DeepSeek Releases DeepSeek V4 Pro (Version 0813) with Significant Performance Gains While Keeping Ultra-Low API Pricing UnchangedLeading Chinese AI research lab DeepSeek has officially rolled out the production build of its flagship model, DeepSeek V4 Pro (version 0813), as previously scheduled. Benchmark evaluation results show noticeable performance improvements across complex coding, terminal execution, and cybersecurity tasks.
The highlight of the release is the company's pricing strategy: despite previous warnings that API rates would increase upon the release of the final build, DeepSeek has decided to hold its ultra-competitive API pricing steady.
Benchmark Performance Comparison
In benchmark evaluations, DeepSeek V4 Pro (0813) demonstrated strong competitive capabilities against rival frontier models:
Terminal Bench 3.1: Reached a score of 87.9, trailing closely behind top-tier competitors Kimi K3 and Fable 5.
DeepSWE: Scored 62.7, placing it behind both Kimi K3 and Fable 5 in real-world software engineering benchmarks.
Cybergym: Narrowly outperformed both rival models, showcasing exceptional proficiency in offensive and defensive cybersecurity challenges.
Disruptive Pricing Advantage
DeepSeek V4 Pro’s API rates remain fixed at $0.435 per million input tokens and $0.87 per million output tokens. This leaves DeepSeek as one of the most cost-effective top-tier models on the market by a massive margin:
DeepSeek V4 Pro (0813): $0.435 / $0.87 per 1M tokens (input/output)
Kimi K3: $3.00 / $15.00 per 1M tokens
Fable 5: $10.00 / $50.00 per 1M tokens
While the API is currently live and accessible to developers, official announcements and technical documentation have so far been published primarily through DeepSeek’s domestic WeChat channels in China, with international web portals awaiting updated public listings.
The way DeepSeek achieves such low API pricing, while Western labs like Anthropic or OpenAI (and regional competitors like Fable) price their solutions to cover high R&D costs and hardware investments, DeepSeek leverages architectural innovations such as Multi-head Latent Attention (MLA) and Mixture-of-Experts (MoE) optimizations to reduce active parameter memory usage during inference. Passing these operational hardware savings directly to developers allows them to lower prices below the market and attract massive global developer interest.
Winning at Cybergym while scoring similarly in Terminal Bench 3.1 demonstrates that DeepSeek V4 Pro is highly optimized for command-line navigation, system administration, and vulnerability assessment. For enterprise Dev-Ops and cybersecurity teams performing thousands of automated terminal calls daily, the low token cost coupled with high command-line accuracy makes this model an attractive backend engine.
Source: @ChristGPT
DeepSeek Releases DeepSeek V4 Pro (Version 0813) with Significant Performance Gains While Keeping Ultra-Low API Pricing UnchangedLeading Chinese AI research lab DeepSeek has officially rolled out the production build of its flagship model, DeepSeek V4 Pro (version 0813), as previously scheduled. Benchmark evaluation results show noticeable performance improvements across complex coding, terminal execution, and cybersecurity tasks.
The highlight of the release is the company's pricing strategy: despite previous warnings that API rates would increase upon the release of the final build, DeepSeek has decided to hold its ultra-competitive API pricing steady.
Benchmark Performance Comparison
In benchmark evaluations, DeepSeek V4 Pro (0813) demonstrated strong competitive capabilities against rival frontier models:
Terminal Bench 3.1: Reached a score of 87.9, trailing closely behind top-tier competitors Kimi K3 and Fable 5.
DeepSWE: Scored 62.7, placing it behind both Kimi K3 and Fable 5 in real-world software engineering benchmarks.
Cybergym: Narrowly outperformed both rival models, showcasing exceptional proficiency in offensive and defensive cybersecurity challenges.
Disruptive Pricing Advantage
DeepSeek V4 Pro’s API rates remain fixed at $0.435 per million input tokens and $0.87 per million output tokens. This leaves DeepSeek as one of the most cost-effective top-tier models on the market by a massive margin:
DeepSeek V4 Pro (0813): $0.435 / $0.87 per 1M tokens (input/output)
Kimi K3: $3.00 / $15.00 per 1M tokens
Fable 5: $10.00 / $50.00 per 1M tokens
While the API is currently live and accessible to developers, official announcements and technical documentation have so far been published primarily through DeepSeek’s domestic WeChat channels in China, with international web portals awaiting updated public listings.
The way DeepSeek achieves such low API pricing, while Western labs like Anthropic or OpenAI (and regional competitors like Fable) price their solutions to cover high R&D costs and hardware investments, DeepSeek leverages architectural innovations such as Multi-head Latent Attention (MLA) and Mixture-of-Experts (MoE) optimizations to reduce active parameter memory usage during inference. Passing these operational hardware savings directly to developers allows them to lower prices below the market and attract massive global developer interest.
Winning at Cybergym while scoring similarly in Terminal Bench 3.1 demonstrates that DeepSeek V4 Pro is highly optimized for command-line navigation, system administration, and vulnerability assessment. For enterprise Dev-Ops and cybersecurity teams performing thousands of automated terminal calls daily, the low token cost coupled with high command-line accuracy makes this model an attractive backend engine.
Source: @ChristGPT
Comments
Post a Comment