DeepSeek V4 Pro 0813
TL;DR Highlight
DeepSeek V4 Pro 0813 is GA with 1M context at $0.2016 in / $8 out on OpenRouter
Who Should Read
A backend developer choosing a low-cost LLM for coding agents and torn between DeepSeek Pro and Flash
Core Mechanics
- V4 Pro 0813 is the GA release
- It is the official release of DeepSeek's large mixture-of-experts model.
- Released Aug 12, 2026, with a 1,048,576-token context and up to 393,216 completion tokens.
- OpenRouter pricing: cheap input, pricey output
- Listed at $0.2016/M input, $8.00/M output, and $0.20/M cache read.
- 21 providers serve it, with different prices and speeds.
- Function calling and structured outputs work
- tools and tool_choice enable function calling.
- A JSON schema in response_format gives structured outputs.
- Artificial Analysis scores lean toward coding
- On Max setting: Coding Index 68.8, Intelligence Index 36.0, Agentic Index 41.3.
- GPQA Diamond 92.8%, Terminal-Bench 2.1 78.7%, HLE 41.0%.
- Hallucination-related scores are low
- AA-Omniscience Accuracy is 49.1%.
Community Reactions
- Many preferred Flash 0731
- One commenter was disappointed because Flash 0731 was such a big jump, and plans to stay on Flash.
- Another found the latest Flash a big step up and Pro feels slow by comparison.
- Gaps showed up on complex tasks
- On a docker-compose generation task via OpenRouter, one user said this model had a few issues while gpt-5.6-terra-high had none.
- In Codex CLI on the same feature, DeepSeek took 12m 02s at $0.12 with a bug, and Grok 4.6 took 3m 18s at $1.41 with no bug, per one commenter.
- Cheap enough to be useful for some
- One user ran it all day for about $12.50 (2B tokens, 50% cache hits) and got meaningful gains on a physics engine.
- Another said it matches or beats earlier models at lower cost.
- Privacy and price changes were concerns
- One commenter noted the only endpoint then required allowing endpoints that train on request data, and would wait before benchmarking.
- A commenter shared that DeepSeek's official API moves to peak/off-peak pricing on Aug 16 16:00 UTC, with off-peak at 50%.
How to Apply
- Compare Pro and Flash on OpenRouter
- Pick one real repo task today and run the same prompt on both by only changing the model slug.
- Record time, cost, and bug count in a table.
- Move batch jobs to off-peak hours
- On the official API, peak is 01:00–04:00 and 06:00–10:00 UTC; all else is off-peak.
- Shift nightly batches and bulk evals to off-peak.
- Check privacy settings and providers
- Before sending sensitive code, check OpenRouter's setting for endpoints that train on data.
- Pin or exclude providers with provider routing.
Try It
# Call DeepSeek V4 Pro 0813 via the OpenRouter API (paste in terminal)
pip install openai
export OPENROUTER_API_KEY={{key}}
python - <<'EOF'
import os
from openai import OpenAI
c = OpenAI(base_url="https://openrouter.ai/api/v1", api_key=os.environ["OPENROUTER_API_KEY"])
for m in ["deepseek/deepseek-v4-pro-0813"]:
r = c.chat.completions.create(model=m, messages=[{"role":"user","content":"{{task}}"}])
print(m, r.usage)
print(r.choices[0].message.content)
EOF
# Next: add a Flash slug to the model list and compareTerminology
Related Papers
Mercury 3: Diffusion at scale
Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality
Timestral: Mistral's time-series foundation models
A 209M time-series model ranks first on three benchmarks and runs on CPU.
DeepSeek V4.1 Flash 1017 (Max)
DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.
If you don't run your own LLM proxy, how do you know whats going on?
Running your own LLM proxy shows which keys, models, and cache misses drive your usage.
DeepSeek: DeepSeek v4.1 Flash 1017
DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents
DeepSeek-v4.1-Flash-1017
DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length
Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware