At Half Price, DeepSeek's New Transcription API Costs Less Than Microsoft's. At List It Costs More, It Stops at Two Minutes per File, Its Voice Has Two Choices, and the Half-Price Offer Has No End Date.
DeepSeek put the features behind its consumer app on the API on November 13: built-in web search and deep-thinking search, URL reading, file parsing, an optional content-safety review, conversation titles, and standalone OCR, speech-to-text, and text-to-speech endpoints, with an open-source demo chat client. Transcription lists at $0.116 per audio hour, $0.058 at a limited-time 50% discount with no end date shown, which at list works out above Microsoft MAI-Transcribe-2's $1.67 per 1,000 minutes on Artificial Analysis and below it only at the promo rate. The speech endpoint takes at most 120 seconds and 8 MiB per request with no streaming, and has no independent accuracy score. Text-to-speech has two voices, no speed control, a content review that can reject text, and costs more than Mistral's Voxtral TTS 3 even at half off. We read DeepSeek's docs and pricing page.
By T.W. Ghost
DeepSeek posted today, November 13, 2026, that "the full DeepSeek App experience" is now available through the API. The chat features that were app-only, built-in web search and deep-thinking search, URL reading, file input, an optional safety check, user metadata, and conversation titles, work on the Messages, Responses, and Chat Completions protocols with the deepseek-flash model, and an open-source demo client shows the wiring. Alongside them are three standalone endpoints: OCR for PDF, Word, PowerPoint, and images; speech-to-text; and text-to-speech. The Flash token prices are unchanged at $0.15 and $0.60 per million off-peak and $0.30 and $1.20 at peak.
The speech prices are the story, and they need two readings. Transcription lists at $0.116 per audio hour, and at a "limited-time" 50% discount, $0.058. By our arithmetic from DeepSeek's list price that is about $1.93 per 1,000 minutes at list and about $0.97 at the promo rate. Artificial Analysis's speech-to-text board, per AA's November 13 thread, lists Microsoft's MAI-Transcribe-2 at $1.67 per 1,000 minutes and Voxtral Small Transcribe 3 at $4.00. So DeepSeek undercuts Microsoft only while the discount lasts, and the discount has no end date, the same pattern as earlier DeepSeek launches. The endpoint also takes a whole file at a time, at most 120 seconds and 8 MiB, with no streaming, reads only the left channel of a stereo file, and has no independent accuracy score that we found. Text-to-speech is the opposite case: at $0.45 per 10,000 characters, $45 per million and $22.50 at half off, it costs more than Mistral's Voxtral TTS 3 at $13 per million characters, per our October 21 read of Mistral's pricing, with two voices and no speed control. We read the X post, the chatbot guide, the pricing page, and the ASR, TTS, OCR, and safety guides. Here is what held up.
What You Are Calling
| DeepSeek app features on the API (November 13, 2026) | |
|---|---|
| Chat features | Web search; deep-thinking search; file input; URL reading; safety check; user metadata; titles; on Messages, Responses, Chat Completions with deepseek-flash |
| Parameter override | With web tools, safety at high, or titles on, the service overrides temperature, top_p, parallel tool settings, and more, per protocol |
| Search and fetch | $0.75 per 1,000 (text or image search; web fetch) |
| File parsing and OCR | $0.225 per 1,000 pages; PDF, Word, PowerPoint, images; 100 MiB, 1,000 pages, 180 s total; file not stored or sent to a chat model |
| Speech-to-text | $0.116 per audio hour list; $0.058 limited-time, no end date; whole file, no streaming; WAV 16-bit PCM or Ogg Opus; stereo uses left channel only; max 8 MiB and 120 s; 21 languages; 50 concurrent requests |
| Text-to-speech | $0.45 per 10,000 chars list; $0.225 limited-time; voices stella and tide only; 10 languages; max 10,000 billed chars and 30 min audio per request; no speed control or resume; 400 if text fails content review ("Content Exists Risk") |
Safety check high | Reviews input and output; may hold the connection open after generation and end in a refusal; docs tell public services to also handle "content safety review, large model registration and AI labeling" |
| Flash tokens | $0.15 / $0.60 off-peak; $0.30 / $1.20 peak per 1M (unchanged) |
| Demo client | cdn.deepseek.com/chat-demo; code at github.com/deepseek-ai/chat-demo |
Two minutes is the ceiling. Fine for voice input to an agent. Useless for a meeting or a podcast without chunking the file and stitching the text, which is engineering the price does not show.
Cheap to hear, pricey to speak. The transcription rate is the bargain, at promo. The voice rate is above Mistral's at list and at half off, and comes with two voices and a content filter that can return a 400 on your text.
The safety switch is written for China's rules. The docs frame it as compliance with content review, model registration, and AI labeling obligations. Turn it on and DeepSeek chooses your sampling parameters.
No board has scored it. DeepSeek's ASR and TTS are not on AA's speech boards as far as we found. Price without a quality number is half a comparison.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Announced Nov 13, 2026; demo client and code | Verified (DeepSeek on X) |
| Chat features and protocols; parameter overrides | Verified (chatbot guide) |
| Search, fetch, OCR, ASR, TTS prices; promo with no end date | Verified (pricing page; ASR guide) |
| ASR limits; TTS voices and limits; OCR limits; safety behavior | Verified (guides) |
| Flash token prices unchanged | Verified (pricing page) |
| MAI-Transcribe-2 $1.67 and Voxtral Small Transcribe 3 $4.00 per 1,000 min | Verified (AA thread, Nov 13) |
| Voxtral TTS 3 $13 per 1M chars | Prior read of Mistral pricing (Oct 21), not rechecked today |
| About $1.93 and $0.97 per 1,000 min | Our arithmetic from DeepSeek's list price |
| Independent accuracy score for DeepSeek ASR or TTS | None found |
What To Do
Use the transcription endpoint for short voice input, where two minutes is plenty and the promo rate is the cheapest row we have seen; chunk anything longer and budget for the stitching.
Skip the voice endpoint on price, and test the content review on your own domain text before shipping either; read the overridden-parameters table before enabling web tools in production.
The frontier is elsewhere: the speech rows with independent scores are on AA's boards, including Mistral's Voxtral 3 line, and the DeepSeek model doing the chat work here is still V4.1 Flash. Our model comparison has the current standings side by side.
Sources
- DeepSeek on X: the full DeepSeek App experience, now via API
- DeepSeek API docs: build a chatbot
- DeepSeek API docs: pricing
- DeepSeek API docs: ASR guide
- DeepSeek API docs: TTS guide
- DeepSeek API docs: OCR guide
- DeepSeek API docs: safety guide
- GitHub: deepseek-ai/chat-demo
- Artificial Analysis on X: speech-to-text and diarization thread, Nov 13
*Not sure whether a cheap transcription endpoint with a two-minute cap fits your pipeline? Take the free 2-minute quiz and get matched. Then read about the usage DeepSeek's own harness is drawing: DeepSeek Harness daily users.*