DeepSeek-V4-Flash Is Now the Cheapest Major AI Model to Run — and It's Built for Agents

DeepSeek has launched DeepSeek-V4-Flash (version tag: V4-Flash-0731), and according to research cited by Reuters, it is now the cheapest well-known AI model available to run — by a significant margin. The release lands at a pivotal moment for enterprises evaluating the economics of scaling agentic AI workloads, where inference costs can compound rapidly as agent loops multiply across systems.

The Numbers That Make It Exceptional

DeepSeek's official API pricing puts V4-Flash at $0.14 per million input tokens (cache miss) and $0.28 per million output tokens. Cache hits drop the input cost to just $0.0028 per million tokens — effectively rounding to zero for heavily repeated contexts such as system prompts or large shared documents.

To put those figures in context: GPT-4o is priced at approximately $2.50 per million input tokens and $10 per million output tokens; Claude Sonnet 4 sits around $3 and $15 respectively. DeepSeek-V4-Flash is roughly 18× cheaper on input and 35× cheaper on output than frontier models from OpenAI and Anthropic. Even DeepSeek's own flagship, V4-Pro, costs $0.435 per million input tokens and $0.87 per million output tokens — making V4-Flash roughly three times cheaper than its sibling for uncached requests.

The model ships with a 1M-token context window and a maximum output of 384K tokens per call, with a concurrency limit of 2,500 simultaneous requests — more than five times the 500-request ceiling on V4-Pro.

Agentic Capabilities: The Strategic Addition

Cost alone does not explain why V4-Flash matters to the agentic AI segment. DeepSeek's own announcement emphasises that agent capabilities have been significantly enhanced in this release. The model supports Tool Calls natively, enables JSON-structured output, and is the first DeepSeek model to support the Responses API — an OpenAI-compatible interface designed explicitly for agentic orchestration, allowing multi-turn agent loops to maintain state and tool context across calls without manual prompt assembly.

It also supports both thinking and non-thinking modes, giving developers the ability to trade latency for reasoning depth at runtime — a meaningful lever for agentic applications where some steps require deliberate multi-step reasoning and others require fast reactive responses. The model additionally supports the Anthropic API format alongside the OpenAI-compatible endpoint, reducing integration friction for teams already running Claude-based agent stacks.

Why Cost Is the Decisive Variable at Agentic Scale

In traditional LLM use cases — single-turn question answering, document summarisation, one-shot code generation — inference cost is a line item. In agentic architectures, it becomes a structural variable. An agent completing a 50-step workflow, each step generating tool calls and receiving tool responses, can consume tens of thousands of tokens on a single task. Multiply that across an enterprise running hundreds of concurrent agent sessions and the difference between $0.28 and $10 per million output tokens is not marginal — it determines whether agentic AI is economically viable at that scale.

This is the market DeepSeek is explicitly targeting. The high concurrency ceiling (2,500 requests) and the Responses API support both point toward use cases involving large fleets of simultaneously running agents, not occasional interactive queries.

The Competitive Pressure This Puts on Western Labs

The gap between DeepSeek's pricing and that of OpenAI and Anthropic has widened with each major DeepSeek release. V4-Flash continues that pattern. Both US labs have reduced prices over time, but they are doing so from a cost base that includes substantially higher US labour and infrastructure costs, and from business models that cross-subsidise safety research, policy engagement, and consumer product development.

For enterprise buyers making infrastructure decisions today, V4-Flash represents a credible production-grade option for high-volume agentic workloads — provided that data residency, compliance requirements, and model provenance concerns can be satisfied. For the wider industry, it sustains the pressure on frontier-model pricing that DeepSeek first applied when it released R1 in early 2025.

DeepSeek notes that a peak/off-peak pricing policy — applying a 2× multiplier during Beijing business hours — will take effect from a date to be announced. Even at peak prices, V4-Flash remains materially cheaper than its Western counterparts.