Repository navigation
Kimi v2 plans: adopt quotas at half of same-price v1 tiers - #142
Conversation
New plans (Go/Plus/Pro/Max/Ultra) mapped to same-price legacy tiers per maintainer ruling 2026-10-26: v2 = v1 x 1/2. Plus/Pro/Max/Ultra get kimi-k3 and kimi-k2.7-code rows at 1.45/7.25/21.77/43.53 and 1.57/7.84/23.52/47.04 yi; Ultra derived as 2x Max per official Agent ratio; Go has no Kimi Code and gets no point. Screenshot weekly k3.1-pro measurement recorded research-only (tier unconfirmed).
|
I'll fix CI failures and address comments from users with write access that start with 'Devin'.
|
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
There was a problem hiding this comment.
Note
Newer findings are available below. Devin Review posted a newer report on this PR, in addition to the findings presented here.
🔍 Devin Review: 1 flag
Not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)
|
UI verification of the 8 new Kimi v2 rows (Devin test agent, local dev server on this branch) — all checks passed.
Verified in a recorded session: https://app.devin.ai/sessions/bcb9743bf6b74e4195fb9b1fc87b8fba |
🚀 Deploying Preview to Cloudflare 🚀Preview URL: https://devin-1793033449-kimi-v2-quota-real-api-pricing.inkcraftmg.workers.dev (commit b631dfc)This URL reflects your latest Preview deploymentPreview Deployments by commit
|
….1-pro measurement Maintainer's gifted Allegretto (v1, CNY 199) account: 37,362,656 raw k3.1-pro tokens vs console monthly 1.92% -> monthly pool 19.46B; the same window's 7-day 9.61% implies 389M/week, month/week = 5.005 matching the monthly = weekly x 5 rule. Replaces the 14.51B basis (July k3-256k-mixed ccusage 243.74M/84% back-calculation); the two samples measure different model-equivalents, so the gap is a rate difference, not necessarily a bigger pool. Cascades: moderato k3 2.90->3.89 (x4/20), allegro k3 43.53->58.38 (x60/20); v2 kimi-k3 half-rows follow the v1 base: Plus 1.45->1.95, Pro 7.25->9.73, Max 21.77->29.19, Ultra 43.53->58.38. kimi-k2.7-code rows unchanged (independent pure-model panel measurement). Evidence: data/research/kimi-allegretto-quota-round2-2026-10-27.json; DECISIONS.md 2026-10-27 entry; updatedAt -> 2026-10-27.
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
Resolve conflicts with #127 and regenerate data, charts and README stats.
There was a problem hiding this comment.
Devin Review found 2 new potential issues.
3 flags not posted on this PR by your GitHub settings — view them in Devin Review. (Configure)
| # 同名档国内外并点：price_usd 统一按国际版标价（KIMI_INTL），¥价为国内实付；Andante ¥49 无海外同名档 | ||
| ("kimi_allegretto_cn", "Kimi 会员 199", 199, "CNY", "kimi-k3", kimi_199_monthly_yi(), "medium", f"本机ccusage {KIMI_199_USED_TOKENS}/{KIMI_199_USED_FRACTION:.0%}反推周额度×Kimi月池{KIMI_MONTHLY_TO_WEEKLY:g}倍；kimi-adoption-round6-2026-09-08.json", "旧11.61亿→14.51亿：确认Kimi月池=周池×5，旧值误套项目通用4周；样本以k3-256k为主且含kimi-for-coding，非纯K3 1M实测；SWE1.7短时面板的模型/统计窗口不同，未替换基准；ACP14.28为旧模型旁证，不直接采用；未折算：ccusage 只存周合计，无分项"), | ||
| ("kimi_moderato_cn", "Kimi 会员 99", 99, "CNY", "kimi-k3", round(kimi_199_monthly_yi() * 4 / 20, 2), "medium", "199档×官方4/20；kimi-adoption-round6-2026-09-08.json", "旧2.32亿→2.90亿：随199档改用周池×5；继承K3-256K为主的混合负载估算，不是K3 1M纯模型实测"), | ||
| ("kimi_allegretto_cn", "Kimi 会员 199", 199, "CNY", "kimi-k3", kimi_199_monthly_yi(), "medium", f"维护者实测：Allegretto(赠送) k3.1-pro 单日 {KIMI_199_USED_TOKENS} ÷ 控制台月总{KIMI_199_MONTHLY_FRACTION:.2%} 直推月池；kimi-allegretto-quota-round2-2026-10-27.json", "19.46亿（2026-10-27换基）：k3.1-pro 单日37.36M ÷ 月总1.92% 直推；同窗7天9.61%得周池3.89亿、月/周=5.005吻合月池=周池×5；替换旧基14.51亿（2026-07 k3-256k混合ccusage 243.74M/84%反推）——模型等效不同（k3.1-pro vs k3-256k混合），非纯池变化；未折算：面板只存合计"), |
There was a problem hiding this comment.
🟡 Kimi K3 prices use another model's quota
On a K3.1 Pro sample, kimi_199_monthly_yi() assigns its raw-token allowance to K3. The models lack a verified credit-rate equivalence, so every derived K3 price uses an unsupported allowance.
Learn more
A point pairs a subscription with the model it serves; the benchmark score and effective price both follow that model (CONVENTIONS.md). The new Allegretto sample counts K3.1 Pro tokens, but this row calls the model K3. The research records that K3.1 Pro's credit rate relative to K3 is unknown (sample caveat). Its 19.46 亿 allowance therefore cannot establish K3's capacity; all K3 rows derived from this base inherit the mismatch.
