DeepSeek ships V4 Flash with strong agent scores and low API prices

Why it matters

Agent products can burn through tokens on reasoning and tool calls. DeepSeek is cutting that variable cost while keeping the weights available for independent deployment.

DeepSeek, founded by Liang Wenfeng, released DeepSeek-V4-Flash-0731 on July 31st, pairing an MIT-licensed model with API prices low enough to change the cost assumptions behind coding agents and other token-heavy software.

DeepSeek describes the release as the official version of V4 Flash, replacing its preview checkpoint with "substantially enhanced agentic capabilities." An early independent evaluation supports part of that pitch: Artificial Analysis gave the model's max-reasoning configuration an Intelligence Index score of 50, ranking it third among 101 large open-weight models in its comparison class.

That placement matters because DeepSeek charges $0.14 per million cache-miss input tokens and $0.28 per million output tokens, according to its current API pricing page. Cached input falls to $0.0028 per million tokens. The model supports a 1 million-token context window and as many as 384,000 output tokens, giving developers room to run long coding, research and tool-use sessions without immediately inheriting the economics of a premium proprietary model.

The release extends the strategy Liang has pursued since creating DeepSeek in 2023: publish capable weights, charge aggressively for hosted inference and let outside developers turn the underlying research into products. Liang came to model development through quantitative finance. He studied information and communication engineering at Zhejiang University, then co-founded High-Flyer.

That background helps explain the focus on efficiency. Quantitative trading rewards improvements in throughput, latency and cost at scale. DeepSeek's model work applies the same discipline to inference, where a small difference in token price can determine whether an agent remains a demo or becomes a viable product.

The benchmark case is strong, with caveats

Independent developer Simon Willison highlighted the model's price-performance position in a July 31st post. He pointed to Artificial Analysis ranking V4 Flash ahead of MiniMax M3 and described it as a contender for the best value per unit of measured intelligence.

Willison also produced a useful reminder that reasoning settings affect the output developers actually receive. Running V4 Flash through OpenRouter at its default reasoning level generated a broken illustration of a pelican riding a bicycle. Raising reasoning_effort to high produced a coherent image with the bird seated on a recognizable bike.

The experiment is anecdotal, yet it exposes an operational tradeoff that aggregate leaderboards can hide. DeepSeek's strongest third-party result comes from the max-reasoning configuration, which can consume more output tokens and time. DeepSeek's low token prices reduce the financial penalty for extra computation, though developers still need to test latency and reliability inside their own workflows.

DeepSeek's model card also requires careful reading. The model card says V4 Flash scored 82.7 on Terminal-Bench 2.1, 54.4 on DeepSWE and 70.3 on Toolathlon-Verified. DeepSeek ran public coding-agent evaluations with a minimal version of DeepSeek Harness that the model card says will be released later. Two other reported results, DSBench-FullStack and DSBench-Hard, use internal DeepSeek test sets. Those numbers describe DeepSeek's own evaluation setup rather than fully reproducible outside validation.

The size figures vary by source as well. The Hugging Face repository reports 304 billion parameters, while Artificial Analysis lists 284 billion total parameters and 13 billion active for each token. DeepSeek says the official checkpoint uses the same structure as the earlier Flash model with an attached DSpark speculative-decoding module. Builders comparing memory requirements should keep the source-specific measurements separate.

Open weights still require serious hardware

The MIT license gives developers broad freedom to inspect, modify and serve V4 Flash. Deployment remains an infrastructure project. The Hugging Face files occupy roughly 167GB. Smaller quantizations may widen access, though the official checkpoint sits far outside the simple laptop workflow that has made compact coding models attractive.

That distinction is important for developers thinking about running the model themselves. DeepSeek is targeting a different layer: developers and infrastructure providers willing to manage a large mixture-of-experts model in exchange for stronger agent performance and control over deployment.

Self-hosting also gives U.S. companies a route to evaluate the weights without sending prompts to DeepSeek's hosted service. That separation will matter for organizations with strict data residency, procurement or security requirements. Open weights leave trust questions in place while giving engineering teams a way to inspect and operate the model inside infrastructure they control.

Liang's pricing bet

DeepSeek's cheap API creates pressure at the point where model vendors expect agents to drive consumption. Coding systems, research agents and automated business processes can generate long prompts, repeated tool calls and large reasoning traces. At higher proprietary-model prices, those loops can produce unpredictable bills before the product reaches meaningful scale.

Liang is offering founders another path: use DeepSeek's hosted endpoint for unusually cheap inference, or take the MIT-licensed weights and operate them independently.

DeepSeek-V4-Flash-0731 still needs broader testing across real repositories, long-running agents and production tool chains. Liang has put a high-scoring open-weight model into the market at a fraction of the token prices charged by many frontier providers, forcing competitors to defend the premium attached to every agent action.