Editly

DeepSeek V4 Pro 0813

DeepSeek-V4-Pro-0813 is the dated general-availability snapshot of DeepSeek's flagship model: a 1.6-trillion-parameter mixture-of-experts with 49 billion active parameters, a 1M-token context window, and up to 384K tokens of output. It thinks by default, exposes a non-thinking endpoint for latency-sensitive calls, ships MIT-licensed weights, and bills $0.435 per million input tokens against $0.87 per million output.

Start creating

What Is DeepSeek V4 Pro 0813?

0813 is a snapshot tag, not a new architecture. It pins the GA build of DeepSeek V4 Pro that appeared in DeepSeek's official API docs on August 12, 2026, sitting beside the cheaper DeepSeek-V4-Flash-0731. Here is what the flagship tier ships with.

1.6T Parameters, 49B Active

A sparse mixture-of-experts that activates roughly 3 percent of its weights per token. DeepSeek pairs Compressed Sparse Attention with Heavily Compressed Attention, adds Manifold-Constrained Hyper-Connections, and trains with the Muon optimizer. Experts run in FP4 while the rest of the stack stays FP8.

1M-Token Context, 384K-Token Output

A full million tokens in, which is roughly a mid-size codebase or a few thousand pages of documents, and up to 384,000 tokens back out in one response. Long-horizon agent runs and whole-repository reviews fit in a single call instead of a chunking pipeline.

Thinking by Default, Three Effort Levels

The model reasons before answering unless you opt out. Effort runs Non-think, Think High, and Think Max; the Max setting is the configuration DeepSeek benchmarks as V4-Pro-Max and the one that closes most of the gap to closed frontier models on reasoning and agentic tasks.

Built for Agent Developers

Structured JSON output, tool calls, the Responses API, and an Anthropic-API-compatible endpoint, so existing clients point at DeepSeek by swapping a base URL. Conversation-prefix continuation and fill-in-the-middle completion are available in beta.

MIT-Licensed Open Weights

DeepSeek publishes deepseek-ai/DeepSeek-V4-Pro and DeepSeek-V4-Pro-Base on Hugging Face under the MIT license. You can serve it yourself, fine-tune it, or use it commercially without asking permission — the constraint is hardware, not terms.

Frontier-Adjacent Scores at Open Prices

80.6 on SWE-bench Verified, 93.5 on LiveCodeBench, a 3206 Codeforces rating, and 90.1 on GPQA Diamond, at output pricing under a dollar per million tokens. The gap to the top closed models is real but no longer an order of magnitude.

DeepSeek V4 Pro 0813 vs DeepSeek V4 Flash 0731

Same 1M context and same 384K output ceiling on both tiers. What you buy with Pro is capacity — more total and active parameters and stronger reasoning — at roughly triple the token price and a fifth of the concurrency allowance.

V4 Flash 0731
V4 Pro 0813Pro
Total parameters284B1.6T
Active parameters per token13B49B
Context window1M tokens1M tokens
Max output384K tokens384K tokens
Thinking modeThinking on by default, non-thinking availableThinking on by default, plus Think High and Think Max
Input per 1M tokens, cache miss$0.14$0.435
Output per 1M tokens$0.28$0.87
API concurrency limit2500500
Best forHigh-volume classification, extraction, chat, cheap long-context passesHard reasoning, agentic coding, deep research, whole-repo work

DeepSeek V4 Pro 0813 Benchmarks

Published scores for DeepSeek-V4-Pro at maximum reasoning effort, the configuration DeepSeek reports as V4-Pro-Max. Treat vendor-reported numbers as a ceiling and re-run your own evals before committing a pipeline.

SWE-bench Verified: 80.6

Real GitHub issues resolved end to end. That is squarely in agentic-coding territory, though the strongest closed models still sit above it on their best configurations. SWE-Bench Pro, the harder variant, comes in at 55.4.

LiveCodeBench: 93.5

Contamination-resistant competitive coding drawn from contests published after training cutoffs, so the score reflects reasoning on unseen problems rather than memorized solutions.

Codeforces: 3206

A rating in grandmaster range on algorithmic contest problems. Useful as a proxy for tight, correct implementations under constraints, less so for navigating a messy production codebase.

GPQA Diamond: 90.1

Graduate-level physics, chemistry, and biology questions written to defeat search. Ninety percent means the model holds up on technical Q&A well outside code.

HMMT February 2026: 95.2

Olympiad-level math from a competition held after the model's training window closed. Multi-step proof-style reasoning, not arithmetic.

Terminal-Bench 2.0: 67.9 and BrowseComp: 83.4

The two agentic benchmarks that matter for tool use: driving a real shell to completion, and finding hard-to-locate facts across the live web. Both are where the 1M context and default thinking mode pay off.

DeepSeek V4 Pro 0813 Pricing

Rates per million tokens on DeepSeek's own API. Context caching is the line item worth engineering around: a cache hit costs about one one-hundred-and-twentieth of a miss, so a stable system prompt or a pinned repository snapshot pays for itself within a few calls. DeepSeek has warned that overall API pricing will rise significantly in the near future — check its pricing page before you model a budget.

V4 Flash 0731
V4 Pro 0813Pro
Grok 4.5
Input, cache hit$0.0028$0.003625Not published
Input, cache miss$0.14$0.435About $2
Output$0.28$0.87About $6
Context window1M tokens1M tokens500K tokens
WeightsOpen, MITOpen, MITClosed
Cost to fill the full context onceAbout $0.14About $0.44About $1 for 500K

How to Use DeepSeek V4 Pro 0813

  1. 01

    Step 1: Pin the Model String

    Call deepseek-v4-pro on DeepSeek's platform for the rolling flagship, and reference the DeepSeek-V4-Pro-0813 snapshot when you need a build that will not shift under a running evaluation. The endpoint speaks the OpenAI format, so most SDKs need only a base URL and a key.

  2. 02

    Step 2: Choose Reasoning Effort and Cache Your Prefix

    Leave thinking on for planning, debugging, and multi-step tool use; switch to the non-thinking endpoint for latency-bound calls like autocomplete or routing. Keep your system prompt and any large fixed documents at the head of every request so they land as cache hits.

  3. 03

    Step 3: Or Route Around DeepSeek Entirely

    Aggregators such as OpenRouter list deepseek/deepseek-v4-pro-0813, and inference providers serve it hosted. Because the weights are MIT-licensed, you can also run it on your own hardware — budget for a 1.6T-parameter MoE, which in practice means a multi-GPU node even in FP4.

  4. 04

    Step 4: Point the Output at Something

    A frontier text model is upstream infrastructure — it writes the plan, the code, or the prompt. If the deliverable is a video or an image, hand the text it produces to a generation model: paste a DeepSeek-written prompt straight into Editly and render it.

Frequently Asked Questions about DeepSeek V4 Pro 0813

Turn a DeepSeek Prompt Into Video

Let a frontier text model write the shot list, then render it. Sign up for Editly and run Seedance, Veo 3, Hailuo, and Qwen Image on one credit balance — no per-model subscription.

Try Editly Free