Mainland AI startup DeepSeek is expected to release a ‘fully-fledged’ V4 version as early as Monday, with the new version boasting performance approaching OpenAI's GPT-5.6 and introducing peak-valley billing for the first time.
The new version will be available in two types: DeepSeek V4 Flash and DeepSeek V4 Pro.
According to mainland media reports, DeepSeek V4’s overall capabilities are drawing close to Anthropic's flagship Opus 4.8 model, with a substantial upgrade in its autonomous task execution. While some developers noted that V4's agent functionalities have been greatly enhanced, it requires more computational iterations than Anthropic's Fable 5 to complete identical complex tasks.
Industry analysts suggest that although V4 might remain slightly behind Moonshot AI’s recently released Kimi K3 in handling long-context window tasks, DeepSeek maintains a significant pricing advantage.
In addition to performance, DeepSeek offers peak-valley pricing for enterprise and API users. Under the new scheme, generating one million words using the high-performance V4 Pro version costs just US$0.87 (HK$6.79) during off-peak hours, doubling to US$1.74 during peak times. The lighter V4 Flash version is priced at US$0.28 during off-peak windows and US$0.56 during peak hours. Furthermore, if a user submits a duplicate query previously processed by the system, the charge is only US$0.0028, the report added.
