Volcengine Ark Coding Plan's overview page was updated on August 17, and both DeepSeek-V4-Flash and DeepSeek-V4-Pro are in its switchable model list (source). Coincidentally, DeepSeek's official API also moved to peak/off-peak pricing the same day. One is a flat monthly fee, the other a per-token hike—let's put the two side by side.
Price comparison: official per-token hike vs Ark flat fee
The core takeaway: DeepSeek's official API now bills per token at peak/off-peak rates, while Ark Coding Plan bills per request at a fixed monthly fee—two separate systems, so the official hike does not directly raise Ark's monthly fee. What you compare isn't unit price; it's which route costs less over a month of your actual usage.
DeepSeek official API: peak/off-peak pricing from today
DeepSeek's official API switches to peak/off-peak pricing from 2026-08-17 00:00, with off-peak at half the peak rate. Peak hours are 9:00–12:00 and 14:00–18:00 Beijing time, everything else off-peak (source). Two models are on sale:
| Model | Input (cache miss) | Output | Cache-hit input | Concurrency |
|---|---|---|---|---|
| DeepSeek-V4-Flash | ¥1.5 off-peak / ¥3 peak | ¥4.5 / ¥9 | ¥0.05 / ¥0.10 | 2500 |
| DeepSeek-V4-Pro | ¥4.5 off-peak / ¥9 peak | ¥13.5 / ¥27 | ¥0.15 / ¥0.30 | 500 |
Prices are per million tokens. This is hard per-token billing with separate input/output charges, suited to backends or low-frequency calls where you control token volume precisely.
Volcengine Ark Coding Plan: fixed monthly fee per request
Ark Coding Plan bills per request—Lite ¥40/mo, Pro ¥200/mo, with about 2.5× off for the first two months (Lite ~¥9.9, Pro ~¥49.9). Its quota is counted in requests, not tokens (source):
| Plan | List price | First 2 months | 5 hours | Weekly | Monthly |
|---|---|---|---|---|---|
| Lite | ¥40/mo | ~¥9.9 | 1,200 | 9,000 | 18,000 |
| Pro | ¥200/mo | ~¥49.9 | 6,000 | 45,000 | 90,000 |
One plan bundles DeepSeek, Doubao, GLM, Kimi, and MiniMax, with manual switching or Auto routing. That's the biggest difference from the official API: the official API only runs DeepSeek's own models, while Ark lets you hop across vendors.
Deduction coefficients: V4-Pro costs more, V4-Flash goes further
Not all models are equal under Ark's request count—each has a deduction coefficient. DeepSeek-V4-Pro is explicitly flagged as "higher deduction coefficient, burns quota faster" and recommended for hard problems; DeepSeek-V4-Flash has no such flag and sits at the normal tier (source).
That means the same Lite 18,000-request monthly quota drains faster on V4-Pro than on V4-Flash. Both default to thinking (can be disabled), share a 1024K (1M) context window, and support up to 384K output; the difference is agent strength and the coefficient. V4-Pro is still a "trial/preview" version—Ark suggests switching away under congestion—while V4-Flash is already GA.
Top-up, deals, and official access
Ark Coding Plan's official entry is the campaign page volcengine.com/activity/codingplan, and you subscribe rather than top up API balance. There are three deals: ~2.5× off for the first two months (Lite ¥9.9, Pro ¥49.9), a 9.5% first-order referral discount, and a 5% voucher for the referrer. From the third month it returns to ¥40/¥200 list price.
By contrast, DeepSeek's official API tops up balance and deducts per token, with gift balance used first and no official monthly coding plan. For a fixed monthly fee, you can also look at Tianyi Cloud token packages or Baidu Qianfan.
Quota, usage, and rate limits
Ark Coding Plan rate-limits through a dedicated key, the official API through account concurrency—don't mix the two. Ark's quota has three limits: a rolling 5-hour window, a weekly reset every Monday 00:00, and a monthly reset on the subscription's first day 00:00; when exhausted you wait for refresh without touching account balance.
Integration uses a dedicated Coding Plan key—Anthropic Base URL /api/coding, OpenAI /api/coding/v3—which only works inside coding tools, not API scripts or backends (source). The official API uses the OpenAI/Anthropic-compatible api.deepseek.com, with concurrency caps of 2500 for Flash and 500 for Pro.
Worth it: how to choose
Bottom line: for high-frequency interactive coding, Ark Coding Plan's flat monthly fee is easier to budget; for self-hosted backends or low-frequency on-demand use, the official API's per-token billing is more precise. Three decision signals:
- If you want one subscription to run DeepSeek, Doubao, GLM, Kimi, and MiniMax together, pick Ark—the ¥9.9–49.9 first-two-month entry is the lowest.
- If you only need DeepSeek, want cache-hit discounts, and prefer exact per-token math, pick the official API—V4-Flash off-peak input is ¥1.5 and cache-hit drops to ¥0.05.
- If you're an enterprise or backend service, skip Coding Plan (API calls are explicitly banned) and use Ark model API or DeepSeek's official API pay-as-you-go.
After today's DeepSeek official hike, Ark, Tianyi Cloud, and Baidu Qianfan fixed-fee plans become ready-made alternatives that aren't affected by peak/off-peak pricing—whether it's worth it depends on your usage pattern.
Caveats
A few traps to keep in mind before you jump on ¥9.9. Specifically:
- DeepSeek-V4-Pro is still a "trial/preview" version on Ark with a higher deduction coefficient; Ark suggests switching to another model under congestion.
- Ark Coding Plan quota works only inside AI coding tools—API scripts, backends, and non-interactive batch calls are banned and may lead to suspension or account bans.
- The 2.5× discount is a limited promo; from the third month it returns to ¥40/¥200, so estimate long-term costs at list price.
- Official deduction coefficients aren't published as numbers—check actual quota consumption in the console management page.