DeepSeek V4 Pro Stable Release Quietly Goes Live: Agent Performance Rivals Claude Fable 5, API Pricing Unchanged for Now

DeepSeek quietly released the stable version of V4 Pro late on August 12, 2026, Beijing time, without any prior announcement. There were no WeChat articles, technical blogs, or social media teasers — the only sign was a change on the DeepSeek API pricing page, where the model version shifted from preview to DeepSeek-V4-Pro-0813.

This release marks the official graduation of the flagship model that had been running as a preview since April 2026. The preview version had been downloaded over 1.4 million times on Hugging Face, while the stable version surpassed 174,000 downloads in its first week.

Model Size and Core Specifications

DeepSeek V4 Pro is a Mixture-of-Experts (MoE) model with 1.6 trillion total parameters, activating approximately 49 billion parameters per token. It features a context window of up to 1 million tokens and a maximum output length of 384,000 tokens, placing it in the top tier of global AI models.

Notably, V4 Pro employs a hybrid attention design that combines compressed sparse attention with new high-compression attention heads for low-cost long-context prefilling.

Key Upgrade: Significantly Enhanced Agent Capabilities

The main focus of this update is improving agent performance. DeepSeek released a benchmark comparison chart in its official group chat, showing that the new model's agent capabilities are close to, and in some areas surpass, Anthropic's Claude Fable 5.

According to benchmark data cited by media, V4 Pro stable exceeded Fable 5 in the following two tests:

• AutomationBench: Tests whether an AI agent can truly execute tasks
• CyberGym: Evaluates the model's offensive and defensive capabilities in cybersecurity

Additionally, V4 Pro stable now supports Responses API and Codex integration, further enhancing its utility as a development tool.

API Pricing and Cost Advantage

Currently, DeepSeek V4 Pro's API pricing has not been raised, maintaining the reduced standard rates. Specific pricing is as follows:

Billing Item	Price (yuan per million tokens)
Input (cache hit)	0.025
Input (cache miss)	3
Output	6

For comparison, its lightweight counterpart V4 Flash charges 0.02 yuan, 1 yuan, and 2 yuan respectively, meaning V4 Pro's output and cache-miss input costs are three times those of V4 Flash.

DeepSeek's pricing strategy is extremely competitive. According to reports, V4 Pro's performance is close to Fable 5, but its price is only about 2% of the latter. However, the DeepSeek website has already indicated a significant overall API price increase in the near future.

Benchmark Data Notes

It should be noted that the benchmark data for V4 Pro stable is currently derived from a comparison chart released in DeepSeek's official group chat and has not yet been independently verified by third parties on a large scale. Therefore, conclusions about "close to/surpassing" other models should be taken with caution.

DeepSeek has not yet published any official changelog or technical blog detailing the specific improvements and performance data of V4 Pro stable.

Summary

The quiet launch of DeepSeek V4 Pro stable showcases significant progress in agent capabilities while maintaining a highly competitive pricing structure. Although official details are limited, the published data suggests V4 Pro has become a strong contender in the top tier of global AI models. For developers and enterprise users, this means a powerful yet cost-effective new option.