本页为机器翻译。如有错误，请发邮件至 [email protected]。
方法
API 价值
“$200 a month for $7,000 of tokens. If you actually use it.”
这个数字的含义
API 价值 = 如果每周都用满起约束作用的额度（通常是每周额度），一个月的套餐内含用量按厂商自己的按量付费 API 标价计算的费用。这是上限，不是典型账单：大多数人用得少得多。
用量低于 Claude 套餐每周额度的约 20% 时，按 token 付费比订阅更便宜（Wokey 的盈亏平衡点，2026 年 10 月）。实际价值取决于你的模型组合、缓存命中率和额度用了多少。每行给出的是区间，而不是一个神奇数字。
我们只采用厂商官方数字或真实实测。如果厂商没公开额度、也没有人做过可信的实测，该行显示 未公开。我们不用猜测填空。
可信度等级
- 高
- 由公开的美元或积分额度按公开的 API 价格扣减计算得出。精确值。
- 中
- 由厂商自己的积分换算加上明确的工作负载组合计算得出，或在大量真实账号上实测得出。
- 低
- 单次受控测试、由其他档位推算、厂商的粗略估计，或正在复核的实测。
最后复核于 。月度 = 每周 × 4.345。
性价比
每 1 美元订阅费能换到多少 API 用量。 Claude Pro：每月 $20，可换到 $738–1,173 的 API 用量。
每周用满额度 · 按低估值排序，其次按高估值
- 低估值
- 至高估值
- 可信度（高、中、低）
Claude Pro 中API 用量相当于价格的 37–59×
每月付 $20 → $738–1,173 的 API 用量
Claude Max 5x 低API 用量相当于价格的 36–59×
每月付 $100 → $3.6k–5.9k 的 API 用量
Claude Max 20x 中API 用量相当于价格的 36–59×
每月付 $200 → $7.1k–11.7k 的 API 用量
Codex Pro 100 低API 用量相当于价格的 10–14×
每月付 $100 → $1k–1.4k 的 API 用量
Codex Pro 200 低API 用量相当于价格的 10–14×
每月付 $200 → $2.1k–2.9k 的 API 用量
Codex Pro 500 低API 用量相当于价格的 10–14×
每月付 $500 → $5.2k–7k 的 API 用量
Codex Plus 低API 用量相当于价格的 8.1–13×
每月付 $20 → $162–253 的 API 用量
MiniMax Build 低API 用量相当于价格的 7.3–9.9×
每月付 $132 → $963–1,309 的 API 用量
MiniMax Explore 低API 用量相当于价格的 7.1–9.7×
每月付 $55 → $393–534 的 API 用量
Codex Business (Standard seat) 低API 用量相当于价格的 6.5–10×
每月付 $25 → $162–253 的 API 用量
MiniMax Go 低API 用量相当于价格的 6–8.1×
每月付 $22 → $131–178 的 API 用量
GLM Max 中API 用量相当于价格的 5.9–13×
每月付 $168 → $992–2,117 的 API 用量
GLM Pro 中API 用量相当于价格的 5.3–11×
每月付 $80 → $425–907 的 API 用量
GLM Lite 中API 用量相当于价格的 3.9–8.4×
每月付 $18 → $71–151 的 API 用量
GitHub Max 高API 用量相当于价格的 2×
每月付 $100 → $200 的 API 用量
GitHub Pro+ 高API 用量相当于价格的 1.8×
每月付 $39 → $70 的 API 用量
GitHub Pro 高API 用量相当于价格的 1.5×
每月付 $10 → $15 的 API 用量
GitHub Business 高API 用量相当于价格的 1×
每月付 $19 → $19 的 API 用量
以表格查看数据
| 套餐 | 价格 | API 价值 /月 | × 价格 | 可信度 | 方法 |
|---|---|---|---|---|---|
| Claude Pro$20/月中 | $738– | 37×–59× | 中 | measured | |
| Claude Max 5x$100/月低 | $3,600– | 36×–59× | 低 | scaled | |
| Claude Max 20x$200/月中 | $7,121– | 36×–59× | 中 | measured | |
| Codex Pro 100$100/月低 | $1,042– | 10×–14× | 低 | scaled | |
| Codex Pro 200$200/月低 | $2,084– | 10×–14× | 低 | measured | |
| Codex Pro 500$500/月低 | $5,210– | 10×–14× | 低 | measured | |
