DeepSeek V4-Flash Operating Costs Far Below Comparable Models
0xBroomberg
DeepSeek released V4-Flash on August 1 with an average benchmark run cost of roughly $0.03 — about one-hundredth of Anthropic's Claude Fable 5, marking a new floor in large-model pricing.
Three cents per benchmark run — how cheap is that?
Research firm Artificial Analysis estimates V4-Flash completes one benchmark run for roughly $0.03 on average.
For comparison: Moonshot's Kimi K3 costs about $0.86, OpenAI's GPT-5.6 Sol about $1.86, and Anthropic's Claude Fable 5 about $3.15.
In plain terms = running the same test suite, V4-Flash spends 3 cents while Claude Fable 5 spends over 3 dollars — a gap of roughly 100×.
Why not just compare listed API prices?
V4-Flash's API is priced at $0.14 per million input tokens and $0.28 per million output tokens.
But Artificial Analysis notes that listed prices alone are misleading — different models process and generate different amounts of data to finish the same task.
This means → measuring "cost to complete one task" captures the user's real bill more accurately, and that is exactly where V4-Flash's edge is sharpest.
How does it actually perform — is cheap also good?
On Artificial Analysis's intelligence index (out of 100), V4-Flash scores 50, tied with Google's Gemini 3.6 Flash and one point below Meta's Muse Spark 1.1 and Zhipu's GLM-5.2.
Kimi K3 scores 57; Claude Opus 5, Fable 5, and GPT-5.6 each beat V4-Flash by 9 points or more.
This means → V4-Flash has not broken into the top performance tier. Its pitch is not "smartest" — it is "best value."
What comes next?
DeepSeek is preparing a higher-performance V4-Pro variant, though no official release date has been announced.
This reflects DeepSeek's own recognition that cost leadership alone is not enough — the performance gap must close.
In plain terms = V4-Flash proved DeepSeek can push prices to the floor. V4-Pro must answer the next question: can you build a ceiling on top of that floor?
Content is for reference only, not financial advice.