Hardcore Reviews
Hardcore Reviews

The Cost of Feeding Audio and Video to AI: Who Is Cheapest

This review runs the math on feeding audio and video to models: it lines up Qwen3.8-Omni-Flash, Gemini 3.8 Flash, ByteDance Doubao Seed series and OpenAI GPT-5.x (data collected 2026-09-19, per official pricing pages) on audio and video input pricing with a capability snapshot. Key findings: Qwen3.8-Omni-Flash bills one flat omni-modal rate in China of 0.8 CNY per million input and 2.7 CNY output tokens, roughly a 98.6% audio-input cut versus the previous generation; Gemini 3.8 Flash charges 0.75 USD per million input (intro price until 2026-12-31, doubling from 2027); OpenAI audio input runs 5-10 USD per million and realtime audio about 32. It closes with three scenario budgets (one-hour meeting, one-hour video notes, realtime support) and a reminder that token-consumption optimization and hidden engineering costs matter more than list price.

Published September 19, 20268 min read
<!-- audio-video-model-cost-comparison-review | review | The Cost of Feeding Audio and Video to AI: Who Is Cheapest -->

This review sits beside three others on the site. agnes-free-multimodal-api-review asks whether free multimodal APIs are good enough. kimi-vs-qwen-vs-glm-long-context-review compares the three leading Chinese models on long context. post-aug31-token-cost-comparison-review compares the real token cost of eleven models after price increases. None focuses on the edge this piece covers: handing raw audio and video to a model and asking what it costs. This review only adds up that bill across four columns: audio input price, video input price, output price, and the relief a large context window buys. Capability is sketched, not scored. If you came for "who is smartest", this is the wrong article; it answers "who is cheapest at pushing an hour of meeting audio or a long video into a model".

Scope, Method, and Selection

This review only adds up the bill for feeding audio and video to a model. Four things matter: audio input unit price, video input unit price, output unit price, and how much a large context window spares you from slicing and stitching. We do not rank models on benchmarks.

Every price comes from an official pricing page, dated 2026-09-19: Alibaba Cloud Bailian and DashScope, Google Gemini developer API, Volcengine Ark, and OpenAI. Overseas models are quoted in US dollars per million tokens; domestic in yuan. We do not convert currencies. Where we could not verify a figure, we write "check the official pricing page" rather than invent one.

Three billing traps deserve a flag. First, are modalities priced the same? Some charge one rate across text, image, audio, and video; others list audio at a premium. Second, token conversion: an hour of identical audio yields different token counts under different tokenizers, so confirm the tokens are comparable before comparing per-token prices. Third, does output include thinking? Gemini 3.8 Flash bills thinking tokens as output, as does part of OpenAI, quietly inflating the bill.

We picked four that compete on the audio-video input edge: Qwen3.8-Omni-Flash, the price-cut protagonist with native omnimodality; Gemini 3.8 Flash, native multimodal with a large context; Doubao on Volcengine, domestic video understanding billed by link; and OpenAI GPT-5.x with the realtime family, mature on audio and frame-based on video. Pure-transcription vendors such as iFlytek sit outside the understanding-multimodal table and appear only in the reflection.

Qwen3.8-Omni-Flash: One Price for Every Modality

Alibaba charges one input rate whether tokens came from text, an image, audio, or video. On the Chinese mainland via Bailian, input is 0.8 yuan per million tokens, cache-hit 0.1 yuan, output 2.7 yuan. In international regions, Model Studio lists input 0.15 US dollars per million tokens, cache-hit 0.016, output 0.47. Source: Alibaba Cloud Bailian pricing page and the official launch blog, dated 2026-09-19.

The cut is the thread of this piece. The prior Qwen3.5-Omni-Plus charged 11 US dollars per million tokens for audio input internationally; the new model drops that to 0.15, about 98.6 percent. By official framing, an hour of audio costs more than 98 percent less, and audio plus video more than 93 percent less, pricing a two-minute sample and multiplying by thirty with video at 720p and one frame per second. On the mainland the prior audio rate was about 18 yuan; the new model is 0.8 yuan.

Capability sketch, from the Qwen launch and ai-bot.cn on 2026-09-18. On AliMeeting diarization DER, Qwen scores 3.4 against Gemini 72.6, lower is better, a wide lead. We state honestly the three rows Gemini wins: OmniVideoBench 63.4 against 65.2, LVOmniBench 63.3 against 70.7, and FLEURS WER 9.3 against 7.9. The cut did not collapse capability; Gemini still leads on long-video reasoning and multilingual ASR.

