Artificial Analysis, an independent benchmarking firm, priced the true cost of running DeepSeek’s V4-Flash through its Intelligence Index test suite at roughly three cents. TNW’s Ana Maria Constantin reported the figure on August 3, 2026, citing the firm’s own measurements. The gap to competitors is not close: Moonshot’s Kimi K3 ran to 86 cents, OpenAI’s GPT-5.6 Sol to $1.86, and Anthropic’s Claude Fable 5 to $3.15 for the identical battery of tasks.
AI Insiders covered V4-Flash’s production release from DeepSeek’s own model card yesterday, but this cost figure comes from outside the company, and it is the harder number to dismiss.
The methodology matters more than the headline figure. Artificial Analysis did not simply compare list prices per token. It tracked what each model actually spent, in dollars, to finish a fixed battery of tasks. That distinction separates a real cost comparison from a marketing one: a model with a low per-token rate can still run up a large bill if it needs many more steps, retries, or tool calls to reach the same answer. Cost per completed task captures that. Price per million tokens does not. That is why this ranking carries more weight than DeepSeek’s own price sheet, which lists V4-Flash at $0.14 per million tokens for input and $0.28 per million for output.
The two facts here do not cancel out. V4-Flash costs roughly a hundred times less to run than Claude Fable 5 on this suite, and it also trails the frontier by about nine points on the same test. It posted 50 on the Intelligence Index, matching Google’s Gemini 3.6 Flash. Muse Spark 1.1 and GLM-5.2 edge ahead at 51 apiece. Kimi K3 reaches 57, and the top tier, Opus 5, Fable 5, and GPT-5.6, clears that by roughly nine points more. A nine-point gap on this index shows up as more failed edge cases and more hand-holding on hard reasoning work, not as a rounding error. Cheapest and best are answers to different questions, and V4-Flash only answers one of them.
The timing is not incidental. DeepSeek turned its 75 percent discount into permanent policy earlier this year, a move that pushed rivals into cutting their own prices in response. OpenAI responded by cutting GPT-5.6 prices sharply. The company also closed its first funding round from outside investors this year, raising more than $7 billion, cash that lets it keep undercutting rivals on price while it grows its user base. Zack Kass, who led go-to-market at OpenAI, put a name to the dynamic: “diminishing model returns.” His argument is that once models cluster close enough in capability, price, not the benchmark, makes the final call.
TNW’s own caveat deserves repeating: benchmark scores are an imperfect stand-in for real-world work, and how much a task costs depends on how much a model has to say to finish it, not just its sticker price. A verbose model with cheap tokens can still lose to a terse model with expensive ones. Artificial Analysis’ number is a measurement of one suite under one methodology, not a universal exchange rate.
For high-volume deployments, coding assistants, backend automation, and consumer chatbots, where millions of calls turn fractions of a cent into a real invoice, V4-Flash’s economics are hard to ignore even with the capability gap intact. The question worth asking before the next vendor renewal is not which model tops the leaderboard, but whether the task at hand needs the nine extra points badly enough to pay a hundred times more for them.
TNW’s Ana Maria Constantin reported these findings on August 3, 2026, based on cost measurements from Artificial Analysis.
