DeepSeek officially released DeepSeek-V4.1-Flash on September 10. Its API model name is deepseek-flash, and the same day at noon the pricing for the entire Flash lineup switched to a new tier. The internal test ID from two days earlier, deepseek-v4.1-flash-expires-on-0910, expired on the date written into its own name.
The company positions this release as "the smallest model in an entirely new model architecture series," with native multimodal vision understanding built in. The older model names deepseek-v4-flash and deepseek-v4-flash-vision-exp won't disappear right away — they've been temporarily routed to V4.1 Flash, so callers get the new version without changing any code.
What the new price list looks like
According to the official pricing page, deepseek-flash has a 1 million token context window and a maximum output of 384,000 tokens. Off-peak pricing per million tokens comes in three tiers: 0.02 yuan for a cache hit on input, 1 yuan for a cache miss, and 4 yuan for output. Peak-hour rates are exactly double — 0.04 yuan, 2 yuan, and 8 yuan.
The peak-hour definition carries over from the scheme DeepSeek has used since July: Beijing time 9:00–12:00 and 14:00–18:00, Monday through Friday, seven hours total; everything else counts as off-peak. That split tracks domestic working hours, so callers in other time zones end up with a lower blended average price than teams based in China.
DeepSeek is calling this change a "cut." Public reporting notes that after the peak-valley pricing adjustment in August, the cache-hit price for Flash during off-peak hours had at one point climbed to 0.05 yuan; the official pricing page shows only the currently active rate and keeps no historical record, so that earlier figure can't be verified directly from the page. What can be confirmed is the current set of numbers — and the other change it brings with it.
The name "Pro" currently has no model behind it
Starting at 12:00 on September 14, every request sent to deepseek-v4-pro will be routed wholesale to V4.1 Flash and billed at Flash rates. On the pricing page, deepseek-v4-pro still points to DeepSeek-V4-Pro-0813, whose off-peak rates are 0.15, 4.5, and 13.5 yuan, with peak rates of 0.30, 9, and 27 yuan.
Lined up side by side, the same model name sees its output price drop from 13.5 yuan to 4 yuan, and its cache-miss input price drop from 4.5 yuan to 1 yuan. For teams that have deepseek-v4-pro hardcoded into their configs, the bill will shrink on its own after Monday — but the model behind that name will have changed.
What the official announcement doesn't spell out is a timeline for V4.1 Pro. The routing rule is worded as "until V4.1 Pro launches," which establishes an order of events but not a date. In other words, the Pro line is currently sitting in a gap: the old Pro has stopped taking requests, the new Pro has no release window, and a Flash-sized model is standing in for both in the meantime. How long that gap lasts can't be confirmed right now.
Architecture and benchmark scores
The architectural details come from the official announcement: 552 billion parameters, a Causal-Encoder-Decoder structure, with input and output activation of 8 billion and 16 billion parameters respectively. That activation ratio is very low relative to the total parameter count — and it's the precondition that lets the model sit in the Flash tier at all.
In the benchmark results from official documentation, the model scores 90.9 on GPQA Diamond, 36.8 on HLE, a Codeforces rating of 3471, and 65.6 on MathArena Apex. Those numbers used to sit in the Pro tier's range; seeing them on the smallest model in the lineup — from a product-positioning standpoint — helps explain why the company felt comfortable rerouting Pro's traffic straight over.
Connecting the recent milestones in the Flash line
V4 launched in mid-July, and peak-valley time-based billing arrived alongside it, giving the Flash series its first off-peak/peak two-tier pricing. Around August 13 came a price adjustment. On the evening of September 8, an interim build valid for under 48 hours and capped at 20 concurrent calls per account opened for internal testing, with its expiration date written directly into the model name. On September 10, the official release and the new pricing arrived together, with the Pro routing change set for four days later.
Five moves in two months — a pace denser than anything this product line has seen before. Whether the new architecture can hold up under Pro's real-world load is a question callers will start answering for themselves after Monday.
Sources: DeepSeek API documentation changelog, CocoLoop, DeepSeek model and pricing page — cross-checked against the pricing page for per-tier rates, context length, and peak-hour definitions for deepseek-flash and deepseek-v4-pro.