| Codex Plus$20/月低 | $162– | 8.1×–13× | 低 | measured | |
| MiniMax Build$132/月低 | $963– | 7.3×–9.9× | 低 | computed | |
| MiniMax Explore$55/月低 | $393– | 7.1×–9.7× | 低 | computed | |
| Codex Business (Standard seat)$25/月低 | $162– | 6.5×–10× | 低 | scaled | |
| MiniMax Go$22/月低 | $131– | 6×–8.1× | 低 | computed | |
| GLM Max$168/月中 | $992– | 5.9×–13× | 中 | computed | |
| GLM Pro$80/月中 | $425– | 5.3×–11× | 中 | computed | |
| GLM Lite$18/月中 | $71– | 3.9×–8.4× | 中 | computed | |
| GitHub Max$100/月高 | $200 | 2× | 高 | computed | |
| GitHub Pro+$39/月高 | $70 | 1.8× | 高 | computed | |
| GitHub Pro$10/月高 | $15 | 1.5× | 高 | computed | |
| GitHub Business$19/月高 | $19 | 1× | 高 | computed |
未列入图表：44 个套餐中的 26 个
没有公开额度，也没有可信的实测，所以没有可画的。我们不做估算。
- Cursor + Grok Free, Plus, Premium, Super, Ultra：The weekly pool is published only relative to Plus (4×/15×/50×), never in dollars or tokens.
- Cursor + Grok Pro (legacy $20), Pro+ (legacy)：The Other Models pool is drawn at API price, but pool sizes are not published.
- Devin Free：Free quota is described only as 'light'; no tokens or dollars published.
- Devin Pro：Quota size isn't published in tokens or dollars, only that it's token-based and varies by model.
- Devin Max：Published only as 'significantly higher' than Pro; no tokens or dollars.
- Devin Teams (full seat)：Pro-equivalent quota per full seat; size not published.
- Devin Windsurf Pro (legacy $15) (closed)：Same unpublished quota as Pro.
- Devin Windsurf Teams (legacy $30/seat) (closed)：Same unpublished quota as Teams.
- GitHub Free：Free credit allowance is not stated in dollars or credits.
- Gemini Standard (closed), Enterprise (closed)：Quotas are model requests per day; Google does not publish tokens per request, and no credible measurement exists.
- Gemini Free / individual (Google login) (retired)：Retired.
- Kimi Plus, Pro, Max, Ultra：Kimi does not publish the monthly credit pool or 5-hour cap in tokens or dollars.
- Qwen Pro (closed)：The cap is 90,000 requests/month, but Alibaba does not publish tokens per request.
- Qwen Lite, Essential, Standard, Pro：Alibaba does not publish how many credits a token costs, so credits can't be converted to API value.
