DeepSeek V4.1 Flash Pricing: What a Task Really Costs

I've watched a lot of AI price cuts. Most of them move a number on a pricing page. This one moves which work is worth doing, but only if you price it right. The easy version of the math prices it wrong.

This guide is part of DeepSeek V4.1 Flash changes the economics of AI agents. Here I stay on one question: what a finished task costs.

The full price sheet

From DeepSeek's pricing page, September 10, 2026. The same page lists a 1M-token context window, 384K maximum output and a concurrency limit of 2,500.

US business hours are off-peak

DeepSeek sets its peak hours in UTC, Monday through Friday. Converted to US time while daylight saving lasts:

Every hour of a normal American workday is off-peak. Weekends are off-peak. The one place US teams get caught is the overnight batch: a nightly job that kicks off at 2 a.m. Eastern runs at double the rate. Move it into the midnight-to-2 a.m. gap or after 6 a.m. and the same job costs half as much.

When US clocks fall back on November 1, each window arrives an hour earlier in local time. Re-check your schedule then.

Cached input is the number agent builders should watch

A cached input token costs $0.003 per million. An uncached one costs $0.15. That's a 50-to-1 gap, and agents live on the cheap side of it when they're built well, because they re-send the same instructions, tool list and history on every turn.

Here's one agent run of 20 turns. Each turn re-reads 40,000 tokens of shared context, adds 2,000 new tokens and writes 800.

Same model, same work, about five and a half times cheaper. Whether a given request hits the cache is up to DeepSeek's system, so treat the right-hand column as the best case and measure your hit rate from your own billing data.

Reasoning effort is a hidden multiplier

V4.1 Flash lets you set reasoning effort anywhere from 1 to 100. Higher effort means more thinking tokens, and thinking tokens bill as output. Every benchmark on DeepSeek's model card was run at 100.

That matters for your budget. The quality you read about came from the most expensive setting. I didn't find a published figure for how many tokens a max-effort answer uses against a low-effort one, so measure it yourself: run the same ten prompts at 20, 60 and 100, and log the output tokens and the answers side by side. You'll likely find a setting below 100 that's good enough for routine work.

Run your own numbers

This calculator prices one task on three models and adds the part most comparisons leave out: what a miss costs when a person has to fix it.

Uses uncached list prices as of September 10, 2026 (Sol at its promotional rate). The success rates are placeholders. Replace them with rates you measured on your own tasks, or the answer means nothing.

With the defaults, V4.1 Flash costs $34.50 a month in tokens against $1,250 for Opus 5. Add $2.00 per miss and it costs about $8,034 a month to finish the work against $4,250. Change the V4.1 Flash success rate to 96 and watch the order flip.

Three pricing mistakes to avoid

1. Quoting the off-peak rate for peak-hour work. Overnight batches and some third-party routes pay double.
2. Comparing token prices, not finished tasks. A model that's cheap per attempt and wrong more often can cost more per result.
3. Comparing against a promotion. OpenAI lists GPT-5.6 Sol's $4 input and $20 output as promotional pricing, available through at least November 21, 2026, so budget for the rate to change after that. Its prompts over 272,000 input tokens bill at twice the input rate and 1.5 times the output rate.