What an Hour of Audio or Video Costs to Process With AI

An hour of recorded speech costs between ten cents and about a dollar to turn into text at list price, depending on the provider and whether it is live. The same hour costs under a cent to feed into the cheapest model that listens to it directly, and between about eleven cents and several dollars to feed in as video, depending on how many pixels the model is shown each second. Those three sentences describe three different tasks, and this page keeps them in three tables for that reason.

Every figure is a list price read from the provider's own pricing page or documentation on September 18, 2026, converted to one hour with the arithmetic shown, using the tokens-per-second rate each provider states. Where a provider's own per-hour figure exists, it is checked against the conversion and the match or mismatch is noted. Commercial speech vendors are named, not linked. Output tokens, the transcript or summary the model writes, are extra in the second and third tables and are billed at each model's output rate.

Speech-to-text transcription turns audio into words and is priced per minute or per hour of audio; the output is the transcript and is included. Audio understanding means sending the recording to a multimodal language model and asking it a question, such as "summarise this call" or "who agreed to what"; it is priced per audio token, at a tokens-per-second rate the provider sets, and the answer is billed separately as output. Video understanding is the same idea with frames added, and its price depends on the resolution and frame rate the model is shown. A transcription row and an understanding row cannot be ranked against each other, because one includes the output and the other does not, and because they produce different things.

The per-hour conversions use one formula for token-priced rows: tokens per second, times 3,600, times the price per million, divided by a million. Per-minute prices are multiplied by 60 and per-second prices by 3,600. Where we had to assume a rate, the row says "est." and the methodology explains why.

Pre-recorded, pay-as-you-go, standard tier, in ascending order of cost per hour. Volume discounts, enterprise tiers and add-ons are noted only where the pricing page shows them.

Live transcription is a separate sub-table because most providers price it differently and one of the rates below is marked promotional.

For a closer look at OpenAI's transcription models against each other, see our transcription model comparison; for the case where the right answer is no per-hour bill at all, see our guide to self-hosted transcription.

Input only. The tokens-per-second column is what the provider documents for audio input; the per-hour figure is that rate times 3,600 times the input price. OpenAI documents its ten-tokens-a- second rate for the Realtime API, and we apply it to the gpt-audio models as an estimate, marked as such.

The Qwen row is the one behind the claim in our Qwen3.8-Omni-Flash post. On this method the model's hour of audio input costs $0.0038 against $0.277 for its predecessor Qwen3.5-Omni-Plus, a reduction of 98.6%, which agrees with Qwen's "more than 98%"; Qwen's chart prints the two figures as under $0.01 and $0.28. Gemini 3.8 Flash at $0.086 also agrees with the $0.09 on Qwen's chart, so the chart's audio panel reproduces from public documentation.

One hour of video with its audio track, input only, at one frame per second unless stated. Google bills video by frame at 70 tokens for its low and medium settings and 280 for high, plus 32 audio tokens a second. Alibaba Cloud's estimator resizes a 720p frame to 594 tokens at 32-pixel tiles, plus 7 audio tokens a second. Change the frame rate or the resolution and every figure below changes with it.

Four things stand out once the conversions are done. First, batch transcription has converged: ten of the sixteen rows sit between $0.15 and $0.27 an hour, and the cheapest published rate, Microsoft's $0.10, is a preview price with an end date. The outlier at the top, Google Cloud's $0.96 for the first half-million minutes, falls to $0.24 at volume, so the spread is mostly a question of tier.

Second, listening is now cheaper than transcribing, if listening is what you need. An hour of audio into Qwen3.8-Omni-Flash costs less than a cent and into Gemini's Flash-Lite models a few cents, against fifteen to thirty-six cents to transcribe it. The catch is in the word "input": the model still has to write something, and a full verbatim transcript as output would cost more than the input did. For summaries, answers and action items, though, the arithmetic favours the multimodal route by an order of magnitude.

Third, video cost is a setting you choose. The same hour into Gemini 3.8 Flash costs $0.275 at low resolution and $0.84 at high, and a single unchunked hour into Gemini 3.1 Pro crosses the 200,000-token threshold that doubles its rate. Anyone quoting a video-understanding price without the resolution and frame rate is quoting half a number.

Fourth, two prices on this page are dated and one is promotional. Gemini 3.8 Flash doubles on January 1, 2027, Microsoft's transcription rate is stated only to December 31, 2026, and Deepgram's streaming rate is marked as a limited-time promotion. A budget built on this table should carry those three notes with it. If you would like the table applied to your own volumes, with the output side estimated for your workload, our AI transformation service does that as a short engagement.