DeepSeek launches V4-Pro, rivaling top AI agents at a fraction of the cost

Chinese AI company DeepSeek released the official version of its flagship V4 Pro model late Wednesday, delivering agent capabilities that approach Anthropic's Claude Fable 5 on multiple benchmarks at a fraction of the cost — intensifying competition in the fast-moving market for AI systems that can autonomously complete complex tasks.

A Stealth Midnight Launch

The model, internally versioned as DeepSeek-V4-Pro-0813, went live across web, mobile, and API without advance notice on the night of August 12. Developers calling the existing `deepseek-v4-pro` endpoint automatically receive the upgraded model with no code changes required. The release features a 1 million-token context window and a maximum output length of 384,000 tokens.globaltimes+3

According to DeepSeek's own benchmarks, the model scored 87.9 on Terminal-Bench, narrowly trailing Fable 5's score of 88, and 83.3 on the cybersecurity benchmark Cybergym, edging past Fable 5's 83.1. On the reasoning benchmark HLE with tools enabled, it reached 60.0 versus Fable 5's 63.0. Third-party assessments were more conservative: benchmarking service Artificial Analysis rated V4 Pro at 53 on its Intelligence Index, four points behind OpenAI's GPT-5.6 Terra and seven behind Moonshot AI's Kimi K3.trendingtopics+2

Price Gap Draws Industry Attention

The pricing has become the model's defining feature. DeepSeek set the cost at 3 yuan (roughly $0.44) per million input tokens and 6 yuan per million output tokens, with cached inputs priced at just 0.025 yuan per million tokens. The 120-fold gap between cached and uncached input pricing means that effective caching architecture can have a larger impact on costs than model selection itself.weex+1

By comparison, Fable 5 costs approximately $50 per million output tokens, making DeepSeek's offering roughly 46 times cheaper on a blended basis. Research firm SemiAnalysis posted on X congratulating DeepSeek, noting the model "massively beats Nemotron 3 Ultra on agentic tasks".globaltimes+1

However, the low pricing may be temporary. On August 6, DeepSeek announced plans for a "relatively significant" API price increase, and beginning August 16 a peak/off-peak pricing structure will take effect, doubling rates during business hours in China.kucoin+1

A Broader Competitive Shift

The release arrived alongside xAI's same-day launch of Grok 4.6, priced at $2 per million input tokens and $6 per million output tokens. The simultaneous moves underscore an industry-wide pivot from conversational AI toward autonomous agents capable of tool use, code execution, and multi-step task completion.longbridge+1

DeepSeek is also building an agent execution framework called DeepSeek Harness, with job postings and a dedicated WeChat account now public, according to Bloomberg. The Business Times of Singapore reported that the company is recruiting for a team focused on building products to compete directly with Anthropic's Claude Code.businesstimes

Tian Feng, former dean of SenseTime's Intelligence Industry Research Institute, told the Global Times that "the next competitive edge may depend less on who answers better in chat, and more on who can complete complex real-world tasks at lower cost and with higher reliability".globaltimes