All models

Page last updated on

DeepSeek logo

DeepSeek-V4-Pro-0813

By DeepSeek

MoE1M contextAgenticResponses API MIT license

The official DeepSeek-V4-Pro build. A 1.6T MoE with 49B active parameters and 1M context, retrained for agents and coding, with 87.9 on Terminal Bench 2.1.

Hugging Face Paper

Specifications

Total parameters
1.6T (49B active)
Active parameters
49B
Architecture
MoE
Architecture class
DeepseekV4ForCausalLM
Attention
Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA)
Context window
1M
Vocab size
129,280
Precision
FP4 + FP8 Mixed
Modality
Text
License
MIT
Released
August 2026
Recommended hardware
8× H2008× B200
Best for
Complex agent workflows, coding, security research, and long-context analysis

Good to know

  • DeepSeek-V4-Pro-0813 is the official release of DeepSeek-V4-Pro, superseding the April 2026 preview. It went live on the API on August 12, 2026, and DeepSeek published the release notes on August 13.

  • DeepSeek open-sourced an agent harness alongside this model, DeepSeek Harness, which is MIT-licensed. It is the framework DeepSeek used to produce the code-agent benchmark scores on this page. Note that version 0.1 is a developer preview and the maintainers warn of compatibility-breaking changes.

  • This build natively supports the OpenAI Responses API format with a one-click Codex setup, matching what V4-Flash-0731 shipped in July.

  • The API price increase takes effect at 16:00 UTC on August 16, 2026, and it is large. See the pricing section for the peak and off-peak tables.

Architecture

DeepSeek-V4-Pro-0813 keeps the V4 architecture unchanged and rebuilds the post-training, the same playbook that produced DeepSeek-V4-Flash-0731 two weeks earlier.

Two things are specific to this checkpoint:

  • DSpark speculative decoding ships attached. The released weights bundle the module. Both vLLM and SGLang can enable it.
  • Selectable reasoning effort. The reasoning_effort parameter takes low, high, or max. DeepSeek suggests temperature 1.0 with top_p 0.95 for agentic work, and allowing up to 384K output tokens at the higher settings.

For more models like this one, browse the full open source LLM directory or read the open source LLM ecosystem statistics.

DeepSeek-V4 transformer block: input tokens are embedded, passed through pre-block mixing into CSA/HCA attention and a DeepSeekMoE feed-forward layer, recombined with residual and post-block mixing, then read out by a prediction head (LM loss) and MTP modules (MTP loss).
Source: DeepSeek-V4 technical report

Benchmarks

DeepSeek published eleven agent and coding benchmarks for the 0813 build, and the jump over the April preview is the story. One detail matters for reading these numbers: the code-agent rows were not run on a bare API. DeepSeek evaluated them inside DeepSeek Harness in minimal mode, a two-tool setup limited to bash and a text editor, at max reasoning effort.

DS-V4-Pro-0813 vs frontier models

Higher is better on every benchmark shown. Scores marked n/a were not reported. The Fable-5 column is the w/ fallback configuration. The V4-Flash Preview column from the model card is omitted here for readability.

Opus-4.8 Fable-5 Kimi K3 GLM-5.2 V4-Pro Preview V4-Flash-0731 DS-V4-Pro-0813

Each model keeps the same color across every benchmark, stacked top to bottom in this order. Hover or focus a row to read all six scores.

HLE wo tools
42.7 #4 of 7
HLE w tools
60.0 #2 of 7
View all scores as a table
Benchmark Opus-4.8Fable-5Kimi K3GLM-5.2V4-Pro PreviewV4-Flash-0731DS-V4-Pro-0813
Reasoning
HLE (wo tools) 49.8 53.3 43.5 40.5 37.7 37.8 42.7
HLE (w tools) 57.9 63.0 56.0 54.7 48.2 51.5 60.0
Agentic & Coding
Terminal Bench 2.1 (Acc) 85.0 88.0 88.3 81.0 72.1 82.7 87.9
NL2Repo (Pass@1) 69.7 n/a n/a 48.9 38.5 54.2 61.5
Cybergym (Pass@1) 78.3 83.1 80.0 n/a 52.7 76.7 83.3
DeepSWE (Resolved) 58.0 70.0 67.5 46.2 12.8 54.4 62.7
Toolathlon-Verified (Pass@1) 76.2 77.9 76.5 59.9 55.9 70.3 74.1
Agents' Last Exam (Pass@1) 25.7 n/a 27.6 23.8 16.5 25.2 25.7
AutomationBench Public (Pass@1) 27.2 29.1 30.8 12.9 12.8 25.1 31.8
DSBench-FullStack (Pass@1) 71.6 77.2 73.7 61.8 41.8 68.7 71.1
DSBench-Hard (Pass@1) 71.7 68.3 63.0 54.5 31.1 59.6 67.2

Source: the DS-V4-Pro-0813 model card. Best score in each row is marked.

