Can DeepSeek-V4-Pro handle 1M-token documents in the API?

Yes. DeepSeek-V4-Pro with a 1M-token context length and up to 384K output tokens, so it is built for very long documents and multi-file workflows.

Does DeepSeek-V4-Pro support thinking mode and tool calls?

Yes. DeepSeek-V4-Pro supports both thinking and non-thinking modes, plus JSON output and tool calls.

When should I use DeepSeek-V4-Pro instead of DeepSeek-V4-Flash?

Use DeepSeek-V4-Pro when accuracy and agentic coding matter more than speed. DeepSeek says V4-Flash is the faster, more economical option, while V4-Pro is stronger on coding and broader agent evaluations.

Is DeepSeek-V4-Pro good for coding agents like Claude Code or OpenCode?

Yes. DeepSeek-V4-Pro configured for Claude Code and OpenCode, with `reasoningEffort` set to `max` and thinking enabled.

How do I integrate DeepSeek-V4-Pro with OpenAI-compatible SDKs?

Use the CometAPI base URL `https://api.cometapi.com` with the model name `deepseek-v4-pro`

Is DeepSeek-V4-Pro suitable for search-heavy research workflows?

Yes. V4-Pro performs strongly on search and retrieval-style tasks, and it outperforms DeepSeek-V3.2 by a substantial margin in both objective and subjective Q&A categories.

Affordable DeepSeek V4 Pro API | text-to-text

Technical specifications

Item	DeepSeek-V4-Pro
Provider	DeepSeek
API model name	deepseek-v4-pro
Base URLs	https://api.deepseek.com and https://api.deepseek.com/anthropic
Input type	Text
Output type	Text, tool calls, reasoning output
Context length	1,000,000 tokens
Max output	384,000 tokens
Reasoning modes	Non-thinking, thinking (default)
Agent/coding defaults	reasoning_effort can be set as high; complex agent requests may use max
Supported features	JSON Output, Tool Calls, Chat Prefix Completion (beta), FIM Completion (beta in non-thinking mode)
Local/open-weights release	1.6T total parameters, 49B activated parameters, FP4 + FP8 mixed precision
License (model card)	MIT
Reference model card	DeepSeek-V4-Pro preview on Hugging Face

What is DeepSeek-V4-Pro?

DeepSeek-V4-Pro is the stronger member of DeepSeek’s V4 preview family. The official model card describes it as a 1.6T-parameter MoE model with 49B activated parameters and a one-million-token context window, aimed at long-horizon knowledge work, code generation, and agent tasks. The API docs expose it through the standard DeepSeek chat-completions surface and support both OpenAI and Anthropic SDK styles.

Main features

Million-token context: DeepSeek documents a 1M-token context length, which makes the model suitable for very large document sets, repositories, and multi-step agent sessions.
Two reasoning modes: The API supports non-thinking and thinking modes; thinking is the default, and the docs note that complex agent requests such as Claude Code or OpenCode may automatically use max effort.
Tool-call capable: DeepSeek’s thinking mode supports tool calls, which is important for agents that need search, file operations, or external functions.
Long-context efficiency: The model card says V4 uses a hybrid attention design with Compressed Sparse Attention and Heavily Compressed Attention to reduce long-context compute and KV cache cost relative to V3.2. citeturn980363view2
Coding and reasoning focus: DeepSeek says the V4-Pro-Max reasoning mode advances coding benchmarks and closes much of the gap with leading closed-source models on reasoning and agentic tasks. citeturn980363view2
SDK flexibility: It can be accessed through standard OpenAI-compatible chat completions or via DeepSeek’s Anthropic-compatible endpoint for tool-oriented workflows.

Benchmark performance

The official DeepSeek model card reports the following evaluation results for the base model family and for the V4-Pro-Max comparison set. In the base-model table, V4-Pro scores higher than V3.2-Base on several knowledge and long-context benchmarks, including MMLU-Pro (73.5 vs. 65.5), FACTS Parametric (62.6 vs. 27.1), and LongBench-V2 (51.5 vs. 40.2).

