Particle.news

DeepSeek Releases V4‑Pro‑0813 and Open‑Sources Agent Runtime to Target Anthropic’s Claude Code

DeepSeek aims to attract developers with very low per‑token prices, MIT‑licensed model weights and an open agent runtime.

Overview

  • DeepSeek on Thursday released DeepSeek‑V4‑Pro‑0813 across its API, app and web and published the model weights under an MIT license on Hugging Face.
  • The company also published Harness v0.1 as an open‑source developer preview, a MIT‑licensed agent runtime designed to turn V4 models into autonomous coding agents and backed by a newly formed ‘Harness’ team.
  • DeepSeek listed API prices of about $0.435 per 1 million input tokens and $0.87 per 1 million output tokens and said it will raise prices for V4 Pro and V4 Flash and introduce peak/off‑peak rates to manage heavy demand.
  • Independent benchmarks show mixed results: V4‑Pro‑0813 trails some frontier models on broad reasoning indices but performs strongly in niche tests such as cybersecurity, while DeepSeek’s comparisons emphasize only small gaps to Claude at a tiny fraction of the price.
  • Rapid token consumption has strained compute capacity, prompting aggressive hiring, plans for onshore data centres and chip design, and a large fundraising push that seeks capital to finance scaled infrastructure and wider commercial adoption amid data‑sovereignty limits.