Third-party evaluations

Independent numbers arrived within a day of the API rollout, and the agent results followed a week later. They broadly confirm the direction DeepSeek reported. The model lands in the top tier of open weights, it is second only to Kimi K3 among open models on real agent tasks, and the cost per task is what puts it on the frontier.

Two Artificial Analysis charts. The top bar chart is Intelligence Index v4.1.1, which combines nine evaluations. Claude Opus 5 (max) leads at 63, Claude Fable 5 (with fallback) 62, GPT-5.6 Sol (max) 61, Grok 4.6 (high) 61, Kimi K3 (max) 60, Qwen3.8 Max 58. DeepSeek V4 Pro 0813 (max) is marked with an arrow at 53, tied with GLM-5.2 (max), ahead of GPT-5.6 Luna (max) at 52 and DeepSeek V4 Flash 0731 (max) at 52, and well ahead of the April DeepSeek V4 Pro (max) at 45. The bottom scatter plot charts Intelligence Index against cost per task on a log scale, with DeepSeek V4 Pro 0813 (max) arrowed at 53 for about $0.25 per task, inside the shaded most attractive quadrant and on the Pareto line.
Artificial Analysis Intelligence Index. On Intelligence Index v4.1.1, DeepSeek V4 Pro 0813 scores 53. That is 8 points above the April V4-Pro build at 45, and it puts the model level with GLM-5.2 and just ahead of GPT-5.6 Luna and V4-Flash-0731, both at 52. It does not lead the open field: Kimi K3 at 60 and Qwen3.8 Max at 58 are open weights and score higher. The second panel is the more interesting one. Plotted against cost per task, the model sits inside the most attractive quadrant at roughly $0.25 per task, on the Pareto line. Claude Opus 5 and Claude Fable 5 score about 10 points higher and cost close to ten times more per task. Note that these cost figures predate the August 16 price increase. Artificial Analysis · Aug 2026
Vals AI Vals Index table filtered to open weights. Kimi K3 first at 74.70% accuracy, DeepSeek V4 Pro 0813 second at 66.25% with a cost per test of $0.14, Qwen 3.8 Max third at 65.47%, GLM 5.2 fourth at 65.02%, DeepSeek V4 Flash 0731 fifth at 63.95%, and DeepSeek V4 seventh at 55.62%.
Vals Index, open weights. In the Vals Index open-weights view, DeepSeek V4 Pro 0813 ranks second at 66.25%. That is more than ten points above the April V4 build at 55.62%, and it moves ahead of Qwen 3.8 Max and GLM 5.2. The cost column separates it further: $0.14 per test against $2.68 for Qwen 3.8 Max and $2.34 for Kimi K3. Three DeepSeek models sit in the top ten. Vals AI · Aug 12, 2026
Arena Code Arena WebDev top 15 bar chart. Claude Opus 5 (Max) leads at 1,691, Kimi K3 (Max) second at 1,674, Qwen-3.8 Max third at 1,669. DeepSeek-V4-Pro (Max) is highlighted at 1,607 with an AutoEval badge, placed between GPT-5.6 Sol (xHigh) at 1,622 and GLM-5.2 (Max) at 1,587.
Code Arena WebDev leaderboard. On the Arena Code Arena WebDev board, DeepSeek-V4-Pro (Max) scores 1,607. It sits between GPT-5.6 Sol (xHigh) at 1,622 and GLM-5.2 (Max) at 1,587, and clears every Claude Opus 4.x entry in the top 15. Arena · Aug 2026
Arena Agent Arena bar chart of the top 15 open models by net improvement against baseline. Kimi K3 (Max) leads at +10.4%. DeepSeek-V4-Pro (High) is highlighted second at +6.3%, ahead of GLM-5.2 (Max) at +5.8% and DeepSeek-V4-Flash (High) at +4.0%. Kimi-K2.7 Code is fifth at +0.4% and the earlier DeepSeek-V4 Pro entry sixth at +0.1%. The remaining nine entries are negative, from GLM-5.1 at 0.0% down to MiniMax-M2.7 at -11.9%.
Agent Arena, top 15 open models. On the Arena Agent Arena board, DeepSeek-V4-Pro (High) ranks second among open models at +6.3% net improvement against baseline. Only Kimi K3 (Max) is ahead at +10.4%. It clears GLM-5.2 (Max) at +5.8% and DeepSeek-V4-Flash (High) at +4.0%. The sharpest comparison is internal: the earlier DeepSeek-V4 Pro entry sits sixth at +0.1%, so the 0813 post-training run is worth more than six points on real agent tasks. Arena · Aug 20, 2026
Arena Agent Arena Pareto frontier scatter plot of net improvement against median cost per task over the last 14 days, on a log scale. A tooltip labels DeepSeek V4 Pro (High) (0813), DeepSeek, MIT, at a score of +6.26% and $0.21 per task, sitting on the green frontier line. Other labelled frontier points are Claude Opus 5 (High) near $2, Kimi K3 (Max), GPT 5.5 (High), GPT 5.5, GLM 5.2 (Max), GPT 5.6 Luna (xHigh), and Mimo V2.5 Pro. A separate DeepSeek V4 Pro point sits below the frontier near $0.05.
Agent Arena Pareto frontier. The same Agent Arena data plotted against median cost per task puts DeepSeek-V4-Pro (High) on the Pareto frontier at +6.26% for $0.21 per task. The frontier climbs steeply to the left: Claude Opus 5 (High) reaches roughly +12% but costs about ten times more per task, and Kimi K3 (Max) buys four more points for around six times the cost. Arena · Aug 20, 2026

API pricing

The price increase DeepSeek warned about previously landed alongside this release. From 16:00 UTC on August 16, 2026, the flat rate is gone and a peak and off-peak structure replaces it. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC. Every other hour is off-peak, at half the peak rate. The table below shows both tiers.

Model Input · cache hit Input · cache miss Output
DeepSeek-V4-Pro-0813 $0.044 $1.32 $3.96
DeepSeek-V4-Pro-0813, off-peak $0.022 $0.66 $1.98
DeepSeek-V4-Flash-0731 $0.014 $0.44 $1.32
DeepSeek-V4-Flash-0731, off-peak $0.007 $0.22 $0.66

Per 1M tokens, as of August 16, 2026. Official pricing.

The jump is steep. V4-Pro output went from $0.87 per million tokens to $3.96 at peak, so even the off-peak rate of $1.98 is more than double the old flat price. Cache-hit input moved from $0.003625 to $0.044 at peak, roughly a twelve-fold increase. V4-Flash took the same treatment, from $0.28 output to $1.32 at peak and $0.66 off-peak.

Even after the increase, the model is inexpensive against the closed frontier. Off-peak output at $1.98 per million tokens is about 92% below Claude Opus 4.8 at $25, and roughly 93% below GPT-5.6 Sol at $30. Arena measured a blended $0.76 per million tokens at the old rates and placed the model on the cost-performance frontier. What changed is the internal comparison: V4-Flash-0731 off-peak output at $0.66 now costs a third of V4-Pro, so the choice between the two models is a real budget decision rather than a rounding error.

Two practical consequences. First, scheduling matters now. A batch job moved out of the two peak windows costs half as much for the same tokens. Second, prompt caching matters more than it did: cache-hit input at $0.022 off-peak is 30 times cheaper than a cache miss, so reusing system prompts, tool schemas, and repository context is where the savings are.

Sources: DeepSeek pricing · DeepSeek V4-Pro release notes · Claude pricing · OpenAI pricing

What people are saying

“We ran DeepSeek v4 Pro 0813 on our cybersecurity benchmark, it outperformed EVERY (!) other model at finding vulnerabilities - At pass@3, it rediscovered 87.5% of the benchmark CVEs. Far above Opus 5 and Qwen 3.8 at 81.3% - The tradeoff is precision.”

Philippe Dourassov

@pilvar222

X Aug 13, 2026

“DeepSeek V4 Pro 0813 delivers the same high-end performance as Pro Preview, at 29% lower cost. we benchmarked Pro 0813, Pro Preview, and Flash 0731 across 100 deep-research questions. 0813 was strongest on: Law: 100% Academic: 83% Medicine: 67%”

GMI Cloud

@gmi_cloud

X Aug 13, 2026

“DeepSeek V4 Pro 0813 is live on OpenRouter. @deepseek_ai reports large agent gains over V4 Pro Preview: DeepSWE 62.7 (+49.9), CyberGym 83.3 (+30.6), NL2Repo 61.5 (+23.0), and Terminal Bench 2.1 87.9 (+15.8) More providers coming online soon”

OpenRouter

@OpenRouter

X Aug 12, 2026

“Deepseek V4 Pro GA scoring only 1 point higher than Flash on Artificial Analysis while being like 5 times larger is crazy work”

Lentils

@Lentils80

X Aug 13, 2026

“We ran DeepSeek v4 Pro 0813 on our cybersecurity benchmark, it outperformed EVERY (!) other model at finding vulnerabilities - At pass@3, it rediscovered 87.5% of the benchmark CVEs. Far above Opus 5 and Qwen 3.8 at 81.3% - The tradeoff is precision.”