Benchmark	V3.2-Base	V4-Flash-Base	V4-Pro-Base
MMLU-Pro (EM)	65.5	68.3	73.5
FACTS Parametric (EM)	27.1	33.9	62.6
HumanEval (Pass@1)	62.8	69.5	76.8
LongBench-V2 (EM)	40.2	44.7	51.5

The same model card also shows V4-Pro-Max remaining competitive with top frontier models on selected tasks. For example, it posts 87.5 on MMLU-Pro, 57.9 on SimpleQA-Verified, 90.1 on GPQA Diamond, and 67.9 on Terminal Bench 2.0 in the published comparison table.

DeepSeek-V4-Pro vs DeepSeek-V4-Flash vs DeepSeek-V3.2

Model	Best fit	Context	Notes
DeepSeek-V4-Pro	Heavy reasoning, coding, agents, large documents	1M	Largest V4 model, 49B activated parameters, strongest overall capacity in the series. citeturn980363view2turn980363view0
DeepSeek-V4-Flash	Faster, lighter general use	1M	Smaller 284B/13B model, still supports thinking and tool calls. citeturn980363view2turn980363view0
DeepSeek-V3.2	Previous-generation long-context baseline	128K in earlier API docs; V4 uses a different 1M context design	Useful as a reference point for efficiency gains; V4-Pro’s model card reports large reductions in long-context FLOPs and KV cache versus V3.2. citeturn321011view1turn980363view2

Best use cases

Repository-scale coding assistants and refactoring tools
Long-document analysis and synthesis
Tool-using agents that need multi-turn reasoning
Technical support workflows that benefit from long memory and structured outputs
Chinese and multilingual knowledge tasks where the model card shows strong benchmark performance

How to access and use Deepseek v4 pro API

Log in to cometapi.com. If you are not our user yet, please register first. Sign into your CometAPI console. Get the access credential API key of the interface. Click “Add Token” at the API token in the personal center, get the token key: sk-xxxxx and submit.

Step 2: Send Requests to Deepseek v4 proAPI

Select the “deepseek-v4-pro” endpoint to send the API request and set the request body. The request method and request body are obtained from our website API doc. Our website also provides Apifox test for your convenience. Replace <YOUR_API_KEY> with your actual CometAPI key from your account. Where to call it: Anthropic Messages format and Chat format.

Insert your question or request into the content field—this is what the model will respond to . Process the API response to get the generated answer.

Step 3: Retrieve and Verify Results

Process the API response to get the generated answer. After processing, the API responds with the task status and output data.Enable features such as streaming, prompt caching, or long-context handling via standard parameters.

Pricing for DeepSeek V4 Pro

Explore competitive pricing for DeepSeek V4 Pro, designed to fit various budgets and usage needs. Our flexible plans ensure you only pay for what you use, making it easy to scale as your requirements grow. Discover how DeepSeek V4 Pro can enhance your projects while keeping costs manageable.

Comet Price (USD / M Tokens)	Official Price (USD / M Tokens)	Discount
Input:$0.416/M Output:$0.832/M	Input:$0.52/M Output:$1.04/M	-20%

Sample code and API for DeepSeek V4 Pro

Access comprehensive sample code and API resources for DeepSeek V4 Pro to streamline your integration process. Our detailed documentation provides step-by-step guidance, helping you leverage the full potential of DeepSeek V4 Pro in your projects.

Python
JavaScript
Curl

from openai import OpenAI
import os

# Get your CometAPI key from https://www.cometapi.com/console/token, and paste it here
COMETAPI_KEY = os.environ.get("COMETAPI_KEY") or "<YOUR_COMETAPI_KEY>"
BASE_URL = "https://api.cometapi.com/v1"

client = OpenAI(base_url=BASE_URL, api_key=COMETAPI_KEY)

stream = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Which number is greater, 9.11 or 9.8? Answer with one sentence."},
    ],
    stream=True,
    max_tokens=256,
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)

thinking = False
for chunk in stream:
    delta = chunk.choices[0].delta
    reasoning = (delta.model_extra or {}).get("reasoning_content") or ""
    content = delta.content or ""

    if reasoning:
        if not thinking:
            print("<reasoning>")
            thinking = True
        print(reasoning, end="", flush=True)

    if content:
        if thinking:
            print("
</reasoning>

<answer>")
            thinking = False
        print(content, end="", flush=True)

print()

