DeepSeek Releases V4-Flash, Priced 100x Cheaper to Run Than Anthropic's Top Model

China's AI price war just produced its cheapest casualty yet. Since DeepSeek's R1 model rattled U.S. tech stocks back in early 2025 by undercutting Western pricing, the company has kept grinding on cost. Its newest release, V4-Flash, launched Friday and now sits at the bottom of the market on price by a wide margin, according to Artificial Analysis, the San Francisco research firm that benchmarks AI models.

The numbers are blunt. Artificial Analysis estimates V4-Flash costs about 3 cents to run through its standard battery of tests. Moonshot AI's Kimi K3 costs 86 cents. OpenAI's GPT-5.6 Sol runs $1.86. Anthropic's Claude Fable 5, currently the industry's top-scoring model, costs $3.15. Reuters reported that works out to V4-Flash running more than 100 times cheaper than Anthropic's flagship in realized terms.

The pricing itself is straightforward: DeepSeek charges 14 cents per million input tokens and 28 cents per million output tokens, according to Artificial Analysis. Tokens are the basic unit of data AI models chew through to read a prompt and generate an answer. What makes V4-Flash's 3-cent benchmark cost notable is that the model is unusually token-hungry. It uses more tokens than competitors to do the same job and still comes out cheapest, because the per-token rate is that much lower. That's a distinction ZeroHedge flagged clearly, noting the low headline price is "doing all the work" once you account for how verbose the model is.

That verbosity detail matters for anyone actually budgeting compute costs. A model that charges less per token but needs three times the tokens to finish a task can end up costing more in practice. That's why Artificial Analysis leans on its benchmark-cost figure rather than sticker price alone. NDTV Profit made the same point: raw pricing without a task-based benchmark can be misleading.

Where V4-Flash falls short

DeepSeek isn't claiming to have built the smartest model. V4-Flash scored 50 out of 100 on Artificial Analysis's Intelligence Index, a composite benchmark covering coding, reasoning, and workplace-style tasks. That ties Google's Gemini 3.6 Flash and lands a point behind Meta's Muse Spark 1.1 and Zhipu's GLM-5.2.

The real gap is with the frontier models built for hard problems. Anthropic's Claude Opus 5 and Fable 5, along with OpenAI's GPT-5.6, all scored at least nine points higher than V4-Flash, according to Digit.fyi's rundown of the same Artificial Analysis data, which put Opus at 61, Fable 5 at 60, and GPT-5.6 Sol at 59. Moonshot's Kimi K3 scored 57.

For routine, high-volume work—summarizing documents, boilerplate code, back-office automation—V4-Flash is now the cheapest credible option on the market. For difficult, multi-step agentic tasks where small errors compound at every step, the frontier models still justify their price.

The bigger fight

V4-Flash isn't an isolated launch. It's the latest move in what ZeroHedge called an 18-day pricing barrage among Chinese AI labs fighting for domestic market share, with the fallout spilling into global pricing. Moonshot shipped its 2.8-trillion-parameter Kimi K3 on July 16 and briefly took the top spot on Arena's Frontend Code leaderboard. Alibaba unveiled Qwen3.8-Max the same week, according to NDTV Profit, matching the scale of Moonshot's model. ByteDance, MiniMax, and Z.AI are all in the same race.

DeepSeek itself is reportedly preparing a more powerful V4-Pro model, though the company hasn't announced a launch date, NDTV Profit reported. The outlet also noted DeepSeek is preparing for a potential initial public offering, a detail that adds context to why the company is pushing hard on visibility and price right now rather than resting on R1's reputation from last year.

For U.S. AI labs, the pressure is now structural rather than a one-time scare. When R1 launched in early 2025, it triggered a stock selloff and a debate about whether American companies were overspending on compute. A year and a half later, there's now a documented, benchmarked price gap of two orders of magnitude between a usable Chinese model and Anthropic's best product. Whether Anthropic, OpenAI, and Google respond by cutting prices on their own lower-tier models, or instead lean harder into arguing frontier capability justifies the premium, is the open question the next earnings cycle and product releases will answer.