The DeepSeek homepage links a web chat, a Harness desktop build, an API platform, and API docs. The English pricing page says prices are per 1M tokens and that DeepSeek will bill based on the total number of input and output tokens. Two model names are deepseek-flash, served as DeepSeek-V4.1-Flash, and deepseek-v4-pro, version DeepSeek-V4-Pro-0813. Both list a 1M context and a 384K maximum output. Fees come off a topped-up balance or a granted balance.
Use DeepSeek when you want these model names directly, either in the linked web chat or through the chat-completions call. The call takes a messages list, a required model id of deepseek-flash or deepseek-v4-pro, and an optional thinking switch that defaults to enabled. Image parts accept a URL or a base64 data URL, with low detail downsampling to 512x512. That is a different setup from a product that hides the model id. Money is token math, not a seat price: off-peak cache-miss input is $0.15 for flash and $0.66 for pro, and off-peak output is $0.6 and $1.98. Peak hours on weekdays double those rates.
A good fit for
People who want the web chat linked from the DeepSeek homepage.
Developers sending a messages array to POST /chat/completions.
Teams comparing flash and pro on context, vision, and the token table.
Know the trade-offs
Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday, excluding Chinese public holidays.
Legacy flash names still answer, but they are billed at the Flash price on V4.1-Flash.
JSON output still needs your own instruction to produce JSON, or the docs warn the call can run until the token limit.
Web chat, Harness plugins, and the chat-completions call.
Start from the homepage links
The homepage title is DeepSeek, and the Chinese navigation labels a web version, Harness, an API platform, API docs, API prices, a download, and research entries that include DeepSeek V4.1, V4, V3.2, V3.1, R1, and V3. The banner text announces DeepSeek-V4.1-Flash and says text and agent performance are stronger, that it has native multimodal vision, and that it is faster and lower cost. Two short labels under the hero are deep thinking and smart search. Buttons point at a conversation, the API platform, and the Harness desktop. The Harness page calls the build a preview you can start from for daily work, code, or your own harness. It says the preview is open for a global test and that plugins you combine extend agent ability. The on-page examples are a new session, plugins, automation tasks, and a workspace. Sample prompts include a quarterly data summary deck, a review agenda, and adding comparison columns to a sales report. A displayed reply says Harness can be a desktop app or be started as a web page from code, and it lists help with documents, tables, reports, repository changes, sourced checks, background scripts, and plugins. One demo line names DeepSeek-V41-Flash High.
The chat-completions page says POST /chat/completions creates a model response for the given chat conversation. The body requires messages, a list of messages comprising the conversation so far, with at least one item. Roles documented on the page are system, user, assistant, and tool. User content is either a string or an array of content parts for image input. An image part is an http(s) URL up to 8192 characters or a base64 data URL. Supported formats are JPEG, PNG, GIF, and WebP. Detail is low, high, original, or auto. The page says low downsamples the image to 512x512 (faster, cheaper), while high, original, and auto keep the original image. A file part uses a file id from the Files API or inline file data, and those two fields are mutually exclusive. The model field is required. The page says the id of the model to use is deepseek-flash or deepseek-v4-pro. Thinking is an object whose type is enabled or disabled, default enabled. Reasoning_effort accepts none, low, high, or max. None disables thinking mode. Low, high, and max enable it. The default effort is high. For older clients, minimal maps to low, and medium and xhigh map to high. Max_tokens must be between 1 and 384K, which the page also writes as 393216. When unset, the default is 8K in non-thinking mode and 64K in thinking mode, or 128K when reasoning_effort is max. A beta prefix flag forces the answer to start from assistant content and requires the beta base URL. JSON output is a response_format of type json_object. The page says that guarantees valid JSON, and that you must also instruct the model to produce JSON yourself via a system or user message.
The pricing page lists deepseek-flash and deepseek-v4-pro side by side. Use deepseek-flash as the model name. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those models have been retired, requests are served by the DeepSeek-V4.1-Flash model, and they are billed at the Flash price. The pro version string is DeepSeek-V4-Pro-0813. Both the rows list thinking mode supports non-thinking and thinking, with thinking as the default. Context length is 1M. Max output is 384K. Json output, tool calls, the Responses API, the Anthropic API, and chat prefix completion are marked for both. FIM completion is non-thinking mode only for both. Vision is marked for flash and not supported for pro. Base URL in OpenAI format is https://api.deepseek.com. Base URL in Anthropic format is https://api.deepseek.com/anthropic. Concurrency on the same table is 2500 in the first price column and 500 in the second, which are the flash and pro columns in that order. The page points to a separate rate-limit document for the details behind those caps. A token, the page says, can be a word, a number, or even a punctuation mark.
