DeepSeek, a China-based AI model developer, may release its V4 model family within days after a limited rollout of V4 Flash and V4 Pro.

In an email to API users, the company said the full V4 version would launch in mid-July.

Its model card says V4 succeeds V3 and uses a mixture-of-experts design and a 1 million-token context window.

DeepSeek said V4 pricing would shift to peak and off-peak billing, with V4 Pro output priced at US$0.87 per million tokens in normal periods and US$1.74 in peak periods.

V4 Flash output starts at US$0.28 per million tokens. Pricing guides list Claude Opus output at US$25 per million tokens.

Developer Pankaj Kumar said early tests put V4 near Anthropic’s Opus models on some tasks and closer to GPT-5.6 Sol on coding, while public benchmark writeups cited about 91.2% on SWE-Bench Verified and a remaining gap on harder reasoning tasks.

DeepSeek’s API changelog says deepseek-chat and deepseek-reasoner will be discontinued on July 24 and now point to the non-thinking and thinking modes of DeepSeek-V4-Flash, suggesting migration mainly requires changing the model name.

The company previously made a permanent 75% cut to V4 Pro pricing as China’s AI model price war continued.

Some federal agencies and state governments in the US have banned DeepSeek on government devices over security concerns.

🔗 Source: 36Kr