The sweet spot is a 1M-token context that holds about an hour of audio-video, with agentic evidence-gathering cutting token use about 45.7 percent. For an hour of video you skip the slice-extract-stitch pipeline and feed it once. For usage and plugins see the Qwen3.8-Omni-Flash hotspot, the Qwen-MM-Plugins resource, and the API access SOP.

Gemini 3.8 Flash: Pricy on Output and Thinking

Google prices Gemini 3.8 Flash with one rate across all input modalities: text, image, video, audio, and PDF bill at 0.75 US dollars per million input, output 3.75 including thinking, cache-hit 0.075. Source: Google Gemini developer API pricing page, dated 2026-09-19. Two traps: 0.75 and 3.75 are an intro rate through 2026-12-31 and double to 1.50 and 7.50 on 2027-01-01; and output includes thinking tokens, so more thinking means a heavier bill.

Against Qwen, Gemini input at 0.75 is five times Qwen international at 0.15, while output at 3.75 is about eight times Qwen at 0.47. Gemini is more expensive on feed and even more on produce. For a long video, where input is huge and output modest, Gemini's cost sits mainly in input.

Gemini wins the three rows above and offers a 1M context. Its quiet advantage is a mature ecosystem and a separate Live API. The realtime voice line is outside the main table because it runs on per-minute rather than per-token rails.

Doubao on Volcengine keeps two ledgers. On understanding, Doubao-Seed-2.0-pro accepts a video link; a public link bills about 8.6 yuan per million input and 28.9 yuan output in the 0-32k tier, while a same-region TOS path drops input to about 6.6 yuan. Plain text, image, and video on Seed-2.1-pro cost 6 yuan input, 30 yuan output, 1.2 yuan cache-hit; Seed-2.1-turbo is 3 yuan input, 15 yuan output. Source: Volcengine Ark pricing page and product page, dated 2026-09-19. Doubao's tiered billing by input length is its biggest difference from one-price models; a long video in a higher tier lifts the unit price.

On recognition, Doubao file ASR is 0.8 yuan per hour, streaming 1.0 yuan. That is a "turn into text" price, not an understanding price, so we do not set it beside Qwen and Gemini input rates; we only note that if you only need transcription, hourly ASR is often far cheaper than feeding whole audio as tokens.

OpenAI GPT-5.x supports audio input natively. GPT-5.4 charges 5 US dollars per million for audio input, 20 audio output, 2.5 text input, 15 output; GPT-5.5 charges 10 audio input, 5 text input, 30 output. Source: OpenAI official pricing page cross-checked with trackers, dated 2026-09-19. Video needs a note: GPT-5.x input is mainly text, image, and audio, and native video understanding requires splitting into frames fed as images, with no official "one price per video" basis, so we estimate by image-frame input. The realtime line, gpt-realtime-2, bills audio input about 32 US dollars per million, with a per-minute translation channel near 0.034 US dollars per minute.

OpenAI is the most expensive here: audio input at 5 to 10 US dollars is thirty to sixty-five times Qwen at 0.15. Its strength is reasoning and ecosystem, not cheap audio-video intake. The site's earlier token-cost review also places OpenAI's cost in output and thinking.

The Summary Table: Audio, Video, Output, Context

All prices are 2026-09-19 official snapshots; currencies are not converted. Gemini shows the intro rate, doubling in 2027.

ModelAudio inputVideo inputOutputCache hitContextBasis
Qwen3.8-Omni-Flash0.15 USD / M (intl); 0.8 yuan / M (mainland)same as audio0.47 USD / 2.7 yuan0.016 / 0.11Mone price
Gemini 3.8 Flash0.75 USD / Msame3.75 USD (thinking)0.0751Mintro, doubles 2027
Doubao Seed-2.1-pro6 yuan / M~8.6 yuan / M link30 yuan1.2256ktiered
OpenAI GPT-5.45 USD / M audioframe15 USD0.25400kaudio high
OpenAI GPT-5.510 USD / M audioframe30 USD0.5400kaudio high

The table shows the edge: Qwen pulled feed-side price to text-model levels, Gemini is cheap on intro but expensive on output and will double, Doubao's link and tier billing is unfriendly to long video, and OpenAI audio is priciest. Doubao and OpenAI "video input" is not a native whole-video rate; one bills by link, the other by frame.

A second table adds the capability direction, so cost and fitness sit in one view.

