty-five. Chirp 3's $16 per 1,000 minutes and Speechmatics Enhanced's $12.50 buy enterprise procurement comfort, not fewer mistakes.

Sources:

AILLMvoice AIPricing