Volcengine Ark Coding Plan limits, measured
Volcengine’s documentation says the Coding Plan has 5-hour, weekly and monthly limits, but gives no numbers and no per-model coefficients. The figures below come from real quota readings in production: Lite gets 2,000 units per 5 hours and Pro 10,000; one unit is 20,000 tokens on DeepSeek-V4-Flash and GLM-5.3-Flash, and about 4,000 on Doubao-Seed-2.1-pro. Tokens are input plus output, and cached input is counted in full.
- Real quota readings
- Split by model
- Valued at API prices
- Lite · one 5-hour window
- 40M / 8M · DeepSeek-V4-Flash / Doubao-Seed-2.1-pro, input plus output tokens
- Pro vs Lite
- 5× · 5× on the 5-hour, weekly and monthly caps, at 5× the price
- Weekly cap
- 3.5 windows · 10 a month; 3.5 full 5-hour windows use up the week
- Cache hits
- No discount · the coefficient is the same from 86% to 99% cached
Quota use by model
“Tokens per unit” is how many tokens (input including cache hits, plus output) one quota unit covers. The remaining columns are the tokens that fit in a window on each plan. The same window holds 5× as many tokens on the Flash models as on Doubao-Seed-2.1-pro.
| Model | Tokens per unit | Readings | Lite · 5 hours | Lite · month | Pro · 5 hours | Pro · month |
|---|---|---|---|---|---|---|
| DeepSeek-V4-Flash | 20,000 | 19,984, 19,999 | 40M | 400M | 200M | 2B |
| GLM-5.3-Flash | 20,000 | 20,000 | 40M | 400M | 200M | 2B |
| Doubao-Seed-2.1-pro | about 4,000 | 3,798, 3,959, 3,996, 4,046 | 8M | 80M | 40M | 400M |
DeepSeek-V4-Pro, DeepSeek-V4.1-Flash, Kimi-K3.1-Pro, GLM-5.3 and Doubao-Seed-2.1-lite have no clean samples yet and are not listed. Volcengine only says GLM-5.3, DeepSeek-V4-Pro and Kimi-K3 have higher coefficients, and runs time-limited discounts on some models, so coefficients can change.
What a Lite month is worth
A full Lite month spent on one model, valued at that model’s official API reference price. The plan counts cache hits in full while pay-per-token cache reads are cheap, so the higher the cache share, the less the plan is worth. The Wokey column is the same tokens paid per token through Wokey.
| Model and usage | Tokens per month | Official API price | Wokey per token |
|---|---|---|---|
| DeepSeek-V4-Flash · coding session (95% cache, 1% output) | 400M | $5.9 | $4.4 |
| DeepSeek-V4-Flash · almost no cache (10% output) | 400M | $78 | $58 |
| GLM-5.3-Flash · coding session (95% cache, 1% output) | 400M | $16 | $7.9 |
| GLM-5.3-Flash · almost no cache (10% output) | 400M | $74 | $37 |
| Doubao-Seed-2.1-pro · coding session (95% cache, 1% output) | 80M | $19 | $3.9 |
| Doubao-Seed-2.1-pro · almost no cache (10% output) | 80M | $97 | $19 |
Lite costs ¥40 a month, about $5.6 (at 1 USD ≈ 7.2 CNY); multiply tokens and value by 5 for Pro. In a coding session a month of DeepSeek-V4-Flash is worth only about $5.9 at API prices, close to the fee; with almost no cache it is worth about $78. Doubao-Seed-2.1-pro has a higher cache-read price, so a coding-session month is worth about $19.
Caps by plan (unpublished, read from the quota endpoint)
Units are the values the quota endpoint returns; Volcengine does not name them. All three limits apply at once. Prices are list monthly prices; the 75%-off promotion (¥9.9 Lite, ¥49.9 Pro for the first two months) ended on 2026-11-08.
| Plan | Per month | Per 5 hours | Per week | Per month (units) | Week ÷ 5 hours | Month ÷ 5 hours |
|---|---|---|---|---|---|---|
| Lite | ¥40 | 2,000 | 7,000 | 20,000 | 3.5 | 10 |
| Pro | ¥200 | 10,000 | 35,000 | 100,000 | 3.5 | 10 |
The Lite caps were read from 1 account and the Pro caps from 5 (identical). Another 10 accounts’ quota endpoint returns only percentages; their tokens per percentage point match these two plans.
