DeepSeek Ships Its Whole App as an API: Search, Files, Voice and Mandatory Safety Checks on a $0.30 Flash Model
On November 13, DeepSeek exposed the features behind its consumer app through its API: server-side web search, file parsing, URL reading, OCR, ASR and TTS, all anchored on deepseek-flash. The price list and the compliance clause are the real story.
For most of the past two years, the story of a Chinese frontier lab has been told in model cards: parameter counts, context windows, benchmark tables. On November 13, 2026, DeepSeek published something that does not fit that template. There is no new set of weights and no leaderboard claim. Instead, the Hangzhou lab announced that the capabilities behind its consumer app are now callable through its developer API. The official release note↗ puts it plainly: search, files, links, safety and voice, "now available through the DeepSeek API."
From a technical-comparison standpoint, this is the more interesting kind of release. It tells us where DeepSeek thinks the value in the stack sits once raw model quality across Chinese and Western labs has converged within a few points. The answer, apparently, is in the scaffolding: the retrieval, parsing and moderation layers that turn a model into a product.
What DeepSeek actually shipped
The anchor is deepseek-flash, the API alias for DeepSeek-V4.1-Flash-1017. According to the pricing page↗, it offers a 1M-token context window, up to 384K tokens of output, vision support, tool calls, JSON output, and compatibility with both the Anthropic-style and OpenAI-style protocols. The new features work across all three of DeepSeek's endpoints: Messages, Responses and Chat Completions.
The built-in capabilities, as listed in the release note and the Build a Chatbot guide↗, are:
- Web search in non-thinking mode, where the model runs web or image searches as the question requires and answers with citations, enabled through a `deepseek_web_search` tool.
- Deep thinking search, where, with thinking enabled, the model searches and reads pages on its own across multiple reasoning steps rather than making a single retrieval call.
- File input, which parses PDFs, Word and PowerPoint files and images into conversation context, using either plain-text extraction or rich-layout parsing.
- URL reading, switched on with `deepseek_read_url`, where the model detects links in a user message, fetches the pages and answers with them in context.
- Safety checks, a parameter that reviews both model input and output and returns a refusal marker when a risk is detected.
Beyond the chat features, Search, Fetch, OCR, ASR and TTS are also exposed as standalone APIs, so a developer who only wants speech recognition or page extraction does not have to route through the language model at all. As TokenPost reported↗, DeepSeek had previously opened Search and Fetch access to developers; this release adds the rest of the bundle and fuses it with the chat endpoints.
Server-side tool execution is the architectural point
The detail that matters most is where the tool loop runs. The guide states that "the service runs the searches and URL reading," and that the client "only shows the returned blocks and passes them back unchanged in the next round." In other words, DeepSeek's servers execute the tool calls and assemble the context. This removes an entire layer that developers otherwise write themselves: the orchestration code that parses a tool request, calls a search provider, truncates results, and feeds them back.
This is a deliberate move away from the "bring your own tools" pattern that dominates Western agent frameworks. It trades flexibility for convenience and, as we will see, for control.
"The full DeepSeek App experience, now via API." (DeepSeek release note, November 13, 2026)
There is a cost to that convenience. When a request declares the web tools, sets safety to high, or enables conversation titles, the service decides a long list of sampling parameters itself. For Chat Completions, `temperature`, `top_p`, `frequency_penalty`, `presence_penalty`, `stop`, `seed` and `n` are all ignored. For the Responses API, `temperature`, `top_p` and `max_tool_calls` are among the overridden parameters. Researchers who need reproducible, seeded runs will find these modes unsuitable for evaluation work, which is a meaningful limitation when comparing against models you can fully control.
The price list is the quiet headline
Chinese labs have used price as a competitive weapon throughout 2026, and this release continues the pattern, but with a twist: the new capabilities are billed as separate line items on top of token charges. The pricing page↗ lists the following for `deepseek-flash`:
- Peak hours: $0.30 per million cache-miss input tokens and $1.20 per million output tokens, with cache hits at $0.006.