BenchmarkQwen3.8-Omni-FlashGemini 3.8 FlashNote
AliMeeting DER (low)3.472.6Qwen lead
WildClawBench-MM (high)71.058.9Qwen
OmniVideoBench (high)63.465.2Gemini
LVOmniBench (high)63.370.7Gemini
FLEURS WER (low)9.37.9Gemini

Scenario Math: What Three Jobs Cost

Scenarios use transparent assumptions for relative gaps, not precise quotes. Domestic models in yuan, overseas in US dollars, not converted.

Scenario one: an hour of meeting audio to minutes. Assume about 60K input tokens, 2K output. Qwen international: input 60K times 0.15 per million about 0.009 US dollars, output 2K times 0.47 about 0.0009, total about 0.01 US dollars, roughly 0.07 yuan. Gemini: input 0.045, output 0.0075 and higher with thinking, above 0.05. Doubao Seed-2.1-pro audio understanding about 60K times 6 per million about 0.36 yuan; pure ASR at 0.8 yuan per hour avoids token conversion. OpenAI GPT-5.4 audio: 60K times 5 per million equals 0.3 US dollars, about thirty times Qwen. Verdict: meetings are cheapest per token on Qwen and Gemini; pure transcription on Doubao ASR is simple but no understanding.

Scenario two: an hour of video to notes. Assume 720p, one frame per second, plus audio, about 200K input, 3K output. Qwen international: input 200K times 0.15 equals 0.03 US dollars, output 3K times 0.47 about 0.0014, total about 0.03 US dollars, roughly 0.2 yuan. Gemini: input 0.15, output higher with thinking, around 0.2. Doubao Seed-2.0-pro video link public: about 200K times 8.6 about 1.72 yuan input, plus output about 1.9 yuan. OpenAI by frame lands in the Doubao class or higher. Verdict: on long video Qwen's one-price-plus-large-context advantage is sharpest, pressing cost to cents.

Scenario three: realtime voice service. This runs on realtime rails, not Chat Completions. OpenAI gpt-realtime-2 per-minute translation is about 0.034 US dollars per minute, or audio tokens near 32 US dollars per million. Gemini Live API bills separately. Qwen3.8-Omni-Flash-Realtime has shipped but had no standalone pricing row as of 2026-09-19, so check the official page. Realtime compares on latency and concurrency more than unit price.

Reflection: Price War, Token Optimization, Hidden Cost

The price war logic: Qwen cut audio input from 11 to 0.15 US dollars, pulling omnimodality from luxury to commodity. Teams that avoided models over audio cost can now feed long meetings directly. But cuts move; Gemini's intro rate doubles in 2027, so "cheapest today" is not "cheapest forever".

Token optimization matters more than unit price. Qwen says agentic evidence-gathering cuts token use about 45.7 percent, a second discount on top of the unit cut. Coarse-scan-then-evidence versus stuffing the whole clip changes the bill by nearly half. Gemini thinking is the same: dropping the level from high to medium can cut output tokens by a third. Compare whether you can compress tokens before comparing unit price.

Hidden costs outside price. Preprocessing: OpenAI video needs frame splitting, Doubao a link, Gemini sampling control, all engineering that costs staff time. A large video on a TOS path is cheaper than public, but you move it first. A multi-model chain of slice, extract, transcribe, and understand looks cheap per unit but carries high maintenance; a one-feed model like Qwen is low on price and low on engineering. Compliance: domestic workloads on Bailian or Volcengine keep data in-region, a cost absent from any price sheet. Pure-ASR vendors like iFlytek still hold a place in transcription-only scenarios but sit outside the understanding-multimodal table.

Closing: this piece adds up cost, not capability. For real selection, read this edge with the Chinese three-model long-context review and the free multimodal API review.

FAQ

question one: How does this differ from the site's token-cost review?

That piece compares the real token cost of eleven models after price increases, mostly text-centered. This piece focuses only on feeding raw audio and video to a model, counting audio input, video input, output, and context. The two are complementary.

question two: Are Qwen's price-cut numbers reliable?

Audio input fell from 11 US dollars per million tokens, the prior Qwen3.5-Omni-Plus international rate, to 0.15, about 98.6 percent, from the Alibaba pricing page and launch. The more than 98 percent hourly audio and more than 93 percent audio-video cuts are official framing, a two-minute sample times thirty. Figures follow the official pricing page, dated 2026-09-19.

question three: Why does OpenAI have no single video price?

GPT-5.x native input is mainly text, image, and audio; video must be split into frames fed as images, with no official "one price per video" basis, so we estimate by image-frame input. That split is itself OpenAI's hidden engineering cost on this edge.

question four: Doubao hourly ASR or token-based understanding, which to pick?

