DeepSeek shipped V4.1-Flash on September 10, and the number on the marketing page is $0.003 per million cached input tokens, off-peak, first-party. That's roughly two orders of magnitude below what Anthropic charges for Claude Opus 5 cache reads ($0.50/M) and OpenAI charges for GPT-5.6 Sol cached input ($0.40/M). Standard input is $0.15/M and output is $0.60/M, putting a full off-peak million at $0.75 against $30 for either frontier competitor.

The "Flash" branding, however, is doing a lot of work.

Under the hood, V4.1-Flash is a 552-billion-parameter mixture-of-experts split across a 20-layer causal encoder and a 20-layer decoder. It activates 8B per token during prefill and 16B during decode. The checkpoint ships as 48 safetensors shards weighing 475 GiB, which isn't a laptop-class artifact. The previous V4-Flash was a 284B backbone with 13B active; the successor is nearly twice the parameter count with a smaller prefill footprint and a separate 196B Engram module handling conditional n-gram memory lookups.

That architecture is where the cache economics come from. Global KV cache is 890 bytes per token, roughly one-quarter of V4-Flash's HBM footprint and one-eighth of its SSD requirement. For anyone running long-context agents (the window is 1,048,576 tokens, with 384,000 max completion), the memory bill is the bill. DeepSeek has essentially rebuilt the transformer to keep the cache small and then priced cache reads accordingly.

The benchmarks are where the story gets textured. V4.1-Flash posts 90.6 on Terminal-Bench 2.1, 88.1 on CyberGym, and 90.9 on GPQA Diamond, within striking distance of Opus 5 (93.4) and GPT-5.6 Sol (94.1) on that last one. But it also scores 30.0 on Terminal-Bench 3.0 against Opus 5's 43.3, and TechTimes flagged DeepSWE v1.1 results ranging from 65.5% to 74.2% depending on which agent harness ran the eval. Stepping reasoning effort from 25 to 100 lifts average scores from 67.1% to 76.3% while consuming 2.5× more output tokens; effort 60–80 recovers most of the accuracy at a fraction of the spend.

Distribution is already broad. Eleven providers serve the model on OpenRouter, with Fireworks at $0.22/$0.66/$0.007 and SiliconFlow at $0.30/$1.20/$0.006. DeepSeek's own endpoint clocks 99.99% uptime and 122 tok/s at the fastest provider. Starting September 14 at 04:00 UTC, all deepseek-v4-pro requests route to V4.1-Flash at Flash rates until V4.1-Pro ships.

The pattern rhymes with Z.AI's GLM-5.3 Flash preview and Qwen3-8/27B's rapid Hugging Face uptake: Chinese labs pricing MIT-weighted MoE checkpoints as though the frontier tier's margin structure is a temporary condition. It's starting to look like one.

Sources

https://deepseek.com/en/news/deepseek-v4-1-flash/
https://api-docs.deepseek.com/news/news260910/
https://openrouter.ai/deepseek/deepseek-v4.1-flash
https://venturebeat.com/technology/deepseek-v4-1-flash-debuts-with-0-003-1m-off-peak-cached-input-rate-and-benchmarks-eclipsing-gpt-5-6-sol-claude-opus-5
https://www.techtimes.com/articles/327163/20260910/deepseek-v41-flash-cuts-agent-memory-costs-fourfold-new-architecture.htm