DeepSeek V4 Pro

DeepSeek V4 Pro is the flagship model in the DeepSeek V4 family. DeepSeek released the official DeepSeek-V4-Pro-0813 build on August 13, 2026, updating its app, website, and API simultaneously. It replaces the April preview while keeping the API model name deepseek-v4-pro.

Pro is a 1.6 trillion parameter Mixture-of-Experts model that activates about 49 billion parameters for each token. It supports a 1 million token context window, up to 384K output tokens through the API, three reasoning-effort modes, tool use, structured JSON, OpenAI-compatible Responses, and an Anthropic-compatible endpoint. DeepSeek also publishes the model weights under the MIT License.

Key specifications

Specification	DeepSeek V4 Pro
Official API build	DeepSeek-V4-Pro-0813
API model ID	deepseek-v4-pro
Release status	General release on app, web, and API
Total / active parameters	1.6T / 49B
Architecture	Mixture-of-Experts with CSA + HCA hybrid attention
Context / maximum output	1M / 384K tokens
Reasoning modes	Non-think, Think High, Think Max
API features	JSON, tools, Responses API, Anthropic API, prefix completion, FIM
Weight precision	FP4 experts plus mostly FP8 remaining parameters
License	MIT

The 49B active count helps explain inference compute, but self-hosters still have to store and distribute the 1.6T-parameter checkpoint. In other words, Pro is open-weight, not small or inexpensive to operate locally. Most individuals and smaller teams will use the hosted API or a specialized inference provider.

Official benchmark comparison

DeepSeek reports the following Pro Max scores against selected frontier models. The table is vendor-reported and configurations are not perfectly uniform: models use their respective high or maximum reasoning modes, and some cells were not reported.

Benchmark	DeepSeek V4 Pro Max	Opus 4.6 Max	GPT-5.4 xHigh	Gemini 3.1 Pro High
MMLU-Pro	87.5	89.1	87.5	91.0
GPQA Diamond	90.1	91.3	93.0	94.3
LiveCodeBench	93.5	88.8	—	91.7
Codeforces rating	3206	—	3168	3052
IMOAnswerBench	89.8	75.3	91.4	81.0
MRCR 1M	83.5	92.9	—	76.3
Terminal Bench 2.0	67.9	65.4	75.1	68.5
SWE Verified	80.6	80.8	—	80.6
BrowseComp	83.4	83.7	82.7	85.9
Toolathlon	51.8	47.2	54.6	48.8

The results show a specialist rather than a universal winner. Pro Max leads this comparison on LiveCodeBench and Codeforces and is strong on math, software engineering, browsing, and tool orchestration. The closed models remain ahead on several science, knowledge, terminal, and long-context tests. Scores should be treated as directional evidence until replicated with the same prompts, budgets, tools, and evaluation harness.

How reasoning modes change performance

Pro offers Non-think for lower latency, Think High for difficult production tasks, and Think Max for the largest available reasoning budget. DeepSeek's own results show large gaps between modes: Pro rises from 56.8 on LiveCodeBench in Non-think to 89.8 in High and 93.5 in Max; Terminal Bench 2.0 rises from 59.1 to 63.3 and 67.9.

More thinking is not automatically the economical choice. Max may consume more output tokens and increase latency, and easy tasks often do not benefit enough to justify it. A production system should route ordinary work to Non-think or High and reserve Max for complex tasks, retries, or cases where correctness has unusually high value.

Pro versus Flash

DeepSeek V4 Flash shares Pro's context window, maximum output, core architecture, API features, and reasoning controls while activating only 13B parameters. Flash is cheaper and better suited to high-volume work. Pro's larger capacity delivers the clearest advantage on factual knowledge, difficult long-context retrieval, browser research, terminal execution, and complicated multi-tool agents.

Flash Max is surprisingly close on some reasoning benchmarks, so Pro should not be selected merely because it is larger. Use workload-level evaluation: if Flash meets the required accuracy, its lower cost and likely higher throughput make it the better production default.

API price

As of August 14, 2026, DeepSeek lists Pro at $0.003625 per million cache-hit input tokens, $0.435 per million cache-miss input tokens, and $0.87 per million output tokens. New rates take effect on August 16: $0.022, $0.66, and $1.98 off-peak, and $0.044, $1.32, and $3.96 during peak hours. DeepSeek defines peak periods as 01:00-04:00 and 06:00-10:00 UTC; verify the live pricing page because rates can change.

When to use DeepSeek V4 Pro

Pro is designed for hard coding, competitive programming, research synthesis, large-repository analysis, long documents, browsing, terminal agents, and workflows where one failed multi-step run costs more than a higher model bill. Open weights can also suit organizations that require deployment control or data residency and have the infrastructure to serve a frontier-scale MoE model.

Flash or a smaller model is usually better for simple chat, classification, extraction, short summaries, and large volumes of routine requests. Pro's benchmark advantage does not remove ordinary model risks: it can hallucinate, follow malicious content, make unsafe tool calls, or lose reliability across very long trajectories. Production agents still need permissions, sandboxing, validation, monitoring, and human escalation.

Bottom line

DeepSeek V4 Pro is no longer a preview. Pro 0813 is DeepSeek's released flagship: a very large, open-weight reasoning and agent model with standout coding results, a million-token context window, and a price and infrastructure footprint aimed at the hardest workloads.