DeepSeek V4 Pro launches with 1.6T parameters at $1.74/M tokens, undercutting Claude Sonnet 4.6 by 42%

DeepSeek released two preview models in its V4 series: DeepSeek-V4-Pro and DeepSeek-V4-Flash. Both are Mixture of Experts models with 1 million token context windows, released under the MIT license.

Model specifications

DeepSeek-V4-Pro has 1.6 trillion total parameters with 49 billion active parameters. DeepSeek-V4-Flash has 284 billion total parameters with 13 billion active. According to DeepSeek, this makes V4 Pro the largest open weights model available, exceeding Kimi K2.6 (1.1T) and GLM-5.1 (754B), and more than double the size of DeepSeek V3.2 (685B).

The Pro model weighs 865GB on Hugging Face. Flash weighs 160GB.

Pricing comparison

DeepSeek's pricing significantly undercuts existing frontier models:

DeepSeek V4 Flash: $0.14/M input tokens, $0.28/M output tokens DeepSeek V4 Pro: $1.74/M input tokens, $3.48/M output tokens

For comparison:

• GPT-5.4: $2.50/$15 per million tokens
• Claude Sonnet 4.6: $3/$15 per million tokens
• Gemini 3.1 Pro: $2/$12 per million tokens
• Claude Haiku 4.5: $1/$5 per million tokens
• GPT-5.4 Nano: $0.20/$1.25 per million tokens

V4 Flash is the cheapest small model available. V4 Pro costs 42% less than Claude Sonnet 4.6 for input tokens and 77% less for output tokens.

Efficiency improvements

DeepSeek attributes the low pricing to substantial efficiency gains. According to their technical paper, in 1M-token context scenarios, V4 Pro achieves only 27% of the single-token FLOPs and 10% of the KV cache size compared to DeepSeek V3.2. V4 Flash achieves 10% of the FLOPs and 7% of the KV cache size.

Performance benchmarks

DeepSeek claims V4 Pro is competitive with frontier models, with one caveat. According to the company's paper: "Through the expansion of reasoning tokens, DeepSeek-V4-Pro-Max demonstrates superior performance relative to GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks. Nevertheless, its performance falls marginally short of GPT-5.4 and Gemini-3.1-Pro, suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately 3 to 6 months."

No independent benchmarks are yet available.

Availability

Both models are available via OpenRouter and downloadable from Hugging Face. The models can be accessed through the standard llm-openrouter plugin.

What this means

DeepSeek's aggressive pricing creates immediate pressure on OpenAI, Anthropic, and Google to justify their premium pricing—or match it. At $1.74 per million input tokens, V4 Pro costs less than half of Claude Sonnet 4.6 while claiming near-frontier performance. If the performance claims hold under independent testing, this represents a fundamental shift in the economics of frontier model deployment. The 90% reduction in KV cache size at 1M tokens also suggests meaningful architectural innovations beyond simple scaling, particularly for long-context applications where memory constraints have been a limiting factor.