The prices listed below are in units of per 1M tokens. We will bill based on the total number of input and output tokens by the model. The expense equals the number of tokens times the price. Fees are deducted from the topped-up balance or the granted balance, with a preference for the granted balance first when both exist. Product prices may vary, and the page says to top up from actual usage and to recheck the page. Off-peak rates are half of the peak rates. Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and Chinese public holidays in full. On the cache-hit input row, off-peak is $0.003 then $0.022 and peak is $0.006 then $0.044. On the cache-miss input row, off-peak is $0.15 then $0.66 and peak is $0.3 then $1.32. On the output row, off-peak is $0.6 then $1.98 and peak is $1.2 then $3.96. In each pair the first figure is the flash column and the second is the pro column, matching the model order on the page.
Is to open the web chat linked from the homepage, use the deep-thinking or smart-search entry the hero shows, and only then decide whether the same model should be called from code. The homepage banner is specifically about DeepSeek-V4.1-Flash, which the pricing footnote says is what the deepseek-flash name serves. If the task needs vision, the pricing table marks vision on flash and not on pro, so the suggested check is that column before you attach an image. Those limits, if they exist inside the app, are not on the pages used here.
Call flash with thinking effort set
Another suggested workflow is a single POST /chat/completions. Set model to deepseek-flash or deepseek-v4-pro. Leave thinking enabled when you want the default, or set the type to disabled. If you keep thinking on, the page says the default effort is high, and max raises the unset output budget from 64K to 128K. Put the task in a user message. Add a system message when you need JSON, because json_object format still requires that instruction. For an image, pass an image_url part and set detail to low when you want the 512x512 downsample the page calls faster and cheaper. Keep high, original, or auto when you need the original image. Count tokens against the 1M context and the 384K output cap before you assume a long transcript will fit.
Schedule a large job outside peak hours
A suggested cost check is to read the cache-miss and output rows before a batch. Off-peak flash cache-miss input is $0.15 per 1M tokens and pro is $0.66. Off-peak output is $0.6 and $1.98. Peak doubles those figures to $0.3 and $1.32 on input and $1.2 and $3.96 on output. Cache-hit input is the smaller row, $0.003 and $0.022 off-peak. The page does not say you can turn the cache off. It tells you to top up and to look again because prices may change. Weekends and Chinese public holidays are off-peak all day. Weekday peak is only the two UTC windows. A granted balance is spent before the topped-up balance when both are present. That is the deduction rule on the pricing page.
How to get started with DeepSeek
1
Pick the surface
Open the web chat, the Harness desktop link, or the API docs from the homepage.
2
Choose the model id
Use deepseek-flash or deepseek-v4-pro. Legacy flash names bill at the Flash price.
3
Set thinking and messages
Send at least one message. Thinking defaults to enabled with high effort.
4
Check the token row
Compare off-peak and peak per 1M tokens, then top up the balance the docs deduct from.
The first number in each pricing pair is the flash column and the second is the pro column, as the table is ordered. A granted balance is used before a topped-up balance. Prices can change.
You want Claude's own app rather than an Anthropic-format base URL on DeepSeek.
DeepSeek publishes an Anthropic-format base URL and still serves DeepSeek model names. Claude is the other assistant. Use Claude when the job is not a DeepSeek token balance and a flash or pro id.
You want a different lab's chat instead of DeepSeek-V4.1-Flash.
DeepSeek's public pages center on V4.1-Flash, V4-Pro, the web chat, and Harness. Pick Qwen when you are not buying the DeepSeek per-1M-token rows.
DeepSeek: common questions
Which model names does the pricing page tell you to send?
Use deepseek-flash as the model name. The chat-completions page also allows deepseek-v4-pro. Older names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work, but they are served by DeepSeek-V4.1-Flash and billed at the Flash price.
Prices are per 1M tokens. The page bills the total number of input and output tokens. Off-peak rates are half of the peak rates. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays, excluding Chinese public holidays.
What are the off-peak cache-miss and output prices?
Cache-miss input is $0.15 for flash and $0.66 for pro. Output is $0.6 for flash and $1.98 for pro. Peak on those rows is $0.3, $1.32, $1.2, and $3.96 in the same column order.
The pricing table marks vision for deepseek-flash and says vision is not supported for deepseek-v4-pro. Image parts on the chat-completions page describe URL and base64 input, plus a low detail mode that downsamples to 512x512.
Thinking defaults to enabled. Reasoning_effort defaults to high. None turns thinking off. Max_tokens defaults to 8K when thinking is off, 64K when it is on, and 128K when effort is max. The allowed range is 1 to 384K.
The pricing page says fees are deducted from the topped-up balance or the granted balance, using the granted balance first when both are available. It does not describe a monthly seat on the homepage capture.
NavHunt did not run DeepSeek. No web chat, Harness session, or API call was sent, and no balance was topped up. Prices and parameters come from the English models-and-pricing page and the chat-completions reference public for this check, plus the homepage and Harness page linked from that site. A search snippet with different yuan figures was not used. This description is based on DeepSeek's official materials reviewed on 2026-10-06, including DeepSeek homepage, DeepSeek Harness, DeepSeek models and pricing, DeepSeek chat completions.