If you only transcribe, Doubao file ASR at 0.8 yuan per hour is often cheaper than feeding whole audio as tokens; if you need watch-and-reason understanding, use the Seed series rate, about 8.6 yuan per million tokens public for video link. Pick by need.

question five: Gemini intro rate doubles in 2027, can we trust the current price?

0.75 and 3.75 are an intro rate through 2026-12-31, and Google docs state 1.50 and 7.50 from 2027-01-01. This is announced, not a rumor. Long-running workloads must budget the post-doubling run rate, not today's price.

This article is AI-assisted and human-edited. Last updated: 2026-09-19

FAQ

How does this differ from the site's token-cost review?
That piece compares the real token cost of eleven models after price increases, mostly text-centered. This piece focuses only on feeding raw audio and video to a model, counting audio input, video input, output, and context. The two are complementary.
Are Qwen's price-cut numbers reliable?
Audio input fell from 11 US dollars per million tokens, the prior Qwen3.5-Omni-Plus international rate, to 0.15, about 98.6 percent, from the Alibaba pricing page and launch. The more than 98 percent hourly audio and more than 93 percent audio-video cuts are official framing, a two-minute sample times thirty. Figures follow the official pricing page, dated 2026-09-19.
Why does OpenAI have no single video price?
GPT-5.x native input is mainly text, image, and audio; video must be split into frames fed as images, with no official "one price per video" basis, so we estimate by image-frame input. That split is itself OpenAI's hidden engineering cost on this edge.
Doubao hourly ASR or token-based understanding, which to pick?
If you only transcribe, Doubao file ASR at 0.8 yuan per hour is often cheaper than feeding whole audio as tokens; if you need watch-and-reason understanding, use the Seed series rate, about 8.6 yuan per million tokens public for video link. Pick by need.
Gemini intro rate doubles in 2027, can we trust the current price?
0.75 and 3.75 are an intro rate through 2026-12-31, and Google docs state 1.50 and 7.50 from 2027-01-01. This is announced, not a rumor. Long-running workloads must budget the post-doubling run rate, not today's price.

Related

Hardcore Reviews

Free Tier Showdown: Six AI Coding Tools at $0 Cost

This review runs the free math only, no model capability: it lines up Qoder, Cursor, Trae, Windsurf, Claude Code and Codex (data collected 2026-09-18, per official pricing pages) on free-tier contents and limits. Key findings: Trae has the thickest paper free tier (1,000 premium plus 5,000 completions monthly), Cursor Hobby gives 2,000 completions plus 50 slow requests, Windsurf offers 25 prompt credits monthly plus 5 Cascade sessions daily; Claude Code and Codex have no real free tier and need a $20/month subscription for full use. During the window, Qoder's free Qwen3.8-Flash plus daily 100 Credits sets the current ceiling for zero-cost usage. It closes with bundle strategies for three audiences (free-rider, light, heavy) and the true cost of free: data, lock-in, and the price hike after the window.

Sep 18, 20268 min read
Hardcore Reviews

Self-Hosted AI Assistants: Form, Ownership, Selection

This review skips model capability and compares form and ownership only: it lines up Octop, Open WebUI, Dify, FastGPT and LibreChat (stars are a 2026-09-17 GitHub API snapshot) across six dimensions: positioning, multi-user capability, deployment shape, open-source license, model access and data ownership. Key findings: the practical gap between MIT (Octop, LibreChat) and custom licenses (Open WebUI, Dify, FastGPT) is real for redistribution and commercial use, and per-member memory isolation for families and small teams is currently implemented as a first design goal only by Octop. It closes with selection advice for five audiences: personal tinkering, family sharing, small teams, knowledge-base apps and workflow orchestration, and reminds enterprises to read each repository's LICENSE before adopting.

Sep 17, 20268 min read
Hardcore Reviews

Realtime video models compared: who edits while you talk

This review skips image quality and compares form and ownership only: it lines up Vidu S2, Kling, Jimeng, Seedance, Sora 2 and HiDream-O1-Video across real-time interaction and editing ability, delivery form (web, API, open weights, commercial product), open versus closed ownership and fit-for-purpose scenarios. The core claim is that what you buy in a real-time video tool is not fidelity but workflow, whether you can edit while talking, how expensive revisions are, and who owns the artifact. It closes with a selection table by scenario, e-commerce try-on, virtual-host livestreaming, ad shorts and personal tinkering, and warns that real-time quality and cost lack a unified third-party benchmark, so do not let the real-time label set the pace.

Sep 17, 202610 min read