GPT-5.5 Takes Text and Images, Not Audio or Video
GPT-5.5 reads text and images and writes text. Audio and video need other models. See per-hour audio prices and a worked example you can recompute.
Correction, October 7, 2026: an earlier version of this post said GPT-5.5 handles audio and video. It does not.
GPT-5.5 reads text and images and answers in text. OpenAI's model page lists no audio or video input.
That settles part of the usual design question, whether to send audio, images and video to one model or split them across specialists. Audio and video cannot go to GPT-5.5 at all, so speech has to become text first, or go to a model built to hear it. What remains is what each step costs and which model should run it.
This post covers what OpenAI, Groq and Google charge for an hour of audio, one worked example you can recompute, and notes on images, video and batch jobs. Prices come from the provider page named in the text and were read on October 30, 2026. Check the page again before you budget.
What GPT-5.5 accepts
- Input: text and images
- Output: text
- Context window: 1,050,000 tokens
- Price: $5 per 1M input tokens and $30 per 1M output tokens
- API snapshot: gpt-5.5-2026-04-23
That price is twice GPT-5.4's $2.50 and $15, a change covered in our post on the price increase. GPT-5.4 mini lists the same input and output types at $0.75 and $4.50. OpenAI's pricing page now also lists a GPT-6 family: GPT-6.1 Sol at $2 and $10, GPT-6 Luna at $0.10 and $0.50, and GPT-6 Astra at $10 and $50. Their model pages also show text and image input and text output, so the reasoning below holds for them.
Three ways to handle audio
Transcribe first, then send text
A speech-to-text model turns the recording into a transcript, and a text model works on the transcript. OpenAI's pricing page lists Whisper at $0.006 per minute and gpt-transcribe at $0.0045 per minute. Groq's speech-to-text page prices its Whisper models per hour. A plain transcript keeps the words but not the tone of voice, so if tone matters, test a model that takes audio.
Use an OpenAI audio model
OpenAI's gpt-audio-1.5 accepts text or audio and can reply in text or audio, through Chat Completions. It is priced per token: $32 per 1M audio input tokens and $64 per 1M audio output tokens, against $2.50 and $10 for text. Run one sample file through it and compare the audio tokens you are billed for with the table below.
Send the audio to Gemini
Google's audio guide says Gemini counts 32 tokens per second of audio, which is 1,920 a minute and 115,200 an hour. A request can mix audio with text, and one prompt can hold up to 9.5 hours. You pay the model's audio input rate for those tokens plus its output rate for the reply.
What one hour of audio costs
The rates come from each provider's pricing page. The last column is the arithmetic for 60 minutes.
| Option | What you get | Rate | One hour of audio |
|---|---|---|---|
| OpenAI Whisper | Transcript | $0.006 per minute | $0.36 |
| OpenAI gpt-transcribe | Transcript | $0.0045 per minute | $0.27 |
| OpenAI gpt-4o-mini-transcribe | Transcript | about $0.003 per minute | $0.18 |
| Groq whisper-large-v3-turbo | Transcript | $0.04 per hour | $0.04 |
| Groq whisper-large-v3 | Transcript | $0.111 per hour | $0.111 |
| Gemini 3.5 Flash-Lite | Answer from the audio, input cost only | $0.30 per 1M tokens, 115,200 tokens per hour | $0.035 |
Notes on the table:
- The first five rows give you a transcript. The Gemini row hears the audio and writes the answer, so it replaces both the transcription and the text call, and its output tokens cost extra.
- OpenAI bills gpt-4o-mini-transcribe by token and shows the per-minute figure as an estimate. Groq bills at least 10 seconds per request, so many short clips cost more than the hourly rate suggests.
- OpenAI's deprecations page lists whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe for removal on February 26, 2027, and names gpt-transcribe or gpt-live-transcribe as the replacements. For new work, plan around gpt-transcribe.
- Some Gemini models price audio above text. Google's pricing page lists Gemini 2.5 Flash at $1.00 per 1M tokens for audio and $0.30 for text, image and video.
- A cheap transcript with more errors can cost more downstream, so test each option on a sample of your own recordings.
Let each request use the cheapest model that's just as good
PromptUnit checks each request your app sends. When a cheaper model is just as good for that request, it uses it; otherwise your model answers. Free until it has saved you money.
Try it free →Worked example: one hour of meeting audio
Assume the hour gives a 12,000-token transcript and you want an 800-token summary. Swap in counts from your own data.
