DeepSeek V4.1 Flash 1017

by DeepSeek · 552B (A8B/A16B) parameters

Current new · 1d budget open weights cheap tier

Email me when DeepSeek V4.1 Flash 1017 gets cheaper

Availability: Call it as deepseek-flash — the model name on DeepSeek's own rate card, whose MODEL VERSION column has read DeepSeek-V4.1-Flash-1017 since 2026-10-17. DeepSeek ships it as an update to V4.1 Flash under the same name, API and rates, but it is a new model in every way a buyer can measure: new weights (MIT, on Hugging Face and ModelScope), higher scores across reasoning, coding, agentic and vision benchmarks, and shorter outputs. The September build (DeepSeek V4.1 Flash, deepseek-v4-1-flash on this site) is no longer served. The retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted and are served by this build, billed at this price. On the same day DeepSeek opened its Search and Fetch tools to every API user at $0.75 per 1,000 calls, free until 2026-10-31 16:00 UTC; those are tool calls billed on top of the token rates shown here. V4 Pro keeps its own service and billing, so it does NOT route here.

Today's price · per 1M tokens

Input

$0.300

per 1M tokens

Output

$1.20

per 1M tokens

Blended

$0.525

blended $/1M — 3:1 weighted input:output

Cached input

$0.006

2% of input — prompt caching

How this price is scoped: DeepSeek bills this model at TWO rates depending on the hour. The tracked figure is the PEAK rate: $0.30 input (cache miss) / $0.006 input (cache hit) / $1.20 output per 1M tokens. Off-peak is exactly half: $0.15 / $0.003 / $0.60. Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday; every other hour, and all of Saturday and Sunday, is off-peak. Peak is the headline here because DeepSeek defines off-peak as half of peak rather than the other way round, so peak is the list rate, and because a caller who does not schedule around the clock needs the published number to be a ceiling rather than a floor. Halve it for a workload that runs entirely outside those hours. Third-party catalogues that quote this model at $0.15/$0.60 per 1M are printing the OFF-PEAK tier, not a different rate card.

Source: official DeepSeek pricing · read Oct 18, 2026 MODELPRICEWATCH.COM · 2026-10-18

Price receipt

We read DeepSeek's own pricing page on Oct 18, 2026 and found DeepSeek-V4.1-Flash-1017 listed with a price on the same column — that read is where the number above comes from, recorded as a new model. We keep the page text we read; its content hash is 87d1681893.

Overview

DeepSeek's 2026-10-17 release of V4.1 Flash: new open weights (MIT, on Hugging Face and ModelScope) on the same 552B-parameter causal encoder-decoder MoE (8B parameters active for input, 16B for output), with native visual understanding, a 1M-token context and 384K max output. It took over the deepseek-flash API name from the September build, which is no longer served, at the same price. DeepSeek reports a Pareto improvement in performance and cost: Terminal-Bench 4.0 31.2 to 42.3, Humanity's Last Exam without tools 36.8 to 42.6, DeepSWE v1.1 74.2 to 77.0, with faster task completion and shorter outputs, so a task costs fewer output tokens at unchanged per-token rates; it was trained with reinforcement learning from self-feedback for end-to-end project building. $0.30/$1.20 per 1M tokens at peak (01:00-04:00 and 06:00-10:00 UTC, Monday to Friday); half that in every other hour.

Run it yourself