DeepSeek Opens App APIs: Search, File Input and Link Reading in a Few Parameters
Two announcements in the same week were not about a model getting stronger. They were about AI being taken apart.
On November 13, DeepSeek opened up the capabilities behind its consumer app to developers. Turning on the relevant parameters in the existing API makes web search, file input and link reading available; tool calls and context assembly are handled on the server side. V4.1-Flash supports them under the Messages, Responses and Chat Completions interfaces. Voice and lower-level abilities have their own endpoints — turning images into text, speech into text, and text back into speech — available individually or in combination. DeepSeek also published an open-source front-end demo that can import a user's chat history.
Between November 12 and 13, Google DeepMind released Gemini 4 Hydrogen, positioned for fast execution of everyday coding and agent workflows. Google called it the largest model upgrade to Search so far, wiring it into search to improve answers to complex questions. It also became the default model for the Antigravity agent inside Gemini Managed Agents, and arrived in Cursor as well as Cognition's Devin Desktop and Devin CLI.
One takes a finished product apart and hands the pieces to builders. The other turns a model into an execution component inside somebody else's workflow. Different directions, same outcome: AI is becoming a set of parts installed into products, rather than a place you go to ask a question.
That is good news. Read the two announcements side by side, though, and only half of the barrier has actually moved.
One: the half that moved
📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.

Consider what no longer has to be built.
An application that could search the web, read an uploaded file and open a link used to require its own tool wiring: calling each service, handling failures, and stitching whatever came back into the model's context. That work now happens on the server. The voice and low-level abilities are separate endpoints, so you take the one you need instead of carrying the whole bundle. The open-source front-end demo removes another chunk of setup by letting existing chat history be imported directly.
Tianyi Cui, an investor who reposted the announcement, put the builder's view plainly: a few parameters are enough to enable built-in functions and recreate the full app experience inside your own application. He added that the next super app might belong to just one person.
Google moved a different kind of work. Hydrogen was given an explicit job — fast execution inside agent workflows — and placed inside Search, inside Antigravity, and inside other people's coding tools. The model is no longer the product. It is the part of the product that does the work.
In both cases, assembly is shifting from something you do to something that has been done for you.
Two: the half that stayed
Now read the shortest, heaviest sentences in the release notes.
DeepSeek states that using extended capabilities incurs tool-call charges on top of token charges, with specific pricing to be listed in a separate announcement. The bill, in other words, changes from a per-token line to a per-token line plus a per-call line — and the per-call column currently has no published rates.
The same release says that developers offering application services to the public must enable the content-safety parameter and take on content-safety review, model registration and AI labelling obligations themselves.
Those two clauses belong together. Capabilities arrive bundled; obligations do not. A handful of parameters is enough to assemble something that searches, reads files and speaks, but once that thing serves other people, who reviews the content, who registers the model and who labels the output is entirely the assembler's problem.
The account is easier to see as two columns.
| Moved to the server | Left with the builder |
|---|---|
| Tool calls and context assembly | Enabling the content-safety parameter |
| Web search, file input, link reading via parameters | Content-safety review, model registration, AI labelling |
| OCR / ASR / TTS as independent, combinable endpoints | Tool-call fees on top of tokens, pricing not yet published |
| An open-source front-end demo and a few parameters | Full responsibility once the service is public |
The left column keeps growing. The right column gets heavier the longer you look at it. That is not a flaw in either release. It is what opening up a capability means: what moves is work, not ownership.
Three: where the edges are

Three lines are worth holding.
Do not invent a price for tool calls. The official wording is that charges apply and pricing will be published separately; there is no public rate card yet. Any figure attached to it is speculation.
Do not read "available individually" as "liable individually". Separate endpoints for OCR, speech-to-text and text-to-speech make integration more flexible. They do not distribute review obligations among the providers.
Do not settle these announcements with benchmark scores. The numbers circulating this week come from third-party evaluators using different methodologies, and the question here is how capability and responsibility are divided, not which model is faster. By the same logic, there is no useful ranking of which company is "more open" — they opened different layers. One scope note also matters: Google's rollout is tiered by user type, with different entry points for all users, paid subscribers, and developers and enterprises. Do not plan an integration on the assumption that everyone gets the same access.
Four: using the two columns
The table generalises well beyond these two releases. The next time a launch promises open capabilities, one-click integration or ready-made tools, draw a vertical line down a page.
On the left, write what the provider finished for you. On the right, write what you still have to do. The right column needs at least three rows: how the bill is calculated, who reviews the content, and who carries the consequences.
Counting only the left column leaves the right one outside the budget. As capabilities turn into parts, the number of people assembling things grows faster than the number of people who can say exactly what they assembled. The barrier is no longer whether you can build it. It is how much you can carry after it is built.
Capabilities can be packaged; responsibility cannot be outsourced. Tool calls can be assembled on a server, but content review, registration and labelling are not included in the box. KAIHE AIBOX keeps the agent and the files on the machine you own, so which tool was called and which file was read stays visible on your side. Whether to wire these parts in is your call. Seeing clearly which half was left with you costs less than reconstructing a compliance checklist afterwards.