- Transcription with Whisper: 60 minutes at $0.006 = $0.36
- Summary on GPT-5.5: 12,000 input tokens at $5 per 1M = $0.060, plus 800 output tokens at $30 per 1M = $0.024, so $0.084
- Summary on GPT-5.4 mini: 12,000 input tokens at $0.75 per 1M = $0.009, plus 800 output tokens at $4.50 per 1M = $0.0036, so $0.0126
Whisper plus GPT-5.5 costs $0.444 per meeting. Whisper plus GPT-5.4 mini costs $0.3726. Over 1,000 meetings that is $444.00 against $372.60, a gap of $71.40. Transcription is $0.36 of the cheaper total, about 97 percent.
What changes the total most
Changing the speech step moves the bill more than changing the summary model. These totals use the GPT-5.4 mini summary unless noted.
- gpt-transcribe: $0.27 + $0.0126 = $0.2826
- Groq whisper-large-v3-turbo: $0.04 + $0.0126 = $0.0526
- One Gemini 3.5 Flash-Lite call that hears the audio and writes the summary: 115,200 tokens at $0.30 per 1M = $0.0346, plus 800 output tokens at $2.50 per 1M = $0.002, so about $0.0366
These are list prices, not quality results. A product that must keep the transcript also cannot use the one-call route.
Images
GPT-5.5 reads images, so a screenshot or a scanned form can go in the same request as your text. GPT-5.4 mini lists image input at $0.75 and $4.50, and GPT-5.5 costs about 6.7 times as much per token on both input and output. Whether the smaller model reads your images well enough is a test you run: send 50 real images to both and compare the fields you rely on.
Video
GPT-5.5 cannot take a video file. One workaround is to sample frames yourself, send them as images and transcribe the sound separately. Gemini accepts video directly. Google's video guide says it samples one frame per second and uses about 100 tokens per second of video at the default resolution, or about 300 at high resolution.
An hour at the default setting is about 360,000 tokens, or about $0.11 each time at $0.30 per 1M input tokens on Gemini 3.5 Flash-Lite. Ten questions in ten separate requests cost about $1.08 in input alone, unless you cache the video. Our long-context routing post covers that load-once, ask-many trade for text.
Batch jobs
OpenAI's Batch guide takes 50 percent off the price of synchronous calls, with each batch finishing within 24 hours. GPT-5.5 and GPT-5.4 mini both list Batch as a supported endpoint. The guide's endpoint list includes /v1/responses and /v1/chat/completions but not /v1/audio/transcriptions, so the summary step can be batched and the transcription step cannot.
In the example, batching the GPT-5.4 mini summary cuts $0.0126 to $0.0063, which saves $6.30 per 1,000 meetings. The summary was never the large part of the bill, so batch helps little here. It matters more when the text step dominates, and our Batch API post covers which workloads can wait.
Where per-request routing fits
The summary in these pipelines is an ordinary chat request, which is the step where choosing a model per request applies. PromptUnit is a proxy you point your app's AI API base URL at. It checks each request and uses a cheaper model when it is just as good; otherwise the request stays on the model you asked for. It works with your OpenAI, Anthropic, Google, Groq and DeepSeek keys, and with your own OpenRouter key.
In our early data from one production account, about 83 percent of live requests were simple question-answering and about 6 percent were heavy reasoning or complex code. In 24 side-by-side comparisons, a cheaper model matched or beat the chosen model 18 times (75 percent), at roughly 89 percent lower cost on those calls. Across a whole AI bill we estimate a 40 to 70 percent reduction for a workload with many simple requests, but that is an estimate, not a measurement. These are overall figures from a small sample, not a measurement of audio pipelines. The method is in What We Measured Routing GPT-4o Traffic to Cheaper Models.
The dashboard shows each request with the model you asked for, the model that answered, the cost and the saving. To keep a request on the model you asked for, send the header x-promptunit-routing: off. PromptUnit is free until it has saved you money ($5 for individuals, $100 for businesses with a work email), then it takes 20 percent of the savings.
For the full routing framework, read LLM Model Routing: The Complete Guide. For the model-size half of the example, see our GPT-5.4 vs GPT-5.4 mini routing post.
Try It Free
See exactly where your AI budget is going. PromptUnit shows every request with the model that answered, what it cost and what it saved, starting with your first one.
Start free, no credit card needed. Or talk to us if you want a walkthrough.