One model, not the range
Start with the limit that decides who can use the built-in version. The new conversation features are tied to deepseek-flash, rather than the separate Pro model. A team whose chat calls run on another model cannot switch on search, file input and URL reading in the same request and keep that model. It can call the standalone Search, Fetch, OCR, ASR and TTS endpoints alongside its chat calls, but then it is back to assembling the context itself, which is the work the built-in version was supposed to remove.
That split is the honest way to read the announcement. DeepSeek's own social post calls this "the full DeepSeek App experience, now via API." What is on offer matches the app in scope: the same tool layer. A developer who reads "full" as "available across the API's models" will be disappointed by the changelog's first paragraph, which names one model and three interfaces and stops there.
The standalone prices are the interesting part
The built-in tools are the headline, but the standalone price list is where a buying decision actually changes. Search and web fetch each cost $0.75 per 1,000 searches or fetches. File parsing and OCR cost $0.225 per 1,000 pages. Speech recognition lists at $0.116 per hour of audio and text-to-speech at $0.45 per 10,000 characters, both at a limited-time 50% discount at the release: $0.058 per hour and $0.225 per 10,000 characters, with no end date stated. Those endpoint charges are separate from model tokens: a chat request on deepseek-flash bills cache-miss input at $0.30 per million tokens at peak hours and output at $1.20, halved off-peak, with cache-hit input at $0.006 per million at peak.
Those are prices a team can compare line by line against the separate search, parsing and speech vendors it already pays. DeepSeek is not asking developers to admire a bundle. It is pricing each part of the bundle so the parts can be bought individually, including by teams that never use a DeepSeek chat model.
Read the caps before running that comparison, because two of them narrow the market. Speech recognition accepts files of at most 120 seconds and 8 MiB, in 16-bit PCM WAV or Ogg Opus, and does not stream. That fits voice notes, short clips and dictated messages. Longer recordings have to be segmented before they are submitted. Text-to-speech offers two voices in 10 languages, which is a functional service rather than a voice library. And the standalone Search API, unlike the built-in search tool, does not do image search yet; DeepSeek's guide says image search is available only through the web tools for now.
Safety, at the level DeepSeek claims for it
The fifth built-in feature is a safety check that reviews input and output and returns a refusal marker when it finds a risk. DeepSeek calls this basic risk control, and its changelog says developers who provide a service to the public should enable it and fulfill obligations such as content safety review, large model registration and AI labeling. In other words, the operator's duties remain the operator's; switching on the parameter does not discharge them.
For a developer deciding this week, the sequence is short. If you already build on deepseek-flash, switch on search and file input in a test branch and measure the multi-step search mode's latency, because a tool that reads links over several reasoning steps will not respond like a single model call. If you build on another model entirely, price the standalone OCR and ASR endpoints against your current bill, within their caps. And if you need these features attached to a different model, note that this release does not offer that, whatever the headline around it suggests.
Sources
- [1] DeepSeek, API changelog, "Releasing the full suite of application APIs" (November 13, 2026)Read source
- [2] DeepSeek, Models & Pricing (as listed November 13, 2026)Read source
- [3] DeepSeek, Speech to Text guideRead source
- [4] DeepSeek, Text to Speech guideRead source
- [5] DeepSeek, Search & Fetch guideRead source
- [6] DeepSeek, Document Parsing guideRead source
- [7] DeepSeek (@deepseek_ai), announcement post (November 13, 2026)