DeepSeek full-stack API: five capabilities and gateway routing

DeepSeek shipped its "full-stack API" on November 13, 2026 — web search, URL reading, OCR, speech-to-text, and text-to-speech as callable endpoints, marketed as a 1:1 replica of the DeepSeek app. If you already route models through an OpenAI-compatible gateway, you probably want to know one thing: can I keep my current setup and just turn these features on?

I ran five capabilities through a gateway to answer that. This article is organized as a decision checklist rather than a feature tour, because the useful question isn't "does it work" but "what do I have to build myself if I stay here."

AI assisted with drafting and structure. Every result below came from a direct API call.

The checklist up front

Before reading the details, here's the build-vs-buy call for each capability:

  • Speech-to-text: buy. It works out of the box.
  • Text-to-speech: buy, with a name change. It works after swapping the model ID.
  • Web search: build. The gateway hands you a tool call and stops. You supply the search.
  • URL reading: build. Same story as search.
  • Official model name: neither — translate it. The documented ID returns 404 on the gateway.

Read that again as a cost estimate: two of the five are free wins, two are engineering work, and one is a one-line fix you'd never guess from the docs. Here's how I got there.

Full-stack capability breakdown

Test setup (so you can reproduce this)

  • Date: November 17, 2026; URL reading re-tested November 20.
  • Gateway: https://api.crazyrouter.com, OpenAI-compatible /v1/chat/completions. An East Asia node (cn.crazyrouter.com) is documented at the same config.
  • Key: passed via environment variable, never written into the article.
  • Scope: the five capabilities that have a gateway counterpart. Deep thinking search, file input, safety checks, and OCR were not tested, so they're outside every conclusion here. That's "not measured," not "broken."

The one-line fix: the model ID

The docs attach full-stack features to deepseek-flash. Sending that ID to the gateway returns:

model_not_found: Model deepseek-flash is not provided by Crazyrouter

The gateway's own ID, deepseek-v4.1-flash, works. Vendor IDs and gateway IDs are different namespaces — a general property of gateways, not a DeepSeek defect. Look up the real ID in the Crazyrouter model catalog instead of copying from the official docs. One catch: search is currently enabled only on the flash tier, so when you translate the name, confirm the target tier still supports tool calls.

Model name 404

Why "buy" works for speech

STT was the cleanest result. A 6-second Chinese clip through whisper-1 returned HTTP 200 with an accurate transcript. Two gotchas if you implement it: WAV timed out at 90 seconds (HTTP 000) while MP3 worked immediately, so convert before upload; and the transcript came back in Traditional Chinese, so a Simplified-Chinese product needs conversion client-side.

STT transcript

TTS works after the ID swap. The documented tts-1 returns 404; minimax-speech-2.8-turbo returned a real 46 KB MP3 with ID3 tags showing ContentProducer=MiniMax and an AIGC marker — playable audio, not an error page. minimax-speech-2.8-hd is available too. Be accurate in your own docs: this is a speech model the gateway integrated, not DeepSeek's TTS.

TTS MP3 details

Why search is "build": the tool-call round trip

This is the part that costs you engineering time. I passed deepseek_web_search as a function tool, asked what DeepSeek released on November 13, and told it not to rely on memory. The gateway returned HTTP 200, finish_reason = tool_calls, and two correctly formed search calls with real query strings in arguments. Looked perfect.

Then I checked content: empty. The gateway had relayed the tool call to me and stopped. It never ran the search or returned a page. All I had was "let me look that up" and a function name I now own. DeepSeek's private protocol does the search and context stitching on its own server; an OpenAI-compatible gateway only carries the tool call back to your client.

The built-in type: web_search form returns a 400 invalid_request_params — that private tool type isn't recognized here.

So the minimal client you'd build is: detect tool_calls, run the query against a search provider yourself, pack the results back as a tool message, and issue a second request. That's a round trip you implement and maintain. The discriminator that tells you whether a search really happened is simple — look for actual page content and citations in the response, not just a finish_reason.

tool_calls with no results

URL reading is the same build. Passing DeepSeek's announcement URL to deepseek_read_url returned tool_calls, one "I'll read that page" line in content, and zero page content. The built-in read_url form returned a 400 parameter error.

The four costs of building that round trip

These are not from my tests — they're from the official docs and first-hand write-ups, with sources. They're the hidden work in the "build" column.

  1. Date injection. Without a current date, the model guesses "today" from its training data. One test asking what DeepSeek shipped "today" returned February news. The official fix is a fixed-format latest_reminder message inserted immediately before the user message on every request — not once at the top, because the date goes stale. A developer who added it cut a query from 12 searches to 3 before hitting the right day (source: Juejin setup guide).
  2. Sampling parameters get dropped. Declaring web search, URL reading, or a safety check makes temperature, top_p, and thinking.budget_tokens silently ignored server-side. If you need sampling control, don't combine it with those switches (source: Juejin parameter analysis).
  3. Return signature unchanged. Read pages come back encrypted in a signature field your client can't read but must echo in multi-turn conversations — otherwise the request succeeds while the model has "forgotten" the page. It looks like a hallucination; it's a framework dropping an unknown field (source: yage.ai interface analysis).
  4. Stacked billing. Search is per-call, voice is per-character or per-duration, both on top of token cost. One estimate put the cheapest search query and the most expensive long read an order of magnitude apart per call (source: Juejin parameter analysis). Check the official DeepSeek announcement for current rates; I'm not freezing numbers here.

Items 1 and 3 are the real maintenance burden: the less the server does, the more your client owns.

So, build or buy?

  • You need the full app experience — search, file input, the whole thing. Buy the official API at api.deepseek.com. The server executes tool calls and stitches context, so it works as advertised.
  • You mainly need chat and speech, and you're already on a gateway. Buy into the gateway. One Base URL reaches multiple vendors, and both speech endpoints work.
  • You need web search but want to keep the gateway. Build the client: wire the standalone Search endpoint or implement the tool-call round trip above.
  • You need sampling control or structured output. Avoid the built-in search switches on either path — enabling them strips your sampling settings.

The gateway's Base URL is https://api.crazyrouter.com/v1; point any OpenAI-compatible SDK at it. Its value is one entry point across vendors, and that convenience simply doesn't extend to DeepSeek's private-protocol features.

My default recommendation: keep the gateway for chat and speech, and add search client-side only for the specific flows that need it. Before you commit either way, run your own real question through and check whether what comes back is a model request or a server-executed result — that one test tells you which column you're actually in.

👉 Try it with a single API key: https://crazyrouter.com/en/pricing?utm_source=devto&utm_medium=article&utm_campaign=offsite_article&utm_content=cr_7d671d92ccc5_bottom_cta