AI Tools

DeepSeek V4 Flash vs V4 Pro: which should you pay for?

Published 2026-09-10 · Researched and price-checked on official sources

DeepSeek V4 Flash vs V4 Pro comes down to a 3x output-price gap: per DeepSeek's official rate card checked on September 10, 2026, Flash costs $0.66 per million output tokens off-peak versus Pro's $1.98, and Flash allows 2,500 concurrent requests against Pro's 500. Both share a 1M-token context window.

Key facts

What matters

Off-peak prices per 1M tokens: Flash is $0.007 cache hit, $0.22 cache miss and $0.66 output; Pro is $0.022, $0.66 and $1.98.
deepseek-v4-flash serves DeepSeek-V4-Flash-0731 and deepseek-v4-pro serves DeepSeek-V4-Pro-0813; both run a 1M-token context with 384K maximum output.
Flash allows 2,500 concurrent requests against Pro's 500, which makes Flash the wider choice for fan-out batch work.
Off-peak is exactly half of peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday.
DeepSeek's changelog puts Pro's Terminal Bench 2.1 score at 87.9 against Flash's 82.7, so the price premium buys a narrow agent lead, not a different tier of model.

What DeepSeek V4 Flash and V4 Pro actually are

deepseek-v4-flash is DeepSeek's low-cost API model, currently served as DeepSeek-V4-Flash-0731. deepseek-v4-pro is the premium tier, served as DeepSeek-V4-Pro-0813. Both are reached from the same base URL, https://api.deepseek.com, and both accept OpenAI-format or Anthropic-format requests.

The two share the same published envelope: a 1M-token context window, a 384K-token maximum output, and support for thinking and non-thinking modes. JSON output, tool calls, the Responses API and chat prefix completion are all marked as supported in both columns of DeepSeek's pricing table. The only capability that splits them is FIM completion, which runs in non-thinking mode only on both.

The verified DeepSeek rate card, per 1M tokens

DeepSeek bills per token and publishes one table with peak and off-peak columns. Off-peak rates are exactly half of peak, and Pro sits at roughly three times Flash across the board.

Cache-hit input: Flash $0.007 off-peak and $0.014 at peak, against Pro's $0.022 and $0.044.
Cache-miss input: Flash $0.22 off-peak and $0.44 at peak, against Pro's $0.66 and $1.32.
Output: Flash $0.66 off-peak and $1.32 at peak, against Pro's $1.98 and $3.96.

Because the multiple holds on every row, the saving does not depend on your input-to-output mix. It does depend on when you run: peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, and every other hour is off-peak at half price.

What the price premium buys: Pro's agent benchmarks

DeepSeek's changelog lists each model's own benchmark results. V4-Pro, generally released on August 13, 2026, reports Terminal Bench 2.1 at 87.9, NL2Repo at 61.5, DeepSWE at 62.7 and Toolathlon-Verified at 74.1. V4-Flash, released July 31, 2026, reports 82.7, 54.2 and 54.4 on the first three.

The gap is real but not lopsided. On terminal-style agent work Pro is about five points ahead; on repo-level code generation the split widens to roughly seven. A buyer paying three times the token price is buying headroom on long, failure-prone agent runs, not a different class of model.

One caveat on those figures: DeepSeek states they were produced with the DeepSeek Harness in minimal mode at max effort, topp=0.95 and temperature=1.0, so they are vendor-run and not independently reproduced.

Concurrency: Flash takes five times the parallel load

The number most comparisons skip is the concurrency limit in the pricing table's footnotes: 2,500 concurrent requests on deepseek-v4-flash against 500 on deepseek-v4-pro.

For a fan-out job that is the difference between a queue that drains and one that stalls. If you run thousands of short classification or extraction calls, Flash is not only cheaper per token, it is five times wider at the same moment.

Pro's lower ceiling fits its use case: fewer, longer, higher-stakes runs where you are not parallelising across thousands of independent requests.