- Off-peak hours: exactly half of peak, at $0.15 input, $0.60 output, and $0.003 for cache hits.
- Search and web fetch: $0.75 per 1,000 calls each, whether invoked through the standalone APIs or by the model via tools.
- File parsing and OCR: $0.225 per 1,000 pages.
- Speech: text-to-speech at $0.45 per 10,000 characters (currently 50% off, at $0.225) and speech-to-text at $0.116 per hour of audio (currently $0.058).
For comparison, the larger `deepseek-v4-pro` (DeepSeek-V4-Pro-0813) sits at $1.32 input and $3.96 output at peak, and does not support vision. Off-peak windows apply on a schedule: peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays, and everything else, including weekends, is off-peak.
A quick worked example shows why the structure matters. A research agent that issues ten searches and fetches ten pages per task pays about $0.015 in tool fees, before token costs. That is trivial next to the tokens themselves, which suggests DeepSeek is pricing tools to remove friction rather than to extract margin. The 50% promotional discount on speech points the same way: the company wants the voice endpoints adopted quickly.
How this compares globally
Western labs already bundle search and file handling into their APIs, so DeepSeek is not inventing the category. What differs is the price floor and the openness of the reference client. A fully working assistant, with search, uploads, voice and history, can be assembled from DeepSeek's endpoints at fractions of a cent per interaction. We have not benchmarked equivalent bundles from Western providers here, and the comparison above rests only on DeepSeek's published rates. Readers pricing a real workload should run their own numbers.
The demo, and why it is MIT-licensed
DeepSeek also released an open-source reference client, deepseek-ai/chat-demo↗, under the MIT license. It is a static app with no backend: the browser connects directly to the configured API, and the API key, settings and conversations stay in the browser's localStorage and IndexedDB. It supports streaming, deep thinking, link reading, citations, attachments, voice input, read-aloud, Markdown with KaTeX and Mermaid diagrams, and both Chinese and English interfaces.
One feature stands out: import of DeepSeek chat history. Users can export their conversations from the DeepSeek app and load them into a client built on the API. The repository notes that an import replaces all existing local conversations. As Crypto Briefing observed↗, the demo gives builders a working reference for how the pieces fit together, and it could shrink the workload for smaller teams considerably.
The strategic reading is straightforward. If a third party can stand up a near-clone of the DeepSeek app in an afternoon, and users can bring their history with them, DeepSeek's API becomes the default backend for a long tail of regional and vertical assistants. That is distribution without the cost of running a consumer product everywhere.
The compliance clause that comes with it
The most revealing sentence in the documentation is not about technology. It reads: "Public-facing apps must enable safety checks. Developers remain responsible for content moderation, applicable model registration, and AI-generated content labeling."
The guide repeats the point in its compliance section: anyone providing a service to the public should enable the safety parameter and "fulfill on your own the obligations of a service provider, such as content safety review, large model registration and AI labeling." This maps directly onto China's regulatory regime for generative AI, in which service providers are expected to register models and label synthetic content.
Practically, this means:
- Safety checks are not optional for public apps, which gives DeepSeek a standardized moderation layer across every downstream deployment and a defensible compliance posture with regulators.
- Responsibility is pushed to the developer, so a startup building on the API inherits registration and labeling duties rather than DeepSeek absorbing them.
- International builders face a choice, because a mandatory Chinese-regime safety layer may be acceptable for some products and a dealbreaker for others, particularly in regulated Western sectors.
"The mandatory safety checks are the piece to watch. They signal that DeepSeek wants a say in how its tools appear in public." (Crypto Briefing, November 13, 2026)
That reading is the outlet's interpretation rather than a DeepSeek statement, but it is consistent with the text of the documentation.
Context: an agent-first month for DeepSeek
The API release lands in a month when DeepSeek has been steadily building toward agents. Bloomberg reported↗ on November 2 that the lab introduced a "creator mode" that lets users turn ideas into plugins using plain-English instructions, with demonstrations including a 3D model renderer and a video editing workbench. The same report said DeepSeek's Harness team, formed earlier this year to build agents that compete with AI coding services, had released an enhanced coding assistant designed to proactively help resolve issues with minimal human attention. Bloomberg defined a harness as "the software scaffolding that wraps around a large language model."
