China's DeepSeek upgrades its V4 Pro model

Chinese artificial intelligence startup DeepSeek released on late Wednesday the official API version of its V4 Pro model, delivering a major upgrade in AI agent capabilities, with benchmark results showing the model closing in on — and in some tests surpassing — Claude Fable 5.

The new V4 Pro retains the same model name for API calls and now supports the Responses API as well as integration with Codex, expanding the ways developers can deploy the model for agent-based applications and coding tasks.

DeepSeek priced V4 Pro at 3 yuan ($0.42) per million input tokens and 6 yuan per million output tokens. Cached input costs just 0.025 yuan per million tokens, maintaining the company's strategy of aggressively lowering the cost of deploying advanced AI models.

The launch marks DeepSeek's latest attempt to challenge leading global AI developers not only on model intelligence but also on the ability of AI systems to independently use tools, write and execute code and complete multi-step tasks.

Benchmark results showed V4 Pro scored 87.9 on Terminal-Bench, a test designed to measure how well AI agents perform tasks in terminal environments, narrowly trailing Fable5's score of 88.

DeepSeek's model surpassed Fable5 on CyberGym, a benchmark for evaluating AI agents on cybersecurity-related tasks, and AutomationBench, which tests agents on difficult automation tasks, according to the benchmark results.