各套餐的 API 价值
| 方法 | ||||||
|---|---|---|---|---|---|---|
| Codex | Plus | $20/月 | $162–253/月 | ~8.1–13× | 低 | measured |
| Codex | Pro 100 | $100/月 | $1k–1.4k/月 | ~10–14× | 低 | scaled |
| Codex | Pro 200 | $200/月 | $2.1k–2.9k/月 | ~10–14× | 低 | measured |
| Codex | Pro 500 | $500/月 | $5.2k–7k/月 | ~10–14× | 低 | measured |
| Codex | Business (Standard seat) | $25/月 | $162–253/月 | ~6.5–10× | 低 | scaled |
| Claude | Pro | $20/月 | $738–1,173/月 | ~37–59× | 中 | measured |
| Claude | Max 5x | $100/月 | $3.6k–5.9k/月 | ~36–59× | 低 | scaled |
| Claude | Max 20x | $200/月 | $7.1k–11.7k/月 | ~36–59× | 中 | measured |
| Cursor + Grok | Free | 免费 | 未公开 | — | — | not published |
| Cursor + Grok | Plus | $8/月 | 未公开 | — | — | not published |
| Cursor + Grok | Premium | $30/月 | 未公开 | — | — | not published |
| Cursor + Grok | Super | $100/月 | 未公开 | — | — | not published |
| Cursor + Grok | Ultra | $200/月 | 未公开 | — | — | not published |
| Cursor + Grok | Pro (legacy $20) | $20/月 | 未公开 | — | — | not published |
| Cursor + Grok | Pro+ (legacy) | $60/月 | 未公开 | — | — | not published |
| Devin | Free | 免费 | 未公开 | — | — | not published |
| Devin | Pro | $20/月 | 未公开 | — | — | not published |
| Devin | Max | $200/月 | 未公开 | — | — | not published |
| Devin | Teams (full seat) | $40/月 | 未公开 | — | — | not published |
| Devin | Windsurf Pro (legacy $15)不再接受新用户 | $15/月 | 未公开 | — | — | not published |
| Devin | Windsurf Teams (legacy $30/seat)不再接受新用户 | $30/月 | 未公开 | — | — | not published |
| GitHub | Free | 免费 | 未公开 | — | — | not published |
| GitHub | Pro | $10/月 | $15/月 | ~1.5× | 高 | computed |
| GitHub | Pro+ | $39/月 | $70/月 | ~1.8× | 高 | computed |
| GitHub | Max | $100/月 | $200/月 | ~2× | 高 | computed |
| GitHub | Business | $19/月 | $19/月 | ~1× | 高 | computed |
| Gemini | Standard不再接受新用户 | $22.80/月 | 未公开 | — | — | not published |
| Gemini | Enterprise不再接受新用户 | $54/月 | 未公开 | — | — | not published |
| Gemini | Free / individual (Google login)已停售 | 免费 | 未公开 | — | — | not published |
| GLM | Lite | $18/月 | $71–151/月 | ~3.9–8.4× | 中 | computed |
| GLM | Pro | $80/月 | $425–907/月 | ~5.3–11× | 中 | computed |
| GLM | Max | $168/月 | $992–2,117/月 | ~5.9–13× | 中 | computed |
| Kimi | Plus | $19/月 | 未公开 | — | — | not published |
| Kimi | Pro | $39/月 | 未公开 | — | — | not published |
| Kimi | Max | $99/月 | 未公开 | — | — | not published |
| Kimi | Ultra | $199/月 | 未公开 | — | — | not published |
| MiniMax | Go | $22/月 | $131–178/月 | ~6–8.1× | 低 | computed |
| MiniMax | Explore | $55/月 | $393–534/月 | ~7.1–9.7× | 低 | computed |
| MiniMax | Build | $132/月 | $963–1,309/月 | ~7.3–9.9× | 低 | computed |
| Qwen | Pro不再接受新用户 | $50/月 | 未公开 | — | — | not published |
| Qwen | Lite | $8/月 | 未公开 | — | — | not published |
| Qwen | Essential | $16/月 | 未公开 | — | — | not published |
| Qwen | Standard | $25/月 | 未公开 | — | — | not published |
| Qwen | Pro | $80/月 | 未公开 | — | — | not published |
区间为每月数值，假设每周都用满起约束作用的额度。Claude Max 另含的 $100/$200 API 额度不计入。
各套餐的假设与来源
Codex
Plus$20/月→ $162–253/月 (~8.1–13×)低
Low end: SemiAnalysis GPT-6 Astra-only test ($162/mo). High end: Wokey measured $58/week on 8 accounts (Oct 12–19). Wokey flags this week as down 63% in tokens and 37% in API value vs Oct 1–8 with no OpenAI announcement, and remeasures Oct 26, so treat it as provisional.
Pro 100$100/月→ $1k–1.4k/月 (~10–14×)低
Not measured directly. SemiAnalysis finds Pro 100/200/500 give the same value per dollar since the Sep 29 change, so this is half of the Pro 200 figures. Wokey is still collecting (1 account).
