DeepSeek V4.1 Flash Pricing: The Forced Migration Nobody Opted Into

On September 10, 2026, DeepSeek released V4.1 Flash and made the cheapest headline in the AI price war cheaper still. Cache-hit input on the off-peak rate now costs $0.003 per million tokens. That is not a typo, and it is not a rounding artifact: it is roughly one sixth of the $0.022 that the outgoing V4 Pro charged for the same unit.

Then DeepSeek did something more interesting than a price cut. It announced that starting September 14, every request for the retiring V4 Pro would be automatically routed to V4.1 Flash, billed at the Flash price. No opt-in, no migration tool, no API key change. If you had pinned deepseek-v4-pro in your config, your traffic changes model underneath you on Monday.

That is the story heavy API users should read carefully. The savings are real. So is the fact that you did not choose them.

What deepseek v4.1 flash pricing actually changed

The official DeepSeek pricing page now lists two tiers, and the comparison explains the whole strategic shift. Prices are per 1M tokens.

Meter	V4.1 Flash off-peak	V4.1 Flash peak	V4 Pro off-peak	V4 Pro peak
Cache-hit input	$0.003	$0.006	$0.022	$0.044
Cache-miss input	$0.15	$0.30	$0.66	$1.32
Output	$0.60	$1.20	$1.98	$3.96

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. Every other hour, including all weekend, is off-peak.

Read the cache-hit column twice. The gap between a cached token and a freshly written one on V4.1 Flash is 50x off-peak ($0.003 versus $0.15). On V4 Pro that gap was 30x. The new model does not just lower the floor. It sharpens the penalty for having a bad cache-hit ratio.

Two other numbers matter for anyone running agents. The concurrency limit moved from 500 to 2,500 requests for Flash, a 5x increase. And the context window stayed at 1M tokens with a maximum output of 384K, now with native vision support that V4 Pro never had.

The architecture bet behind the number

DeepSeek is not buying market share with a loss leader. It rebuilt the model to make long agent sessions structurally cheaper.

V4.1 Flash uses what DeepSeek calls a Causal Encoder-Decoder split. The 552B-parameter Mixture-of-Experts backbone activates only 8B parameters during prefill (reading your input) and 16B during decode (writing the answer). The decoder’s global KV cache is projected directly from the encoder’s hidden states instead of being derived layer by layer.

The practical result: the KV cache is compressed to about 890 bytes per token. That is roughly a quarter of the GPU-memory footprint of V4 Flash and one eighth of the offloaded portion that sits on SSD or host RAM. Against DeepSeek V1, the per-token global cache is 437x smaller.

For a chatbot this is a nice efficiency win. For an agent that runs 40 tool calls across a 600K-token session, it is the difference between a session that fits and one that does not. The whole release is aimed at the cost driver that actually hurts heavy users: the memory bill for holding context open, not the price of a single prompt.

The benchmarks that justify the migration

The cost story only matters if quality holds. DeepSeek’s own technical report says it does, on the benchmarks that track agentic work.

V4.1 Flash scores 90.6 on Terminal-Bench 2.1, 74.2 on DeepSWE v1.1, and 88.1 on CyberGym. It beats V4 Pro on every agentic benchmark while using 3x fewer active parameters. On DeepSWE v1.1, that 74.2 narrowly edges both Claude Opus 5 and GPT-5.6 Sol.

The gaps are equally clear and DeepSeek does not hide them. On GPQA Diamond, a pure reasoning test, V4.1 Flash lands at 90.9 against Opus 5 at 93.4 and GPT-5.6 Sol at 94.1. The report also acknowledges a measurable gap on complex scientific agent tasks and on reading complex images. On ProgramBench it trails badly.

So the honest routing profile is: strong on coding agents and tool-calling loops, weaker on expert scientific reasoning and hard vision. That is the shape of the migration most teams should plan.

What the forced migration means for heavy users

Three concrete consequences follow from routing every V4 Pro request into V4.1 Flash on September 14.

First, your cache-hit discipline becomes your bill. If 80% of your input is cached reads, the input line is nearly free and your cost is dominated by output at $0.60 off-peak. If your prompts churn so that cache misses dominate, you pay $0.15 per million on input instead of $0.003, a 50x swing on the same workload. The model rewards stable prefixes, long system prompts, and consistent tool schemas more than any predecessor.

Second, your schedule is now worth money. Peak is double off-peak. A batch job that runs at 02:00 UTC pays full price; the same job at 12:00 UTC pays half. For teams already pushing background agent work to Saturday and Sunday, the weekend is entirely off-peak.

Third, quality regressions hide in the long tail. You swapped a model without a benchmark run on your own tasks. The vendor benchmarks say agentic coding improved. Your specific workload, especially anything reasoning-heavy or vision-heavy, may not have. The only way to know is to run your own eval before you assume the cheaper model is also the better one.

A migration checklist before September 14

Do this before the switch flips, not after.

• Export your current V4 Pro outputs for 20 to 50 representative tasks from the last month. That is your regression baseline.
• Re-run the same tasks against deepseek-flash after Monday and diff the results. Score them pass or fail against your own definition of done, not the vendor’s.
• Measure your real cache-hit ratio from token usage logs. If it is below 60%, restructuring your prompt prefixes is worth more than any provider discount.
• Move bursty, token-heavy batch work to off-peak hours and weekends. The 50% discount is the single largest lever on the bill.
• Keep a second provider warm. DeepSeek now carries a $71B valuation and a habit of changing routing on a Monday. A provider that can rewrite your model binding without asking is a provider you should not depend on alone.

The cheap model got cheaper and, on agentic coding, better. The catch is that cheaper is only cheaper if your workload matches the architecture it was rebuilt for. Measure the migration on your own tasks, and the savings are real. Skip that step and you have simply moved a cost you were not tracking onto a model you did not choose.