That definition explains the November 13 announcement. A harness needs retrieval, file understanding, voice and guardrails. DeepSeek has now productized those parts and made them addressable by anyone. Bloomberg also noted the lab has raised about 100 billion yuan (roughly $15 billion) in its latest round ahead of an initial public offering, which gives context for why a company with a famously lean research culture is now investing in developer-platform breadth.
What to watch and how to try it
Practically, anyone with a DeepSeek API key can test this today. The model name is `deepseek-flash`, the tools are declared in the request body, and the open-source demo accepts an API key directly in its settings. Several open questions will determine whether the release matters beyond developer convenience:
- Adoption of standalone OCR, ASR and TTS, which will show whether these endpoints compete on their own merits against specialist providers or only serve as add-ons to the chat model.
- Quality of server-side search, which DeepSeek has not benchmarked publicly in the release materials, so independent evaluations of citation accuracy are needed.
- Reproducibility trade-offs, since the ignored sampling parameters limit use in research settings that depend on deterministic behavior.
- Reaction to the safety mandate, particularly from non-Chinese developers who may route around it by keeping the tools off and using the base model alone.
None of these is answered by the announcement. What the announcement does establish is a direction: DeepSeek is competing on the surrounding platform as well as the weights. For a lab that built its reputation on efficiency, that is a notable broadening of ambition.
Bottom line
The November 13 release is modest in headline terms and significant in structure. A 1M-context flash model, server-side tools priced at $0.75 per thousand searches, an MIT-licensed reference app, and a compliance clause that travels with every public deployment add up to a package aimed at becoming the default backend for assistants built inside China's regulatory frame, and priced to tempt builders beyond it. Whether the rest of the world accepts the bundle, safety clause included, will say a great deal about how far Chinese AI infrastructure can travel as the model race itself flattens out.
Links & Resources
External links — opens in a new tab

Electrophysiological Biomarkers of Neuropsychiatric Brain Dynamics Vol 1
by Richard Murdoch Montgomery
EEG-based biomarkers for schizophrenia and bipolar disorder — frequency band power, event-related potentials, and neural connectivity patterns.

The HP 17BII Financial Calculator
by Richard Murdoch Montgomery
A 50-chapter treatise integrating financial mathematics, business reasoning, and Solver-based modeling — from annuities to investment analysis.

Artificial Intelligence: Origins and Developments
by Richard Murdoch Montgomery
A comprehensive survey of AI from Turing machines to deep learning — neural networks, expert systems, and the philosophical debates that shaped the field.

Glioblastoma Growth Modelling
by Richard Murdoch Montgomery
Mathematical oncology meets computational neuroscience — reaction-diffusion models, imaging-driven simulations, and treatment optimisation.
Comments
Open discussion — no account needed. Be respectful.
More from Chinese Models Desk
Meituan Gives Away LongCat-2.5, a 1.6T-Parameter Agent Model, Under MIT Terms
Meituan open-sourced LongCat-2.5, a 1.6-trillion-parameter multimodal MoE trained on domestic Chinese chips, under the MIT license with no revenue thresholds. The food-delivery giant is betting that permissive weights, not a token business, buy it a place in the developer toolchain.
Baidu's ERNIE 5.5 Flash Preview Bets on 14B Active Parameters, With Apache 2.0 Weights Still a Promise
Baidu's 280B-total, 14B-active ERNIE 5.5 Flash arrives as an open base model and a Qianfan-hosted Preview, with a 1M-token context and Apache 2.0 weights reportedly due November 25. We separate what Baidu has documented from what is still a claim.
Alibaba Opens Qwen4-Max Preview: 1M Tokens, Two Weeks, and No Benchmarks
Alibaba's Qwen4-Max preview offers a 1M-token window and video input through Token Plan until November 25, but ships without benchmarks, pricing, weights or a license.