What changes on September 10, 2026

A Hacker News thread on September 9, 2026 reported an in-product banner on DeepSeek's platform usage page announcing that a V4.1 Flash model would launch around September 10, that Pro requests would be routed to it at Flash's price until V4.1 Pro ships, and that new Flash rates take effect at 12:00 Beijing time on September 10: $0.003 cache hit, $0.15 cache miss and $0.60 output off-peak, with peak at double.

ToolBistro checked DeepSeek's official pricing page, changelog and the public API model list before that deadline. All three still listed only deepseek-v4-flash, deepseek-v4-pro and deepseek-v4-flash-vision-exp, and the changelog's newest entry remained August 21, 2026. Treat the V4.1 Flash numbers as reported, not published, until the official pages move.

If the change lands as described, a Flash output rate of $0.60 would sit about 9 percent below today's $0.66 and roughly 70 percent below Pro's $1.98, which would make the Pro column hard to justify for anything short of the hardest agent tasks.

Which DeepSeek model should you pick?

Pick deepseek-v4-flash for high-volume extraction, classification, summarisation and most coding-assistant traffic, where a five-point agent-benchmark difference does not change the outcome. It is about a third of the price and five times wider on concurrency.

Pick deepseek-v4-pro when a single failed long-horizon agent run costs more than the tokens you save: multi-step repository work, production debugging, or anything where DeepSWE-style performance decides the result. The premium is small next to a retry loop.

If you are cost-sensitive rather than latency-bound, schedule batch work outside 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays and take the automatic 50 percent cut.

The honest catch

Two things the rate card does not advertise. First, off-peak is half of peak rather than half of list price, so DeepSeek's framing rewards moving work into the quiet window instead of cutting your baseline bill. Second, if the reported Pro routing takes effect, a request sent to deepseek-v4-pro may be answered by Flash weights at Flash prices, so a benchmark you validated against Pro is not guaranteed to be what answers.

DeepSeek's own footer warns that product prices may vary and that buyers should re-check the pricing page regularly. Every figure above comes from that page and the changelog, both checked on September 10, 2026.

At a glance

Metric	deepseek-v4-flash	deepseek-v4-pro
Model version	DeepSeek-V4-Flash-0731	DeepSeek-V4-Pro-0813
Context length	1M tokens	1M tokens
Maximum output	384K tokens	384K tokens
Input, cache hit (off-peak / peak)	$0.007 / $0.014	$0.022 / $0.044
Input, cache miss (off-peak / peak)	$0.22 / $0.44	$0.66 / $1.32
Output (off-peak / peak)	$0.66 / $1.32	$1.98 / $3.96
Concurrency limit	2,500	500
Terminal Bench 2.1 (vendor-reported)	82.7	87.9
NL2Repo (vendor-reported)	54.2	61.5
DeepSWE (vendor-reported)	54.4	62.7

DeepSeek V4 Flash vs V4 Pro, verified September 10, 2026 (per 1M tokens unless noted)

FAQ

Is DeepSeek V4 Flash worse than V4 Pro?

On DeepSeek's own numbers, V4 Pro leads Flash by about five points on Terminal Bench 2.1 (87.9 against 82.7) and roughly seven points on NL2Repo and DeepSWE. Flash is not a different class of model, it is the same 1M-token envelope with narrower agent headroom for about a third of the token price.

When is DeepSeek off-peak pricing?

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. Every other hour is off-peak and priced at exactly half the peak rate, on both deepseek-v4-flash and deepseek-v4-pro.

Has DeepSeek V4.1 Flash launched?

DeepSeek's platform banner announced a September 10, 2026 V4.1 Flash launch with lower Flash rates, as reported on Hacker News. When we checked the official pricing page, changelog and API model list the same day, all three still listed only deepseek-v4-flash, deepseek-v4-pro and deepseek-v4-flash-vision-exp.