DeepSeek OCR, TTS and ASR API pricing: what each one costs

DeepSeek opened standalone OCR, text-to-speech and speech-to-text APIs on Nov 13, 2026, per its launch post. These are the same tools that power the DeepSeek app, now billed per page, per character or per hour of audio on top of normal token charges.

DeepSeek OCR, TTS and ASR API prices

All prices are confirmed on DeepSeek's Models and Pricing page, read Nov 14, 2026. TTS and ASR are at a limited-time 50% discount. DeepSeek hasn't said when it ends, so budget at list price for anything long-running. The page lists one price for these APIs, with no peak or off-peak split like the one on chat tokens. Search and fetch billing has its own guide: DeepSeek Search and Fetch API pricing.

What real jobs cost

Our estimate, using the prices above. Your bill can differ if pages fail or audio rounds up.

The OCR number is the one that stands out. A quarter of a cent per page puts bulk document scanning close to free. TTS is the line that adds up fastest, because every character of a long answer counts, and Chinese text counts double.

How DeepSeek bills each API

• OCR: you pay for pages recognized successfully (the page_count field). Failed or timed-out jobs aren't billed. Closing the connection doesn't cancel a job, so a finished run still bills, per the OCR guide.
• File parsing: you pay the page price once when the upload parses. After that, every chat that references the file pays input tokens at the model's rate, per the file parsing guide. On DeepSeek V4.1 Flash that's $0.30 per 1M input tokens at peak and $0.15 off-peak (from our API price table).
• TTS: billed per character. A Chinese character counts as 2, anything else (letters, digits, spaces) as 1, per the text to speech guide. Markdown is stripped before reading by default.
• ASR: billed by audio length to the second, with any fraction of a second rounded up, per the speech to text guide.

Limits to know before you build

The ASR cap is the catch. Each request takes at most 120 seconds of audio, there's no streaming, and stereo files only get the left channel transcribed. That suits voice notes and chatbot voice input. A podcast or a meeting recording has to be cut into 2-minute chunks first, which means 30 requests for every hour.

TTS has two voices (Stella, female, the default, and Tide, male) and no speed control. The language is detected from the text, so there's no parameter to force one.

OCR API or file parsing: which one to use

• You want the text out of a document: use the OCR API. It returns text page by page, with layout boxes on recognized pages, and doesn't store the file.
• You want the model to answer questions about a document: upload it to the Files API with a parsing purpose and reference the file_id in chat. Same page price, but you pay input tokens on every question.
• The charts matter: pick rich_layout_recognition. The model then sees cropped figures, which use more input tokens. For plain text documents, text_recognition is cheaper to chat with.

How DeepSeek ASR compares with OpenAI

OpenAI's API pricing page lists live transcription (gpt-live-transcribe) at an estimated $0.017 a minute, or about $1.02 an hour (our math). DeepSeek ASR is $0.058 an hour at today's price, roughly 18 times cheaper.

It isn't a like-for-like match. OpenAI's figure is for live streaming transcription, and DeepSeek only takes finished clips of up to 2 minutes. If you need real-time captions, DeepSeek doesn't offer that yet. If you're transcribing short recordings in batches, DeepSeek wins on price by a wide margin.

Verdict

For bulk OCR and short-clip transcription, DeepSeek is now about as cheap as these APIs get, and the docs are clear about what bills and what doesn't. The two things to plan around are the 2-minute ASR limit and the TTS discount, which has no end date. If you already run DeepSeek for chat, see DeepSeek off-peak hours to cut the token side of the bill, and DeepSeek vs Claude API cost or the cheapest LLM APIs to compare models.