DeepSeek Releases V4 Pro, Claims 5% Gap to Claude Fable at 1/46th the Cost

What happened: Chinese AI lab DeepSeek released DeepSeek-V4-Pro-0813, its latest large language model, with pricing at $0.435 per million input tokens and $0.87 per million output tokens—roughly 1/46th the price of Anthropic’s Claude Fable 5. DeepSeek’s own benchmarks show V4 Pro trailing Fable 5 by an average of 5.3% on nine agent tasks, while outperforming on certain cybersecurity benchmarks. However, independent tests highlight that while V4 Pro excels at vulnerability discovery (87.5% pass@3), its precision is only 65.6%, indicating a high false positive rate.

Why it matters: The dramatic price-performance gap could pressure incumbents in the agentic AI space, though DeepSeek’s results are largely self-reported and independent verification is still emerging. The model’s strengths in cybersecurity and agent tasks may attract enterprise users, but general-purpose performance and precision limitations remain. The release exemplifies the rapid pace and cost compression in the global LLM race, with China-based labs increasingly competitive on both price and specialized benchmarks.

Source: Decrypt