Example: A reader compares Allegretto × K3 against another model using K3's benchmark score. Its 19.46 亿 allowance instead comes from a K3.1 Pro session; a different credit rate changes the actual K3 token limit and price.
Recommended fix: Keep the existing K3 basis until K3.1 Pro's rate can be converted to K3-equivalent tokens, or add a separately scored K3.1 Pro point without relabeling it as K3. Recalculate the dependent legacy and v2 K3 rows consistently.
Was this helpful? React with 👍 or 👎 to provide feedback.
There was a problem hiding this comment.
Noted — this was adopted by explicit maintainer ruling, with the caveat recorded: data/research/kimi-allegretto-quota-round2-2026-10-27.json flags the unverified k3.1-pro→k3 credit-rate equivalence, and DECISIONS.md (2026-10-27 section) notes the prior basis was itself a k3-256k-mix sample, so model-equivalence was non-identical on both sides. Keeping as ruled for now; a separately scored k3.1-pro point can be a follow-up if a benchmark score becomes available.
| # 2026-10-27 换基：Allegretto(赠送)账号 k3.1-pro 单日 37.36M ÷ 控制台月总1.92% 直推月池； | ||
| # 同窗 7天9.61% 得周池3.89亿，月/周=5.005 与月池=周池×5 吻合（kimi-allegretto-quota-round2-2026-10-27.json） | ||
| return round( |
There was a problem hiding this comment.
🟡 Kimi allowance skips workload normalization
When kimi_199_monthly_yi() uses the K3.1 Pro sample, it divides raw tokens by the monthly percentage. The sample includes cache, input, and output counts, so this bypasses the required workload normalization and skews dependent prices.
Learn more
Samples with token-type breakdowns must first convert each segment to list-price worth, then divide by the plan percentage and the project's standard blended price (token convention). This sample records approximately 36.7M cached, 548K uncached-input, and 120K output tokens (observed breakdown). Dividing the raw total directly by 1.92% treats this 98% cached workload as the standard 97% cached workload and changes both Kimi's reported allowance and its derived real price.
Example: Allegretto's 37,362,656 raw tokens consume 1.92% of the month. The current calculation publishes 19.46 亿 tokens regardless of the three segment prices, even though the standard-workload equivalent changes with those prices.
Recommended fix: Obtain the applicable official Kimi Code prices for this exact served model and convert the observed cache/input/output breakdown to list-worth before dividing by the dashboard percentage and standard blended price. If that pricing cannot be established, do not replace the K3 basis with this unnormalized sample.
Was this helpful? React with 👍 or 👎 to provide feedback.
There was a problem hiding this comment.
Deliberate per maintainer ruling: Kimi's console percentage is computed over the same raw token consumption shown in the screenshot, so raw ÷ percentage measures the pool in the unit the dashboard itself uses. The cache/input/output breakdown is preserved in kimi-allegretto-quota-round2-2026-10-27.json for worth-normalization if official list pricing for this model becomes available; the un-converted status is flagged as 未折算 in the row note.
Summary
Kimi launched new subscription plans (official 2026-10-19 pages): Go ¥49 / Plus ¥99 / Pro ¥199 / Max ¥699 / Ultra ¥1399 (international $19/$39/$99/$199), replacing the 7-day cap with a rolling 5-hour limit plus a shared monthly pool. They map 1:1 to the same-price legacy musical-name tiers (Andante→Go, Moderato→Plus, Allegretto→Pro, Allegro→Max, Vivace→Ultra).
Maintainer ruling (2026-10-26): each new-plan monthly allowance = the same-price legacy (v1) allowance × 1/2. This adds 8 adopted rows (all derived, medium):
kimi_plus_cn"Kimi Plus 99" (plan_gen=v2): kimi-k3 1.45亿, kimi-k2.7-code 1.57亿kimi_pro_cn"Kimi Pro 199" (plan_gen=v2): kimi-k3 7.25亿, kimi-k2.7-code 7.84亿 — now aMAIN_PLANSrepresentative tierkimi_max_cn"Kimi Max 699" (plan_gen=v2): kimi-k3 21.77亿, kimi-k2.7-code 23.52亿kimi_ultra_cn"Kimi Ultra 1399" (plan_gen=v2): kimi-k3 43.53亿, kimi-k2.7-code 47.04亿 — no same-price v1 tier, derived as 2× Max via the official Agent-credit ratio (28× vs 14×)Go gets no rows (no Kimi Code).
plan_gen_ofnow returnsv2for the four new plan ids (KIMI_V2_PLANS); data dates inherit from each same-price v1 counterpart. International USD prices merge into the same points viaKIMI_INTL.A maintainer-supplied dashboard screenshot (k3.1-pro, ~10.95亿 k3.1-pro-equivalent weekly pool) is recorded in
data/research/kimi-v2-plans-round1-2026-10-26.jsonas research-only — the account's plan tier is unconfirmed, so it adopts nothing.Also:
web/src/domain.test.ts"kimi 199" expectation now includes the twokimi_pro_cnrows;data/conventions.jsonupdatedAt→ 2026-10-26;data/i18n/adopted.en.jsonfilled for the 10 new zh strings;.gitignorenow covers/.venv312/; all pipeline checks and the 76 web tests pass. Charts were re-rendered on macOS with Noto Sans CJK (glyph outlines differ from the previous Windows/Microsoft YaHei render — data is identical, verified by all chart checks).Link to Devin session: https://app.devin.ai/sessions/bcb9743bf6b74e4195fb9b1fc87b8fba
Open in Devin Desktop: https://app.devin.ai/desktop/session/bcb9743bf6b74e4195fb9b1fc87b8fba?variant=devin
Requested by: @Fei2411