DeepSeek V4 Pro 0813 Release: Performance, Pricing, and Benchmarks

DeepSeek V4 Pro 0813 is a high-performance model designed for complex reasoning and scientific tasks, providing a cost-effective alternative to top-tier frontier models. While it excels in graduate-level scientific reasoning and specialized agentic tasks, community feedback suggests a disparity between benchmark performance and real-world coding reliability.

Technical Performance and Benchmarks

DeepSeek V4 Pro 0813 demonstrates strong capabilities in high-complexity reasoning, particularly in scientific and conversational agent scenarios. According to data from Artificial Analysis, the model achieves the following scores:

  • Scientific Reasoning: 88.8% on GPQA Diamond (graduate-level scientific reasoning).
  • Conversational Agents: 96.2% on $\tau^2$-Bench Telecom for dual-control scenarios.
  • Instruction Following: 76.5% on IFBench.
  • Long Context Reasoning: 70.0% on AA-LCR.
  • Coding: 50.0% on SciCode (scientific computing) and 46.2% on Terminal-Bench Hard.

However, the model shows significant weaknesses in research-level physics reasoning (12.9% on CritPt) and a low non-hallucination rate (5.9% on AA-Omniscience), suggesting a tendency toward speculative responses in certain knowledge domains.

Pricing and API Economics

DeepSeek V4 Pro 0813 is positioned as a highly aggressive value proposition, often described by users as being approximately 20x cheaper than competitors like Opus 4.8 while remaining competitive in performance.

OpenRouter Pricing

On OpenRouter, the model features the following weighted average costs:

  • Input Price: $0.03942 / M tokens
  • Output Price: $0.8697 / M tokens

Official DeepSeek Pricing Structure

DeepSeek is implementing a peak/off-peak pricing model effective August 16, 2026, to optimize load and cost:

Model Period Cache Hit (Input/1M) Cache Miss (Input/1M) Output (/1M)
V4-Flash Off-peak $0.007 $0.22 $0.66
V4-Flash Peak $0.014 $0.44 $1.32
V4-Pro Off-peak $0.022 $0.66 $1.98
V4-Pro Peak $0.044 $1.32 $3.96

Peak hours are defined as 01:00–04:00 UTC and 06:00–10:00 UTC.

User Insights and Real-World Application

Community feedback reveals a divide between the model's benchmark success and its practical utility in software development.

Coding and Reliability

Several developers report that the model struggles with complex, multi-file repository tasks compared to other frontier models. One user noted that while the model is capable for simple projects, it introduces issues in complex deployment scenarios (e.g., Docker-compose with Caddy) where other models like GPT-5.6-Terra performed flawlessly.

Another user highlighted a reliability issue with "pass@1" performance:

"Deepseek V4 Pro 0813 is the most unreliable model I have tried... it's horrendous at pass@1 very prone to going wrong and doing horribly at most benches."

Comparison with V4 Flash

There is a recurring sentiment that the V4 Flash 0731 model provides better value and more consistent results for general development tasks. Users have described Flash as a "massive jump in capability" that often suffices for prototyping and interactive architecture work, making the upgrade to Pro feel marginal or even disappointing for some workflows.

Specialized Use Cases

Despite coding criticisms, the model is praised for research and financial analysis. One user reported that the model goes "head-to-head with the most expensive models" for stock market and forex searches, and another successfully used it to find significant gains in a distributed physics engine simulator.

Operational Metrics

For developers integrating the model via OpenRouter, the current operational baseline is as follows:

  • Throughput: 58 tokens per second (P50).
  • Latency: 1.24 seconds (P50).
  • Uptime: 100% over the last 3 days.
  • Privacy Note: Some users have noted that current available endpoints may require enabling "Allow paid endpoints that train on request data" in privacy settings.

Sources

Related