DeepSeek-V4-Pro-0813Open Weight

Released

DeepSeek-V4-Pro-0813 is an AI model released by DeepSeek on Thursday, Aug 13 2026, 13 days after DeepSeek-V4-Flash-0731. It is an open-weight model — the trained weights are available to download and run. Benchmark results (shown below) cover BullshitBench v2, SWE-Bench Verified, Terminal-Bench 2.1, Toolathlon-Verified, BrowseComp, CyberGym, and 2 more.

API pricing

Off-peak (all hours outside the published peak windows)
Input
$0.66
Cached input
$0.022
Output
$1.98
Peak (01:00–04:00 and 06:00–10:00 UTC)
Input
$1.32
Cached input
$0.044
Output
$3.96
  • Thinking and non-thinking modes use the same published price.
USD per 1M tokens

Available from

Baidu—$0.2442$0.73261Mfp8
StreamLake—$0.264$0.7921M—
Ionstream—$0.2528$1.95841M—
DeepSeek—$0.66$1.981M—
DeepInfra—$1.30$2.601Mfp8
Phala—$0.957$2.87761M—
Novita—$0.99$2.971Mfp8
GMICloud—$1.056$3.1681Mfp8
NextBit—$1.056$3.1681Mfp8
Alibaba—$1.122$3.3661M—
Wafer—$0.245$3.501M—
CoreWeave—$1.31$3.961Mfp8
AtlasCloud—$1.32$3.961Mfp8
Cloudflare—$1.32$3.961M—
DigitalOcean—$1.32$3.961M—
Parasail—$1.32$3.961Mfp8
SiliconFlow—$1.32$3.961Mfp8
Together—$1.32$3.961M—
Sail Research—$0.40$4.301Mfp4
Sail Researchus$0.40$4.301Mfp4
Venice—$1.65$4.951M—
USD per 1M tokens

Benchmarks

Coding

SWE-Bench Verifiedvia BenchLM
80.6%
#10 of 57

Terminal & CLI

Terminal-Bench 2.1
87.9%
#8 of 30

Agentic & tool use

Toolathlon-Verified
74.1%
#3 of 9
BrowseCompvia BenchLM
83.4%
#15 of 28
CyberGym
83.3%
#5 of 13
AutomationBench
31.8%
#9 of 13

Reasoning & science

Humanity's Last Exam
42.7%no tools
60%with tools
no tools
#8 of 22
with tools
#9 of 44

Robustness

BullshitBench v2
35%
#44 of 79

Source: BenchLM, retrieved 31 August 2026. Other sources are identified on the linked benchmark pages.

Compare DeepSeek-V4-Pro-0813 with

DeepSeek-V4-Pro-0813

About

DeepSeek-V4-Pro-0813, released August 13, 2026, brought the top tier of the V4 line up to the agentic post-training that had landed on Flash two weeks earlier. On DeepSeek's own harness it scored 87.9 on Terminal-Bench 2.1, against 82.7 for V4-Flash-0731 and 72.1 for the April V4-Pro-Preview build, and 83.3 on CyberGym — the highest figure in the lab's launch comparison at the time, marginally ahead of Claude Fable 5. It also posted 42.7% on Humanity's Last Exam without tools and 60.0% with them, 74.1 on Toolathlon-Verified and 31.8 on the public AutomationBench split. As with the July Flash release, DeepSeek ran the code-agent evaluations through its own unreleased harness in minimal mode, and its numbers for rival models did not always match those labs' published figures, so the chart reads best as a within-family comparison.

The headline changes were operational rather than architectural. V4-Pro and V4-Flash both gained a selectable reasoning effort — low for simple prompts, high for everyday agent work, max for hard tasks — replacing the single fixed thinking budget of the earlier builds, and the release added native OpenAI Responses API support with one-click Codex setup. Distribution followed the pattern the Flash refresh established: the checkpoint reached the app and web product under an "Expert Mode" toggle and shipped through the API under unchanged model names, so existing callers were moved onto it without a code change. Weights for this checkpoint are available under the MIT License on Hugging Face.

Frequently asked questions

DeepSeek-V4-Pro-0813 was released by DeepSeek on Thursday, Aug 13 2026.

All DeepSeek releases

23 tracked