Overview
- DeepSeek disclosed Thursday that it had launched DeepSeek‑V4‑Pro‑0813, an official build that the company says adds agent capabilities and supports a 1 million‑token context window using a mixture‑of‑experts architecture with open weights.
- The company published Harness v0.1 as an MIT‑licensed developer preview, a runtime framework designed to turn its models into agentic coding tools that can run multi‑step workflows and call external tools.
- DeepSeek announced a peak/off‑peak pricing plan effective Aug. 16 that raises rates sharply for V4‑Pro and V4‑Flash, with V4‑Pro output rising from $0.87 to $3.96 per million tokens at peak and off‑peak set at half the peak rate.
- Independent tests and indexes show mixed results for the new Pro build, with the model trailing several frontier Western systems on broad reasoning benchmarks but outperforming in niche tasks such as cybersecurity and certain coding benchmarks.
- To handle heavy token volumes and greater commercial demand DeepSeek is accelerating hiring, restarting a large fundraising round, and investing in data‑centres and in‑house chip design, a push that will shape developer costs and enterprise decisions about where to run sensitive workloads.