Philippe Dourassov

@pilvar222

X Aug 13, 2026

“DeepSeek V4 Pro 0813 delivers the same high-end performance as Pro Preview, at 29% lower cost. we benchmarked Pro 0813, Pro Preview, and Flash 0731 across 100 deep-research questions. 0813 was strongest on: Law: 100% Academic: 83% Medicine: 67%”

GMI Cloud

@gmi_cloud

X Aug 13, 2026

“DeepSeek V4 Pro 0813 is live on OpenRouter. @deepseek_ai reports large agent gains over V4 Pro Preview: DeepSWE 62.7 (+49.9), CyberGym 83.3 (+30.6), NL2Repo 61.5 (+23.0), and Terminal Bench 2.1 87.9 (+15.8) More providers coming online soon”

OpenRouter

@OpenRouter

X Aug 12, 2026

“Deepseek V4 Pro GA scoring only 1 point higher than Flash on Artificial Analysis while being like 5 times larger is crazy work”

Lentils

@Lentils80

X Aug 13, 2026

Run it yourself

vLLM Recipe → 
vllm serve deepseek-ai/DeepSeek-V4-Pro-0813
SGLang Cookbook → 
sglang serve --model-path deepseek-ai/DeepSeek-V4-Pro-0813

Frequently asked questions

Is DeepSeek-V4-Pro-0813 open source?

Yes. The weights ship under the MIT license on Hugging Face, which allows unrestricted commercial use, modification, and redistribution. The API went live on August 12, 2026 and the weights followed shortly after.

What is DeepSeek Harness?

DeepSeek Harness, or dsh, is an open-source agent harness from DeepSeek, released under the MIT license alongside DeepSeek-V4-Pro-0813. It turns a language model into a coding agent, and it is built so that every capability is a plugin, including models, tools, skills, sessions, sandboxes, and storage. It ships four modes: standard for full coding work, code for letting the model orchestrate multiple rounds of tool calls, minimal for benchmarking with just bash and a text editor, and creator for building custom presets. Every run is written to an append-only session log that you can resume, fork, search, and replay. Version 0.1 is a developer preview, and the maintainers warn of compatibility-breaking changes.

How much does DeepSeek-V4-Pro-0813 cost?

From 16:00 UTC on August 16, 2026 the API uses peak and off-peak rates. Peak is $0.044 per million cache-hit input tokens, $1.32 per million cache-miss input tokens, and $3.96 per million output tokens. Off-peak is half of each: $0.022, $0.66, and $1.98. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, and every other hour is off-peak. The previous flat rate was $0.87 per million output tokens.

Is DeepSeek-V4-Pro-0813 better than DeepSeek-V4-Flash-0731?

On the eleven benchmarks DeepSeek published, yes, on all of them. Terminal Bench 2.1 is 87.9 against 82.7, DeepSWE 62.7 against 54.4, and Cybergym 83.3 against 76.7. The gaps are meaningful but not huge, and V4-Pro now costs three times more per output token, so Flash remains the better default for high-volume work.

How does DeepSeek-V4-Pro-0813 compare to Claude Opus 4.8 and Fable 5?

It trades wins. DeepSeek reports 87.9 on Terminal Bench 2.1 against 85.0 for Opus 4.8 and 88.0 for Fable 5, and 83.3 on Cybergym against 78.3 and 83.1. Opus 4.8 keeps a clear lead on NL2Repo, 69.7 against 61.5, and on DSBench-Hard, 71.7 against 67.2. Fable 5 leads on DeepSWE, 70.0 against 62.7. All of these are vendor-reported numbers from the DeepSeek model card.

Keep exploring

vllm.ai

DeepSeek V4 in vLLM: Efficient Long-context Attention

lmsys.org

DeepSeek-V4 on Day 0: From Fast Inference to Verified RL with SGLang and Miles

developer.nvidia.com

Build with DeepSeek V4 Using NVIDIA Blackwell and GPU-Accelerated Endpoints

inferencex.semianalysis.com

MI355X DeepSeek-V4-Pro on SGLang: 110.5x Throughput per GPU in 26 Days

From the blog

Is DeepSeek Better Than ChatGPT? 25+ DeepSeek Stats (2026)

From the blog

The Best Open Source LLMs in 2026