Quota rules
- The 5-hour limit starts at the first request in a window and refreshes 5 hours later.
- The weekly limit resets every Monday at 00:00 Beijing time; the monthly limit resets with the subscription month.
- When quota runs out you wait for the next period; resource packs and account balance are not charged.
- Volcengine restricts the plan to AI coding tools; using it as an API can get the subscription suspended.
- A week holds only 3.5 windows: one full 5-hour window a day hits the weekly cap on day four.
How to stretch the quota
- Use DeepSeek-V4-Flash or GLM-5.3-Flash for everyday work: the same quota covers 5× the tokens of Doubao-Seed-2.1-pro.
- Cache hits save no quota, so every turn of a long context is charged in full. Compacting or starting a new session matters more here than on other plans.
- Keep high-coefficient models for work that needs them; Volcengine flags GLM-5.3, DeepSeek-V4-Pro and Kimi-K3 as fast to drain.
- If you only hit the limit now and then, cover the overflow per token: DeepSeek-V4-Flash through Wokey costs $0.112 input / $0.448 output per million tokens and $0.00224 for cache reads, with no 5-hour window.
Method and limits
- Data from 2026-11-07 to 11-14: 40 stretches of a single 5-hour window in which the used share moved at least 15 points, each matched with the tokens Wokey sent to that account in the same stretch.
- Coefficients use only stretches where one model carried at least 90% of the tokens. Three accounts read clearly low, meaning their owners were also using them, and were left out.
- The coefficients are round numbers (19,984, 19,999 against 20,000) and do not move between 86% and 99% cache share, so we read them as Volcengine’s fixed coefficients rather than statistical averages.
- Limits: the Flash coefficients come from 3 stretches on 2 accounts; Doubao-Seed-2.1-pro read 3,798–4,046 and is rounded to 4,000; other models are not measured. Volcengine can change coefficients or caps at any time; the date on this page is the last check.
FAQ
How many tokens does the Volcengine Ark Coding Plan include?
Volcengine does not say. Measured by Wokey, Lite gets 2,000 units per 5 hours, 7,000 a week and 20,000 a month; Pro gets 10,000, 35,000 and 100,000. One unit is 20,000 tokens on DeepSeek-V4-Flash and GLM-5.3-Flash, so a Lite 5-hour window is about 40M tokens and a month about 400M; on Doubao-Seed-2.1-pro a unit is about 4,000 tokens, so about 8M per window and 80M a month.
How do the Lite and Pro plans differ?
Pro has exactly 5× the quota on all three limits, at ¥200 against ¥40 a month, also 5×. Quota per yuan is the same, so choose by volume.
Is the Coding Plan quota short-changed?
Not in our readings: quota is deducted by a fixed coefficient. Three stretches on the Flash models read 19,984, 19,999 and 20,000. It feels small for two reasons: cache hits are counted in full, so every turn of a long coding session is charged for the whole context, and a week holds only 3.5 5-hour windows.
Do cache hits count toward the Coding Plan quota?
Yes, in full. Tokens per unit were the same in stretches from 86% to 99% cached. Kimi Code barely counts cache hits and Zhipu’s GLM Coding Plan counts them at about a quarter.
Which model uses the least quota?
Of the three measured, DeepSeek-V4-Flash and GLM-5.3-Flash cover 20,000 tokens per unit and Doubao-Seed-2.1-pro about 4,000, a 5× difference. Volcengine also flags GLM-5.3, DeepSeek-V4-Pro and Kimi-K3 as fast to drain; this page has not measured them yet.
How long does a week of quota last?
The weekly cap is 3.5× the 5-hour cap and the monthly cap 10×: at most 3.5 full 5-hour windows a week and 10 a month.
Is the Coding Plan worth it, or should I pay per token?
It depends on the cache share. A Lite month (¥40, about $5.6) spent on DeepSeek-V4-Flash is 400M tokens: in a coding session at 95% cache that is worth only about $5.9 at API prices, about the fee; with almost no cache, about $78. A coding-session month on Doubao-Seed-2.1-pro is worth about $19. If you do not use the full quota, or your sessions are mostly cache, paying per token is cheaper.