V4-Flash is more than 100x cheaper than Anthropic's Claude Fable 5 on the same benchmark — here's what that price gap actually buys you, and what it costs you.
DeepSeek dropped a new model on July 31, and the number turning heads isn't a benchmark score — it's the receipt. V4-Flash averages 3 cents to run a single benchmark task, according to independent testing firm Artificial Analysis. Anthropic's Claude Fable 5 needs $3.15 to do the same job. That's not a rounding difference — it's the same task costing over 100 times more depending on which model answers it.
The actual price gap, in one table
Cost per benchmark task (Artificial Analysis)
DeepSeek V4-Flash
Cost per task
$0.03
Intelligence Index
50
Kimi K3 (Moonshot)
Cost per task
$0.86
Intelligence Index
57
GPT-5.6 Sol (OpenAI)
Cost per task
$1.86
Intelligence Index
59+
Claude Fable 5 (Anthropic)
Cost per task
$3.15
Intelligence Index
59+
Cost per task
Intelligence Index
DeepSeek V4-Flash
$0.03
50
Kimi K3 (Moonshot)
$0.86
57
GPT-5.6 Sol (OpenAI)
$1.86
59+
Claude Fable 5 (Anthropic)
$3.15
59+
None of this is DeepSeek claiming to have built the smartest model on the planet. A 50 on the Intelligence Index is respectable, not remarkable — it ties Google's Gemini 3.6 Flash and sits roughly nine points behind Claude Opus 5 and GPT-5.6 on the same nine-benchmark suite spanning coding, reasoning, and workplace-style tasks. The claim is narrower: for a huge share of real workloads — classification, extraction, first-draft generation, anything run at volume — that nine-point gap matters a lot less than the 100x price gap. Enterprises have been saying exactly that all year, which is a big part of why OpenAI just cut GPT-5.6 pricing by 80% three weeks after launch.
Why it's this cheap
V4-Flash is a mixture-of-experts model. Hugging Face lists the repo at 304B total parameters, though only a fraction activates per token — roughly 284B forms the base network, with the rest a speculative-decoding draft module that predicts several tokens ahead before the full model verifies them. That's the same logic behind why MoE models can outrun dense ones of a similar size: you're paying compute for the experts that actually fire, not the whole network, on every token.
Running inference at $0.14 per million tokens is a hardware-economics story as much as a model-quality one. · Unsplash
This is also DeepSeek doing what DeepSeek does. Its R1 model triggered a real selloff in Western AI stocks in early 2025 by proving a frontier-adjacent model could be trained and served far cheaper than the US labs assumed. The company is reportedly prepping an IPO, and V4-Flash reads like a return to that playbook right as Kimi K3 rattled Nvidia's market cap the same way a few weeks ago, and Alibaba pushed Qwen3.8-Max out the door too. Cheap, capable-enough, open-weight Chinese models aren't a one-off story anymore. They're a pattern.
Can you run it yourself
The API pricing is DeepSeek hosting the model for you, and at hundreds of billions of active parameters with MoE routing, this isn't something you're loading onto a single consumer GPU. If you want a private, locally-run alternative in this price-conscious spirit, the realistic move is a smaller open model sized to your actual hardware — see what genuinely runs well on an 8GB card rather than chasing V4-Flash's parameter count at home. The MIT license means quantized, distilled versions will show up eventually; nobody's published one yet.
Quick answers
Is DeepSeek V4-Flash open source?
The model weights are published on Hugging Face under an MIT license, so yes — you can download, modify, and self-host it, though its size makes that impractical on consumer hardware.
How does V4-Flash compare to DeepSeek R1?
R1 triggered a global tech stock selloff in early 2025 by proving a frontier-adjacent model could be trained cheaply. V4-Flash continues that low-cost strategy on a newer architecture, not a direct upgrade of R1 itself.
Is V4-Flash actually as good as Claude or GPT-5.6?
No — Artificial Analysis scored it 50 on its Intelligence Index, roughly nine points behind Claude Opus 5 and GPT-5.6 on the same benchmark suite. It's a cost play, not a capability play.
Why is V4-Flash so much cheaper to run?
It's a mixture-of-experts model, so only a portion of its parameters activate per token, and DeepSeek has a track record of undercutting Western labs on serving costs going back to R1.
Verdict
Worth switching to?
If your workload is high-volume and tolerant of a mid-tier reasoning ceiling — support triage, tagging, first-draft copy — V4-Flash's price makes it close to a no-brainer to test. If you're running anything that needs careful multi-step reasoning or high-stakes coding, the nine-point benchmark gap will show up, and it's worth paying for a frontier model there.
Best for: Teams running high-volume, low-complexity API calls where cost per call actually matters
Watch what OpenAI and Anthropic do next. A 100x price gap on a model that's only nine benchmark points behind is the kind of number that forces a response — a price cut, a cheaper tier, or a public argument about why the benchmark doesn't capture what matters. DeepSeek has forced that response twice already. V4-Flash is trying to be the third.