Python Code Example

from openai import OpenAI
import os

# Get your CometAPI key from https://www.cometapi.com/console/token, and paste it here
COMETAPI_KEY = os.environ.get("COMETAPI_KEY") or "<YOUR_COMETAPI_KEY>"
BASE_URL = "https://api.cometapi.com/v1"

client = OpenAI(base_url=BASE_URL, api_key=COMETAPI_KEY)

stream = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Which number is greater, 9.11 or 9.8? Answer with one sentence."},
    ],
    stream=True,
    max_tokens=256,
    reasoning_effort="high",
    extra_body={"thinking": {"type": "enabled"}},
)

thinking = False
for chunk in stream:
    delta = chunk.choices[0].delta
    reasoning = (delta.model_extra or {}).get("reasoning_content") or ""
    content = delta.content or ""

    if reasoning:
        if not thinking:
            print("<reasoning>")
            thinking = True
        print(reasoning, end="", flush=True)

    if content:
        if thinking:
            print("\n</reasoning>\n\n<answer>")
            thinking = False
        print(content, end="", flush=True)

print()

JavaScript Code Example

import OpenAI from "openai";

// Get your CometAPI key from https://www.cometapi.com/console/token, and paste it here
const api_key = process.env.COMETAPI_KEY || "<YOUR_COMETAPI_KEY>";
const base_url = "https://api.cometapi.com/v1";

const client = new OpenAI({
  apiKey: api_key,
  baseURL: base_url,
});

const stream = await client.chat.completions.create({
  model: "deepseek-v4-pro",
  messages: [
    { role: "system", content: "You are a helpful assistant." },
    { role: "user", content: "Which number is greater, 9.11 or 9.8? Answer with one sentence." },
  ],
  thinking: { type: "enabled" },
  reasoning_effort: "high",
  max_tokens: 256,
  stream: true,
});

let thinking = false;
for await (const chunk of stream) {
  const delta = chunk.choices[0]?.delta ?? {};
  const reasoning = delta.reasoning_content ?? "";
  const content = delta.content ?? "";

  if (reasoning) {
    if (!thinking) {
      process.stdout.write("<reasoning>\n");
      thinking = true;
    }
    process.stdout.write(reasoning);
  }

  if (content) {
    if (thinking) {
      process.stdout.write("\n</reasoning>\n\n<answer>\n");
      thinking = false;
    }
    process.stdout.write(content);
  }
}

process.stdout.write("\n");

Curl Code Example

#!/usr/bin/env bash
# Get your CometAPI key from https://www.cometapi.com/console/token
# Export it as: export COMETAPI_KEY="your-key-here"

if ! command -v jq >/dev/null 2>&1; then
  echo "jq is required to parse streamed reasoning_content in this shell example." >&2
  exit 1
fi

thinking=false

curl --silent --no-buffer --location --request POST "https://api.cometapi.com/v1/chat/completions" \
  --header "Authorization: Bearer $COMETAPI_KEY" \
  --header "Content-Type: application/json" \
  --data-raw '{
    "model": "deepseek-v4-pro",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Which number is greater, 9.11 or 9.8? Answer with one sentence."}
    ],
    "thinking": {"type": "enabled"},
    "reasoning_effort": "high",
    "max_tokens": 256,
    "stream": true
  }' | while IFS= read -r line; do
    case "$line" in
      data:\ *) data=${line#data: } ;;
      *) continue ;;
    esac

    [ "$data" = "[DONE]" ] && break

    reasoning=$(printf '%s' "$data" | jq -r '.choices[0].delta.reasoning_content // empty')
    content=$(printf '%s' "$data" | jq -r '.choices[0].delta.content // empty')

    if [ -n "$reasoning" ]; then
      if [ "$thinking" = false ]; then
        printf '<reasoning>\n'
        thinking=true
      fi
      printf '%s' "$reasoning"
    fi

    if [ -n "$content" ]; then
      if [ "$thinking" = true ]; then
        printf '\n</reasoning>\n\n<answer>\n'
        thinking=false
      fi
      printf '%s' "$content"
    fi
  done

printf '\n'

