DeepSeek V4 Pro
Flagship 1M-context reasoning for agentic coding
DeepSeek's flagship V4 model: a 1.6T-parameter Mixture-of-Experts with 49B active parameters for agentic coding, math reasoning, and long-document Q&A across a 1M-token context, with optional thinking mode and native tool calling.
Live playground
Try DeepSeek V4 Pro in your browser
Run the model, inspect its schema, then move the same request into production with one imageat API key.
Sign in to use the live model.
Your response will appear here
DeepSeek V4 Pro at a glance
DeepSeek's flagship V4 model: a 1.6T-parameter Mixture-of-Experts with 49B active parameters for agentic coding, math reasoning, and long-document Q&A across a 1M-token context, with optional thinking mode and native tool calling.
- Developer
- DeepSeek
- Released
- April 2026
- Architecture
- Mixture-of-Experts, hybrid compressed sparse attention
- Parameters
- 1.6T total, 49B active per token
- Context window
- 1,000,000 tokens
- Input / output
- Text in, text out (no image, audio, or video input)
- Reasoning
- Thinking mode with low, medium, or high effort, or off
- Tool calling
- Native function calls, parallel calls
- Weights
- Open, MIT license
- imageat model ID
- deepseek/v4-pro
DeepSeek V4 Pro API pricing
API-key requests are billed on the tokens each request actually uses, rounded up to the cent. Reasoning tokens count as output, so thinking mode costs more than direct answers. In the playground, website credits are reserved and settle to the measured cost; most short answers use 1 credit.
Input
$2.64 / 1M tokens
Output (incl. reasoning)
$7.92 / 1M tokens
Need it cheaper and faster? Try DeepSeek V4 Flash.
See all API pricesV4 Flash vs V4 Pro
DeepSeek's published results for both V4 models in their deepest thinking mode. Pro leads everywhere, by the widest margins on factual recall and terminal-based agent work; Flash stays close on math and competitive coding at a third of the price.
| Benchmark | Metric | V4 Flash | V4 Pro |
|---|---|---|---|
| MMLU-Pro | EM | 86.2 | 87.5 |
| GPQA Diamond | Pass@1 | 88.1 | 90.1 |
| Humanity's Last Exam | Pass@1 | 34.8 | 37.7 |
| SimpleQA-Verified | Pass@1 | 34.1 | 57.9 |
| LiveCodeBench | Pass@1 | 91.6 | 93.5 |
| Codeforces | Rating | 3052 | 3206 |
| HMMT 2026 Feb | Pass@1 | 94.8 | 95.2 |
| SWE-bench Verified | Resolved | 79.0 | 80.6 |
| Terminal Bench 2.0 | Accuracy | 56.9 | 67.9 |
| MRCR 1M | MMR | 78.7 | 83.5 |
Source: DeepSeek V4 model card on Hugging Face. Scores are DeepSeek's own and were not reproduced by imageat.
Call the DeepSeek V4 Pro API
One OpenAI-compatible endpoint and your imageat API key. Point any OpenAI SDK, opencode, Cline, or the Vercel AI SDK at https://api.imageat.com/v1; the /v1 is required.
messagesSystem, user, assistant, and tool messages, OpenAI format.enable_thinkingfalse skips reasoning for the fastest, cheapest answer.reasoninglow, medium (default), or high thinking effort.max_tokensCaps the reply when thinking is off (up to 4,000).tools & tool_choiceOpenAI function definitions; replies with tool_calls.parallel_tool_callsAllow several tool calls in one turn (default true).streamOpenAI-style server-sent events.user & session_idOptional identifiers for your end users and sessions.
curl https://api.imageat.com/v1/chat/completions \
-H "Authorization: Bearer iat_live_YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek/v4-pro",
"messages": [
{"role": "system", "content": "You are a senior software engineer."},
{"role": "user", "content": "Find the bug in this function and fix it."}
],
"reasoning": "high"
}'from openai import OpenAI
client = OpenAI(base_url="https://api.imageat.com/v1", api_key="iat_live_YOUR_KEY")
response = client.chat.completions.create(
model="deepseek/v4-pro",
messages=[{"role": "user", "content": "Summarize this contract in five bullet points."}],
extra_body={"enable_thinking": False}, # fastest answer; max_tokens applies
max_tokens=800,
)
print(response.choices[0].message.content)DeepSeek V4 Pro questions
How much does the DeepSeek V4 Pro API cost on imageat?
+
API-key requests are billed on measured usage at $2.64 per 1M input tokens and $7.92 per 1M output tokens, rounded up to the cent per request. Reasoning tokens count as output, so thinking mode costs more than direct answers. With website credits, short answers usually settle to 1 credit.
How big is DeepSeek V4 Pro?
+
DeepSeek V4 Pro is a Mixture-of-Experts model with 1.6T total parameters and 49B active per token. At a 1M-token context it needs 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek V3.2.
How good is DeepSeek V4 Pro at coding?
+
In DeepSeek's published results it resolves 80.6% of SWE-bench Verified, scores 67.9 on Terminal Bench 2.0 and 93.5 on LiveCodeBench, and holds a 3206 Codeforces rating in its deepest thinking mode.
Is DeepSeek V4 Pro OpenAI-compatible?
+
Yes. Send model "deepseek/v4-pro" to https://api.imageat.com/v1/chat/completions with your imageat API key. OpenAI SDKs, opencode, Cline, and the Vercel AI SDK work by setting the base URL to https://api.imageat.com/v1, and streaming is supported.
How do I control thinking?
+
Set "reasoning" to low, medium, or high to choose how long the model thinks before answering, or send "enable_thinking": false for direct answers capped by max_tokens. Only the final answer is returned in message.content.
Does DeepSeek V4 Pro support tool calling?
+
Yes. Pass OpenAI-format tools and the response returns tool_calls with finish_reason "tool_calls", including several parallel calls in one turn.
DeepSeek V4 Pro vs Flash: which should I use?
+
Pro for the hardest agentic coding, research, and factual recall: it leads Flash on every published benchmark, by the widest margin on SimpleQA-Verified (57.9 vs 34.1). Flash costs one third as much and is the better fit for high-volume chat and fast agent loops.
Can DeepSeek V4 Pro read images?
+
No. It is a text-only model, and requests with image content parts are rejected before any charge.
Reason with context
Use the model for analysis, planning, and multi-step knowledge work.
Build and code
Apply it to software development, structured output, and agent workflows.
Generate text
Create clear answers, summaries, drafts, and conversational experiences.
What DeepSeek V4 Pro is built for
DeepSeek's flagship V4 model: a 1.6T-parameter Mixture-of-Experts with 49B active parameters for agentic coding, math reasoning, and long-document Q&A across a 1M-token context, with optional thinking mode and native tool calling.
- Agentic coding and repository-scale changes
- Competition-level math and multi-step reasoning
- Long-document Q&A and synthesis across 1M tokens
- Native function calling with parallel tool calls
Call DeepSeek V4 Pro from your app
Use the OpenAI-compatible chat endpoint with your imageat API key and pay-as-you-go balance.