Deepseek.ai is an independent website and is not affiliated with, sponsored by, or endorsed by Hangzhou DeepSeek Artificial Intelligence Co., Ltd.

    DeepSeek V4 Pro API: 75% Price Cut Made Permanent

    Announced 23 May 2026 · Reviewed 25 July 2026

    Disclaimer: deepseek.ai is an independent community and fan site, not officially affiliated with DeepSeek AI. Rates reflect DeepSeek's first-party API pricing as reported by Artificial Analysis and may change. See our pricing page for the current live rates and the newer surge-pricing model.

    Announced 23 May 2026: DeepSeek has made its temporary 75% price cut on the first-party V4 Pro API permanent. The move puts V4 Pro firmly on the Pareto frontier of intelligence-vs-cost, alongside V4 Flash.

    The New V4 Pro Pricing

    Token Type Previous / 1M New / 1M Change
    Input$1.74$0.435−75%
    Output$3.48$0.87−75%
    Cached input$0.0036~99.8% off input

    Using Artificial Analysis's standard 7:2:1 blended pricing, V4 Pro lands at roughly $0.18 per 1M blended tokens.

    Running the Intelligence Index: What It Costs

    • DeepSeek V4 Pro: ~$268
    • Gemini 3.1 Pro Preview: ~$892 — V4 Pro is ~3× cheaper
    • GPT-5.5 (xhigh): ~$3,357 — V4 Pro is ~12× cheaper
    • Claude Opus 4.7 (max): ~$5,117 — V4 Pro is ~19× cheaper

    V4 Pro sits in the upper-right "most attractive quadrant" of the intelligence-vs-cost chart while costing a fraction to run.

    Why DeepSeek Can Do This

    1. MoE efficiency. V4's sparse 1.6T-parameter MoE architecture activates only a small slice of the network per token; inference cost scales with active parameters, not total parameters.
    2. Compressed attention. The compressed-attention mechanism cuts KV-cache pressure — the dominant cost of long-context inference.
    3. Aggressive prefix caching. The $0.0036/1M cached-input rate (a ~99.8% discount vs. fresh input) makes agentic workflows and long system prompts almost free on the input side. This is the lever pulling blended pricing down to ~$0.18.

    Combined with DeepSeek's recently-closed ~$7.4B funding round, the lab has both the capital runway and architectural edge to sustain frontier-tier pricing pressure.

    What This Means for Builders

    • Agentic systems: Multi-step agents that cost $5–20 per task on Claude Opus or GPT-5.5 can run on V4 Pro for cents.
    • Long-context RAG: Prefix caching at $0.0036/1M makes 100K+ token system prompts economically viable.
    • Reasoning at scale: Workloads that used to fall back to R-series for cost reasons no longer have to compromise.
    • Open-weight fallback: Because V4 is open-weight, you can self-host the same architecture if first-party pricing ever changes.

    Note on Surge Pricing

    DeepSeek has since layered a peak-hour surge premium on top of these permanent base rates and retired the legacy deepseek-chat / deepseek-reasoner endpoints. Off-peak base rates below still apply; peak-hour multipliers now apply on top.

    Bottom Line

    Making the 75% cut permanent reframes the entire frontier-AI cost curve. At ~$0.18 blended per 1M tokens, V4 Pro is meaningfully redefining what frontier-tier reasoning costs to operate. Western labs face a choice: match the pricing, differentiate on capability, or watch high-volume workloads migrate.