Versions of DeepSeek V4 Pro

The reason DeepSeek V4 Pro has multiple snapshots may include potential factors such as variations in output after updates requiring older snapshots for consistency, providing developers a transition period for adaptation and migration, and different snapshots corresponding to global or regional endpoints to optimize user experience. For detailed differences between versions, please refer to the official documentation.

version
deepseek-v4-pro

Technical specifications

Item	DeepSeek-V4-Pro
Provider	DeepSeek
API model name	deepseek-v4-pro
Base URLs	https://api.deepseek.com and https://api.deepseek.com/anthropic
Input type	Text
Output type	Text, tool calls, reasoning output
Context length	1,000,000 tokens
Max output	384,000 tokens
Reasoning modes	Non-thinking, thinking (default)
Agent/coding defaults	reasoning_effort can be set as high; complex agent requests may use max
Supported features	JSON Output, Tool Calls, Chat Prefix Completion (beta), FIM Completion (beta in non-thinking mode)
Local/open-weights release	1.6T total parameters, 49B activated parameters, FP4 + FP8 mixed precision
License (model card)	MIT
Reference model card	DeepSeek-V4-Pro preview on Hugging Face

What is DeepSeek-V4-Pro?

Main features

Million-token context: DeepSeek documents a 1M-token context length, which makes the model suitable for very large document sets, repositories, and multi-step agent sessions.
Two reasoning modes: The API supports non-thinking and thinking modes; thinking is the default, and the docs note that complex agent requests such as Claude Code or OpenCode may automatically use max effort.
Tool-call capable: DeepSeek’s thinking mode supports tool calls, which is important for agents that need search, file operations, or external functions.
Long-context efficiency: The model card says V4 uses a hybrid attention design with Compressed Sparse Attention and Heavily Compressed Attention to reduce long-context compute and KV cache cost relative to V3.2. citeturn980363view2
Coding and reasoning focus: DeepSeek says the V4-Pro-Max reasoning mode advances coding benchmarks and closes much of the gap with leading closed-source models on reasoning and agentic tasks. citeturn980363view2
SDK flexibility: It can be accessed through standard OpenAI-compatible chat completions or via DeepSeek’s Anthropic-compatible endpoint for tool-oriented workflows.

Benchmark performance

Benchmark	V3.2-Base	V4-Flash-Base	V4-Pro-Base
MMLU-Pro (EM)	65.5	68.3	73.5
FACTS Parametric (EM)	27.1	33.9	62.6
HumanEval (Pass@1)	62.8	69.5	76.8
LongBench-V2 (EM)	40.2	44.7	51.5

DeepSeek-V4-Pro vs DeepSeek-V4-Flash vs DeepSeek-V3.2

Model	Best fit	Context	Notes
DeepSeek-V4-Pro	Heavy reasoning, coding, agents, large documents	1M	Largest V4 model, 49B activated parameters, strongest overall capacity in the series. citeturn980363view2turn980363view0
DeepSeek-V4-Flash	Faster, lighter general use	1M	Smaller 284B/13B model, still supports thinking and tool calls. citeturn980363view2turn980363view0
DeepSeek-V3.2	Previous-generation long-context baseline	128K in earlier API docs; V4 uses a different 1M context design	Useful as a reference point for efficiency gains; V4-Pro’s model card reports large reductions in long-context FLOPs and KV cache versus V3.2. citeturn321011view1turn980363view2

Best use cases

Repository-scale coding assistants and refactoring tools
Long-document analysis and synthesis
Tool-using agents that need multi-turn reasoning
Technical support workflows that benefit from long memory and structured outputs
Chinese and multilingual knowledge tasks where the model card shows strong benchmark performance

How to access and use Deepseek v4 pro API

Step 2: Send Requests to Deepseek v4 proAPI

Insert your question or request into the content field—this is what the model will respond to . Process the API response to get the generated answer.

DeepSeek V4 Pro

More Models

Claude Opus 4.7

Claude Sonnet 4.6

GPT 5.5 Pro

GPT 5.5

GPT Image 2 ALL

GPT 5.5 ALL

Related Blog

How to Run DeepSeek V4 Locally

How to Use Deepseek V4 API

Deepseek v4 released: What is and How to Access

DeepSeek v4 is now available on the web: How to access and test it

DeepSeek Update: what changed, what’s new, and why it matters