DeepSeek ships V4 Pro, its finished flagship — closing in on Claude Fable 5 at a fraction of the price

The Chinese AI firm ends a four-month preview with V4-Pro-0813, posting agent benchmarks that nearly match Anthropic's flagship at roughly 46 times lower cost.

Chinese AI firm DeepSeek has quietly shipped the finished version of its flagship model, V4 Pro, ending a preview period that has run since April.

The tell was a single table cell: the version listed for "deepseek-v4-pro" on DeepSeek's API page now reads V4-Pro-0813. The model now supports the Responses API and tool calls, with a 1-million-token context window and a maximum output of 384,000 tokens, TechNode reports. DeepSeek confirmed the rollout on its WeChat account, pointing to "significantly enhanced agent capabilities" and Codex integration, with access through the API, mobile apps, and web via an "Expert Mode," according to the Global Times.

The headline number is the benchmark gap. On Terminal-Bench, a test of how well AI agents operate in a terminal, V4 Pro 0813 scored 87.9 — just shy of Anthropic's Claude Fable 5 at 88. Across nine agent benchmarks where both are scored, DeepSeek's own table puts Fable 5 ahead by an average of 5.3%, with DeepSeek winning outright on two, Decrypt reported.

Where the comparison stops being close is price. Fable 5 runs $10 per million input tokens and $50 per million output; V4 Pro runs $0.435 and $0.87, with cached input at just $0.003625 — roughly 46 times cheaper to run on blended rates, per Decrypt's math. Research firm SemiAnalysis congratulated DeepSeek, posting that V4-Pro-0813 "massively beats Nemotron 3 Ultra on agentic tasks."

The launch landed within a day of xAI's Grok 4.6 and just ahead of OpenAI's GPT-5.6 Sol preview — an industry pivot away from chatbots and toward long-running agents that use tools, write code, and finish multi-step jobs on their own.