DeepSeek API Prices Changed Today: Which Line Rose 1,100%

Introduction

Headlines shouting that the DeepSeek API jumped 1,100% make it sound like every bill just multiplied by twelve. That is not what happened.

The 1,100% figure applies to one line only: V4 Pro cache-hit input during peak hours, from about $0.0037 to $0.044 per million tokens. The web app and the mobile app stay free. What actually changes is how you structure API spend and when you run jobs.

The new rates took effect at 00:00 Beijing time on Aug 17, 2026 (16:00 UTC on Aug 16). This site is an independent community. Treat the official pricing page as source of truth if prices move again.

I. Line up the timeline first

Three dates matter; the rest is noise:

1. Aug 6: official docs warned of a sizable overall API price increase.
2. Aug 12–13: DeepSeek-V4-Pro-0813 shipped on app, web, and API. The model ID is still deepseek-v4-pro. Thinking effort gained low / high / max, plus native OpenAI Responses API (Codex-friendly).
3. Aug 17, 00:00 Beijing: peak / off-peak rates went live. Off-peak is half of peak.

The feature surface is not the story here, and the call pattern did not change: OpenAI Chat Completions still hit https://api.deepseek.com; Anthropic format uses https://api.deepseek.com/anthropic. Wiring details live on the DeepSeek API page. For OpenAI Codex (CLI / desktop / VS Code), see DeepSeek Codex setup.

II. New price sheet

USD figures below follow the official English pricing page. “Before” is the public list from the V4 Pro GA window on Aug 12. Percent changes match the CNY sheet — the 1,100% headline is the same line in either currency.

Peak hours: 09:00–12:00 and 14:00–18:00 Beijing time. Everything else is off-peak. Unit: USD per million tokens.

DeepSeek V4 Pro (deepseek-v4-pro)

Meter	Before	Off-peak	Peak	Peak vs before
Input · cache hit	$0.0037	$0.022	$0.044	+1,100%
Input · cache miss	$0.44	$0.66	$1.32	+200%
Output	$0.88	$1.98	$3.96	+350%

Build DeepSeek-V4-Pro-0813. Concurrency cap 500.

DeepSeek V4 Flash (deepseek-v4-flash)

Meter	Before	Off-peak	Peak	Peak vs before
Input · cache hit	$0.0029	$0.007	$0.014	+400%
Input · cache miss	$0.15	$0.22	$0.44	+200%
Output	$0.29	$0.66	$1.32	+350%

Build DeepSeek-V4-Flash-0731. Concurrency cap 2,500 — better for high throughput.

III. Which line actually rose 1,100%

The wild percentage sits on cache-hit input that was already cheap, not on the output line that usually dominates a bill. Going from $0.0037 to $0.044 looks huge as a multiple and still small as dollars.

Most workloads feel output + cache-miss input. Pro peak output moved from about $0.88 to $3.96 (+350%). Off-peak output at $1.98 is still 2.25× the old flat rate. Across models, token types, and hours, this round lands roughly between +50% and +1,100%.

After the hike, V4 Pro peak output is still far below overseas flagships (public estimates put it around one-thirteenth of a Claude-class output price). The label shifts from “extreme bargain” to “still cheap, but you now schedule.” It is not free-to-unaffordable.

IV. Who is hit — and who is not

Match the change to the job:

What you are doing	What changes
Web / app chat and writing	No change — still free
Live support or an online agent that must run in peak hours	The bill jumps; retune model tier and caching
Nightly batches, evals, data cleanup	Shift everything off-peak and pay half
Stable system prompts plus reused long context	Cache hits stay the cheapest line — stop budgeting against $0.0037

One line: V4 Pro is the brain, the API is the meter, Harness is the hands. This update retunes the meter, not the web door and not DeepSeek Harness itself.

V. How developers can cut the new bill

Change the spreadsheet before you change production code:

1. Budget peak first. Size worst-case on the new peak sheet; size batch jobs on off-peak.
2. Move what can wait. Evals, backfills, and daily reports after 18:00 Beijing, or outside the two daytime peaks.
3. Prefer Flash unless you need Pro. High concurrency and simple tool calls belong on Flash (2,500-way). Hard reasoning and coding agents stay on Pro.
4. Make prefixes cache-hit shaped. Keep system prompts, tool schemas, and long-doc prefixes stable.
5. Turn thinking effort down by task. Use low for easy labels, high for daily agents, max only when the problem is actually hard — output tokens are the expensive ones.
6. Do not rename the model overnight. Keep deepseek-v4-pro / deepseek-v4-flash. The legacy IDs deepseek-chat and deepseek-reasoner already retired on Jul 24.

The request shape did not change. Minimal probe:

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -d '{"model":"deepseek-v4-flash","messages":[{"role":"user","content":"Hello"}]}'

Key signup and model choice: 2026 DeepSeek getting-started guide. July posts that said peak pricing “was coming” are now historical background only — see the older peak-pricing explainer.

VI. FAQ

Will the web app start charging? No. The change is DeepSeek API only. Web and app chat stay free.

Did every call just 12×? No. +1,100% is V4 Pro cache-hit input at peak. Peak output is about +350%; off-peak output about +125%.

Do I need to change the model ID tonight? No. Keep calling deepseek-v4-pro or deepseek-v4-flash and you get the current build.

Is this official documentation? No. This is an independent community briefing. Rates, peak windows, and feature flags follow the official pricing page.

Summary

The DeepSeek API new rates are live. Keep three facts: 1,100% is one meter line, not a 12× invoice; the web app is still free; off-peak, cache hits, and Flash vs Pro splitting are how you live with the new sheet.

Further reading

• DeepSeek API docs
• What is DeepSeek Harness?
• Web app and API key basics
• July peak-pricing preview (historical)
• DeepSeek web app