DeepSeek V4 Pro — high-capability model with 1M context window. Superior reasoning, coding, and agent performance with tool calling and structured output.
DeepSeek-V4-Pro is the flagship tier of DeepSeek's April 2026 V4 preview release. It is a 1.6T-total / 49B-active Mixture-of-Experts model with a full 1M-token context window and Compressed Sparse Attention (CSA) architecture. Released alongside the smaller V4-Flash (284B / 13B active), Pro targets workloads that need frontier-level agentic coding, deep reasoning, and factual reliability where the extra parameter budget delivers measurable gains.
For TheRouter operators, V4-Pro is the open-weight choice when hallucination-sensitive enterprise stacks, multi-step agent loops, or graduate-level STEM tasks justify the higher per-token cost. It supports the same thinking-mode API as V4-Flash (non-thinking / high / max) and is fully OpenAI-compatible via api.therouter.ai/v1. DeepSeek reports that at 1M context, Pro uses 27% of V3.2's single-token FLOPs and 10% of KV cache size — still dramatically more efficient than prior generations despite its larger active parameter count.
DeepSeek serves V4-Pro first-party at api.deepseek.com and api.deepseek.com/anthropic, with OpenAI Chat Completions and Anthropic-compatible request formats, 1M context, 384K maximum output, thinking enabled by default at high effort, and no Responses API support for Pro on the cited pricing page.
TheRouter exposes the same model through api.therouter.ai/v1/chat/completions as an OpenAI-compatible routed endpoint. Requests inherit TheRouter account routing, upstream selection, regional policy, tool-call normalisation, and TheRouter pricing rather than DeepSeek's first-party concurrency limits or first-party billing rates exactly.
| Type | Rate |
|---|---|
| Input | $0.7128 / 1M tokens |
| Output | $2.14 / 1M tokens |
| Release date (preview) | 2026-04-24api-docs.deepseek.com ↗ | verified |
| Architecture | 1.6T total / 49B active MoE; Compressed Sparse Attention (CSA) + DeepSeek Sparse Attention (DSA); 1M context nativeapi-docs.deepseek.com ↗ | verified |
| Context window | 1,000,000 tokensapi-docs.deepseek.com ↗ | verified |
| Max output | 384,000 tokensapi-docs.deepseek.com ↗ | verified |
| Reasoning modes | Non-thinking, Thinking (high / max); reasoning_content returned separatelyapi-docs.deepseek.com ↗ | verified |
| License — code | MIThuggingface.co ↗ | verified |
| License — weights | DeepSeek Model License (commercial use permitted)huggingface.co ↗ | verified |
| Open weights | Yes — available on Hugging Facehuggingface.co ↗ | verified |
| Training tokens | Not publicly disclosed in DeepSeek's V4 launch notes, model card, pricing page, or thinking-mode guideapi-docs.deepseek.com ↗ | unknown |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
Artificial Analysis Intelligence Index DeepSeek's official V4 launch says Pro rivals top closed-source models, but the cited vendor pages do not publish an Artificial Analysis Intelligence Index score for V4-Pro. | — | Not publicly disclosed | — |
LiveCodeBench Pass@1 DeepSeek's official V4 launch describes Pro as open-source SOTA on agentic coding benchmarks, but does not publish a LiveCodeBench Pass@1 score on the cited vendor page. | — | Not publicly disclosed | — |
MMLU-Pro The official V4 announcement claims rich world knowledge for Pro, but the cited DeepSeek pages do not disclose a named MMLU-Pro score. | — | Not publicly disclosed | — |
GPQA Diamond DeepSeek's official V4 launch says Pro beats current open models in Math/STEM/Coding, but the cited vendor pages do not publish a GPQA Diamond score. | — | Not publicly disclosed | — |
SWE-bench Verified The official launch names agentic coding strength but does not disclose a SWE-bench Verified score for V4-Pro in the cited vendor sources. | — | Not publicly disclosed | — |
Terminal-Bench 2.0 DeepSeek's launch positions Pro for enhanced agentic capabilities, but no Terminal-Bench 2.0 score is publicly disclosed in the cited vendor sources. | — | Not publicly disclosed | — |
SimpleQA-Verified The official V4 launch says Pro leads current open models on world knowledge, but the cited vendor pages do not disclose a SimpleQA-Verified score. | — | Not publicly disclosed | — |
BrowseComp DeepSeek's official launch calls out agent capabilities generally, but does not publish a BrowseComp score for V4-Pro in the cited vendor pages. | — | Not publicly disclosed | — |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "deepseek/deepseek-v4-pro",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard OpenAI-compatible chat endpoint. Supports both non-thinking and thinking modes.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "deepseek/deepseek-v4-pro",
"messages": [{"role":"user","content":"Explain quantum error correction in 3 sentences."}]
}'Effective immediately, deepseek-v4-pro input/output pricing moves from the promotional $0.435/$0.87 to the permanent rate. Cache-hit input also reduced to $0.003625/1M. TheRouter continues to price at $2.50/$5.00.
DeepSeek released V4-Pro (1.6T/49B) and V4-Flash (284B/13B) simultaneously. Both support native 1M context, thinking mode, and OpenAI/Anthropic compatible APIs. deepseek-chat and deepseek-reasoner will be retired July 24, 2026.
Choose V4-Pro when factual accuracy, deep agentic loops, or graduate-level STEM tasks are critical enough to justify the Pro tier. DeepSeek's official launch positions Pro as stronger on agentic coding, rich world knowledge, and Math/STEM/Coding; for most high-volume production work, V4-Flash remains the economical default.
Yes. Both models use identical extra_body: { thinking: { type: "enabled" } } and reasoning_effort (high/max) parameters. The reasoning_content field is returned at the same level as content.
| source | URL | retrieved | |
|---|---|---|---|
| Release date (preview) | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| Architecture | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| Context window | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| Max output | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| Reasoning modes | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| License — code | huggingface.co ↗ | 2026-05-25 | verified |
| License — weights | huggingface.co ↗ | 2026-05-25 | verified |
| Open weights | huggingface.co ↗ | 2026-05-25 | verified |
| Training tokens | api-docs.deepseek.com ↗ | 2026-08-07 | unknown |
| Artificial Analysis Intelligence Index | api-docs.deepseek.com ↗ | 2026-08-07 | unknown |
| LiveCodeBench Pass@1 | api-docs.deepseek.com ↗ | 2026-08-07 | unknown |
| MMLU-Pro | api-docs.deepseek.com ↗ | 2026-08-07 | unknown |
| GPQA Diamond | api-docs.deepseek.com ↗ | 2026-08-07 | unknown |
| SWE-bench Verified | api-docs.deepseek.com ↗ | 2026-08-07 | unknown |
| Terminal-Bench 2.0 | api-docs.deepseek.com ↗ | 2026-08-07 | unknown |
| SimpleQA-Verified | api-docs.deepseek.com ↗ | 2026-08-07 | unknown |
| BrowseComp | api-docs.deepseek.com ↗ | 2026-08-07 | unknown |
| DeepSeek makes V4-Pro 75% discount permanent | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| DeepSeek V4 series launches with 1M context and dual-tier MoE design | api-docs.deepseek.com ↗ | 2026-05-25 | verified |
| When should I choose V4-Pro over V4-Flash? | api-docs.deepseek.com ↗ | 2026-08-07 | to verify |
| Does V4-Pro support the same thinking mode API as V4-Flash? | api-docs.deepseek.com ↗ | 2026-05-25 | to verify |