DeepSeek Chat Completions

Call DeepSeek V4.1 Flash, V4 Pro and their high-cache versions through one OpenAI Chat Completions endpoint with a BeatAPI key.

• One endpoint, POST /v1/chat/completions, for every current DeepSeek model. Change model to switch.
• Text input and text output, streaming, function calling, and 1M-token context on the models in the table below.
• Works with the OpenAI SDK: change the base URL and use a BeatAPI key.
• The same models also answer on OpenAI Responses and Anthropic Messages.

Opens this endpoint in the API Explorer. Enter your BeatAPI API key and send: the call is real and is billed to your own balance.

Models

Send one of these IDs as model. Each is a separate model; do not treat them as aliases of each other.

• Prices are the base tier, discount already applied. A long-context tier, cached-input rates and cache writes are not in this table; see the price table.
• — means the catalogue does not publish the number.
• * A peak-time multiplier applies in the model’s peak window; the price shown is the off-peak base price.

Older IDs of this family still work and are listed on Legacy models.

Quick start

Install the openai package on your server, keep the key in BEATAPI_API_KEY, and never put it in browser code. These are request examples, not recorded model outputs.

Read the answer from choices[0].message.content, then check finish_reason and usage. An HTTP 200 alone does not prove the answer satisfies your task: validate it in your application. Text requests return directly; do not poll /v1/tasks for them.

curl --request POST \
  --url https://direct.beatapi.io/v1/chat/completions \
  --header "Authorization: Bearer $BEATAPI_API_KEY" \
  --header 'Content-Type: application/json' \
  --data '{
  "model": "deepseek-v4.1-flash",
  "max_tokens": 2000,
  "messages": [
    {
      "role": "user",
      "content": "Design a resilient webhook retry strategy for a payments API."
    }
  ]
}'

Free versions

These models have a free version. It costs nothing per token, is called on the same endpoint with the free model ID, and has its own per-minute limit: 1 successful request per minute per account before the first top-up and 10 per minute after. Its stability is not guaranteed: it can be rate-limited or withdrawn at any time. See Free models.

deepseek-v4.1-flash-free

curl --request POST \
  --url https://direct.beatapi.io/v1/chat/completions \
  --header "Authorization: Bearer $BEATAPI_API_KEY" \
  --header 'Content-Type: application/json' \
  --data '{
  "model": "deepseek-v4.1-flash-free",
  "messages": [
    {
      "role": "user",
      "content": "Say hello in one sentence."
    }
  ]
}'

deepseek-v4-flash-0731-free

curl --request POST \
  --url https://direct.beatapi.io/v1/chat/completions \
  --header "Authorization: Bearer $BEATAPI_API_KEY" \
  --header 'Content-Type: application/json' \
  --data '{
  "model": "deepseek-v4-flash-0731-free",
  "messages": [
    {
      "role": "user",
      "content": "Say hello in one sentence."
    }
  ]
}'

Endpoint and authentication

POST https://direct.beatapi.io/v1/chat/completions

Request parameters

Standard OpenAI Chat Completions fields are accepted. Request-format compatibility does not mean every model-specific option is supported, so test the options you rely on.

High-cache (HC) models

deepseek-v4-pro-0813-hc and deepseek-v4.1-flash-hc are the same two models served on a high-cache line. When requests share a prompt prefix, such as a long system prompt, a tool list or a conversation that keeps growing, the repeated prefix is billed at the cached-input rate. The regular line costs less per token but does not guarantee a cache hit.

Pick an HC model when most of each prompt is a prefix you send again and again. For prompts with little repetition the regular model is cheaper. Requests are identical to the regular models: only the model ID changes. Check the cached-input count in usage instead of assuming a hit.

Rates by time of day

The regular models, deepseek-v4.1-flash and deepseek-v4-pro-0813, bill one rate at every hour of every day.

The two HC models have a peak window. Every rate doubles on weekdays 09:00–12:00 and 14:00–18:00 Beijing time (GMT+8): input, cached input and output alike. Outside those hours, and all day on Saturday and Sunday, they bill the off-peak rate, which is the one in the model table above and on the price table.

Response

A non-streaming call returns one JSON object. This is an abbreviated example.

{
  "id": "chatcmpl-example",
  "object": "chat.completion",
  "created": 1791993600,
  "model": "deepseek-v4.1-flash",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Retry with exponential backoff and jitter, and make every handler idempotent."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 19,
    "completion_tokens": 16,
    "total_tokens": 35
  }
}

Response fields

Streaming

Set "stream": true. The answer arrives as chat.completion.chunk events and ends with data: [DONE].

• The response is text/event-stream; without streaming it is application/json.
• Text arrives in choices[0].delta.content. A reasoning model sends its reasoning in delta.reasoning_content first.
• A failure before the model is called is returned as a normal JSON error with an HTTP status, not as a stream event.
• Set explicit timeouts and reconnect a streamed response safely.

data: {"id":"chatcmpl-example","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant","content":"Retry"},"finish_reason":null}]}
data: {"id":"chatcmpl-example","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":" with backoff"},"finish_reason":null}]}
data: {"id":"chatcmpl-example","object":"chat.completion.chunk","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}
data: [DONE]

Errors

HTTP errors at a glance

Authentication, routing, rate-limit and unknown-model failures are answered before the model is called, in the BeatAPI envelope: error.code, error.message, retryable, request_id. Every other failure keeps the OpenAI SDK shape {"error": {"message", "type", "code"}}: the code is in error.code and the request id ends message. See Errors.

Full code table and retry policy: Errors.

Notes

• DeepSeek models also answer on OpenAI Responses and Anthropic Messages.
• These models reason before they answer. Reasoning tokens are counted in the output tokens and billed as output, so budget max_tokens for the reasoning as well as the answer.
• Input is text only: image input is not supported on these models.
• Keep API keys on trusted servers, log the response request ID, and retry only 429, 500 and 503.