Kimi Code limits, measured
Kimi does not publish quota figures for its Kimi Code plans. Wokey ran a controlled measurement on a Kimi Code Max account (¥699, new plan): one 5-hour window holds about 3.8M uncached input tokens of K3.1 Pro, or about 441K output tokens, and the monthly total is about 25 five-hour windows. For the same tokens K3.1 Pro uses about 50% more quota than K3 on input and 82% more on output, Highspeed uses exactly twice as much as Pro, and cache hits barely count.
- Controlled measurement, not an estimate
- Broken down by model
- Valued in API dollars
- K3.1 Pro · one 5-hour window
- 3.8M / 441K · uncached input only / output only
- Monthly total
- ≈ 25 windows · each 5-hour window is about 4% of the month
- K3.1 Pro vs K3
- +50% / +82% · quota for the same tokens (input / output)
- Highspeed vs Pro
- 2.00× · exactly double on input and output
Quota use by model (measured on Max)
For each model we sent requests with only uncached input and requests that were almost all output, then solved the weight of each token type from the quota readings. The first two columns are the share of the 5-hour quota per million uncached input and output tokens; the rest are how many tokens fit, with a month counted as 25 windows. Output includes thinking tokens.
| Model | Input / 1M | Output / 1M | Window: input | Window: output | Month: input |
|---|---|---|---|---|---|
| K3.1 Pro | 26.4% | 227% | 3.8M | 441K | 94.7M |
| K3.1 Pro Highspeed | 52.9% | 454% | 1.9M | 220K | 47.3M |
| K3 | 17.6% | 125% | 5.7M | 800K | 142.0M |
Cache: with all 45K tokens served from cache the quota reading barely moved, so cache hits effectively do not count; the first cache write counts as normal input. Kimi says k3 (1M context) uses about twice the quota of k3-256k; k3-256k was not measured here.
What one window is worth
Valued at K3.1 Pro’s official API price ($2 input / $10 output / $0.20 cache read per million tokens). Coding sessions are mostly cache reads, so a window’s value depends mainly on the cache share. The Wokey column is the same tokens paid per token through Wokey (30% of the official rate).
| Usage | Window: API price | Month: API price | Month: Wokey |
|---|---|---|---|
| Uncached input only | $7.6 | $190 | $57 |
| Output only | $4.4 | $110 | $33 |
| Coding session (93% cache, 0.3% output) | $14 | $350 | $105 |
| Long session (99% cache) | $25 | $625 | $188 |
The Max plan costs ¥699 a month, about $97 (at 1 USD ≈ 7.2 CNY). At each model’s own API price, a Highspeed window is worth the same as a Pro window ($7.6 of pure input), while a K3 window is worth far more ($17 input, $12 output): K3.1 Pro’s API price fell by a third but its quota use rose by more than half.
What each plan includes (official rules)
Since October 2026 the new plans have no weekly quota, only the rolling 5-hour window and the monthly total shared with the Kimi web app; legacy plans keep their rules. Prices are monthly billing.
| Plan | Per month | Membership quota | Kimi Code | K3.1 Pro / K3 | 1M context | Highspeed | Weekly quota |
|---|---|---|---|---|---|---|---|
| Go | ¥49 | base | — | — | — | — | — |
| Plus | ¥99 | 2× | ✓ | ✓ | — (256K) | — | none |
| Pro | ¥199 | 4× | ✓ | ✓ | ✓ | ✓ | none |
| Max | ¥699 | 14× | ✓ | ✓ | ✓ | ✓ | none |
| Ultra | ¥1,399 | 28× | ✓ | ✓ | ✓ | ✓ | none |
| Andante (legacy) | — | — | ✓ | — | — | — | yes |
| Moderato (legacy) | — | — | ✓ | ✓ | — (256K) | — | yes |
| Allegretto / Allegro (legacy) | — | — | ✓ | ✓ | ✓ | ✓ | yes |
“Membership quota” is the Agent quota multiple on Kimi’s purchase page (Go as the base, checked 2026-10-24). Kimi Code shares the monthly total with the membership; if that total scales the same way, Plus gets about 1/7 of Max and Ultra twice Max, but that is an inference. The Max window comes from the controlled run; other plans are in the production table below, updated weekly.
Window size by plan, measured in production (updated every Monday)
From 2026-10-19 – 2026-10-26: quota readings of 25 Kimi Code accounts on the Wokey gateway and the requests that went through them. Request tokens are converted to the share of a Max window they would use, with the per-model weights from the Max measurement, and compared with what each account actually used, giving each plan’s window size relative to Max. Medians are taken per cycle, account and owner; a plan is published once it has at least 5 accounts.
| Plan | Accounts | Window vs Max | K3.1 Pro input per window | Output per window | Windows per month | From Kimi’s multiple |
|---|---|---|---|---|---|---|
| Max (controlled run) | 1 | 1.00× | 3.8M | 441K | ≈ 25 | 1.00× |
| Allegretto (legacy) | 6 (6 owners) | 0.35× | 1.3M | 154K | — | — |
| Pro | 19 (14 owners) | 0.28× | 1.1M | 125K | ≈ 13 | 0.29× |
Owners’ own use of Kimi also counts against the quota, so production figures may read slightly low. “From Kimi’s multiple” converts the Agent quota multiple on Kimi’s purchase page, for comparison only.
Quota rules and the 5-hour limit
- The 5-hour window rolls: when it runs out the API returns 429 “You’ve reached your 5-hour usage limit. Your quota will reset when the current 5-hour window ends.” Retrying does not help; wait for the window to roll over, and the Kimi Code console shows when.
- The monthly total is shared with the Kimi web app, Kimi Claw and other features; when it runs out, Kimi Code is frozen until it refreshes.
- Legacy plans (Andante to Allegro) also have a weekly quota that refreshes every 7 days; new plans do not.
- Once the quota runs out you can turn on Extra Usage, billed by use at what Kimi calls close to its API price; for heavy use, upgrading is usually cheaper.
- Highspeed is about twice as fast as Pro (1.9× in Wokey’s measurement) and uses exactly twice the quota.
How to stretch the quota
- Keep the cache: cache hits barely count, so avoid restarting long sessions or switching model or thinking effort mid-session (both invalidate the cache).
- Use K3 for simple tasks: for the same tokens it uses about two thirds of K3.1 Pro’s quota.
- Use Highspeed only when someone is waiting on the answer; run background tasks and sub-agents on Pro.
- When you do not need 1M context, pick k3-256k for K3; Kimi says it uses about half the quota of k3.
- If you only hit the limit now and then, cover the overflow per token: K3.1 Pro through Wokey costs $0.60 input / $3.00 output per million tokens, with no 5-hour window.
Method and limits
- When: 2026-10-23 23:15 to 10-24 00:30 UTC. Account: Kimi Code Max (new plan, China region), with the 5-hour window almost unused at the start.
- We read Kimi Code’s /usages endpoint before and after every request (5-hour and monthly readings to 0.01%) and recorded a reading only once two reads 6 seconds apart agreed. Each request type ran 2–3 times, and readings within a type differed by at most 0.01%.
- Uncached input used about 45K tokens of non-repeating text; output used a long-writing request producing 5,000–12,000 tokens; cache used the same prompt four times in a row.
- Limits: the controlled run covers one Max account at one time of day (other plans come from the production data below); k3-256k and requests above 256K context were not measured; Kimi can change its quota rules at any time.
FAQ
How much quota does Kimi Code have, and how many tokens fit in a 5-hour window?
Kimi does not publish quota figures. Wokey measured the Max plan (¥699): K3.1 Pro uses about 26.4% of the 5-hour quota per million uncached input tokens and about 227% per million output tokens, so a window holds roughly 3.8M uncached input or 441K output tokens; K3 holds about 5.7M / 800K. Cache hits barely count, so cache-heavy long sessions get far more tokens through than these figures. The monthly total is about 25 five-hour windows. Pro and other plans are in the production table on this page.
Why does K3.1 Pro use more quota than K3?
For the same tokens K3.1 Pro used about 50% more quota than K3 on input and 82% more on output, even though its API price is a third lower ($2 / $10 vs $3 / $15). Kimi does not say how quota is charged, so we can only report the measurement: at each model’s own API price, a K3 window is worth $17 of pure input and a K3.1 Pro window only $7.6. When quota is tight, keep simple tasks on K3.
How is K3.1 Pro Highspeed’s quota counted?
Exactly twice Pro: about 52.9% per million uncached input tokens and 454% per million output tokens, so a window holds about 1.9M input or 220K output tokens. Highspeed is limited to Pro and above (legacy: Allegretto and above), and in Wokey’s measurement it runs about 1.9× as fast as Pro.
Kimi Code says the 5-hour quota is used up: what now?
The error is 429 “You’ve reached your 5-hour usage limit”. Retrying does not help: wait for the window to roll over (the console shows when), turn on Extra Usage (billed by use at close to the API price), upgrade, or switch to a pay-per-token API. When quota is tight, move from Highspeed back to Pro and simple tasks to K3.
Does Kimi Code still have a weekly quota?
The new plans since October 2026 (Plus, Pro, Max, Ultra) have no weekly quota, only the rolling 5-hour window and the monthly total. Legacy plans (Andante, Moderato, Allegretto, Allegro) keep a weekly quota that refreshes every 7 days. Both are subject to the shared monthly total.
Do cache hits count toward the Kimi Code quota?
Barely. When the same 45K-token prompt was sent several times in a row, the quota reading hardly moved once it hit the cache; the first cache write counts as normal input. Keeping long sessions and not switching model or thinking effort is the most effective way to save quota.
Is it true the Max plan is worth about $25 per 5 hours and $825 a month?
Only at the extreme. At K3.1 Pro’s API price a window is worth about $7.6 of pure input or $4.4 of pure output, about $14 for a coding session with 93% cache hits and about $25 at 99%. A month measured at about 25 windows, so $110–$625; $825 would take 33 windows, which does not match the measurement.
Is Kimi Code Max (¥699) worth it: subscription or pay per token?
If you use the full quota every month and your sessions are cache-heavy, the subscription wins: a month is worth about $350–$625 at API prices, far above ¥699 (about $97). If you do not use it all, keep hitting the 5-hour limit, or mostly send uncached input and long output, pay-per-token is easier to plan: the same tokens through Wokey cost 30% of the official rate, about $33 for a month of pure output and $105 for coding sessions.