Pro 200$200/月→ $2.1k–2.9k/月 (~10–14×)低
SemiAnalysis controlled tests on the new allowance: GPT-6.1 Sol ≈$2,084/mo, GPT-6 Astra only $2,897/mo (Sol figure via secondary quotes of the paywalled article). Grandfathered subscribers have ~2× until Oct 29. Wokey is still collecting (2 accounts).
Pro 500$500/月→ $5.2k–7k/月 (~10–14×)低
SemiAnalysis Astra-only test $6,955/mo; Sol end scaled to 2.5× Pro 200.
Business (Standard seat)$25/月→ $162–253/月 (~6.5–10×)低
OpenAI gives Business the same 5-hour estimates as Plus, so assumed equal to Plus value. Not measured.
Claude Code
Pro$20/月→ $738–1,173/月 (~37–59×)中
Low end: Wokey measured $170/week of real traffic on 474 Pro accounts (Oct 12–19) × 4.345 weeks. High end: SemiAnalysis controlled test, Max 20x on Opus 5.5 only ($11,726/mo) ÷ 10, using its finding that every Claude tier gives the same value per dollar.
Max 5x$100/月→ $3.6k–5.9k/月 (~36–59×)低
No direct measurement yet (Wokey: 1 account, publishes at 5). Scaled from tiers that were measured: Pro and Max 20x both measure ~36–37× price on Wokey, and SemiAnalysis finds equal value per dollar across Claude tiers, so low = 36× $100; high = SemiAnalysis Max 20x Opus-only figure ÷ 2. Excludes the separate $100/mo of API credits Max includes.
Max 20x$200/月→ $7.1k–11.7k/月 (~36–59×)中
Low end: Wokey measured $1,643/week on 17 accounts (Oct 12–19; weekly cap ≈9.7× Pro, not 20×). High end: SemiAnalysis controlled Opus 5.5-only agentic test. Excludes the separate $200/mo of API credits Max includes.
Cursor + Grok
Free免费→ 未公开
The weekly pool is published only relative to Plus (4×/15×/50×), never in dollars or tokens. SemiAnalysis tracks Cursor but its numbers are paywalled. Third-party models are billed at API list price, so they add no subsidy.
Plus$8/月→ 未公开
The weekly pool is published only relative to Plus (4×/15×/50×), never in dollars or tokens. SemiAnalysis tracks Cursor but its numbers are paywalled. Third-party models are billed at API list price, so they add no subsidy.
Super$100/月→ 未公开
The weekly pool is published only relative to Plus (4×/15×/50×), never in dollars or tokens. SemiAnalysis tracks Cursor but its numbers are paywalled. Third-party models are billed at API list price, so they add no subsidy.
Ultra$200/月→ 未公开
The weekly pool is published only relative to Plus (4×/15×/50×), never in dollars or tokens. SemiAnalysis tracks Cursor but its numbers are paywalled. Third-party models are billed at API list price, so they add no subsidy.
Pro (legacy $20)$20/月→ 未公开
The Other Models pool is drawn at API price, but pool sizes are not published.
Pro+ (legacy)$60/月→ 未公开
The Other Models pool is drawn at API price, but pool sizes are not published.
Devin (ex-Windsurf)
Free免费→ 未公开
Free quota is described only as 'light'; no tokens or dollars published.
Pro$20/月→ 未公开
Quota size isn't published in tokens or dollars, only that it's token-based and varies by model. Windsurf's per-day message estimates (Mar 2026) are too wide and too old to price. Extra usage is billed at API list price, so it adds no subsidy.
Max$200/月→ 未公开
Published only as 'significantly higher' than Pro; no tokens or dollars.
Teams (full seat)$40/月→ 未公开
Pro-equivalent quota per full seat; size not published.
Windsurf Pro (legacy $15)$15/月→ 未公开
Same unpublished quota as Pro.
Windsurf Teams (legacy $30/seat)$30/月→ 未公开
Same unpublished quota as Teams.
GitHub Copilot
Free免费→ 未公开
Free credit allowance is not stated in dollars or credits.
Pro$10/月→ $15/月 (~1.5×)高
1,500 AI credits × $0.01. GitHub charges credits at provider API list rates (e.g. Opus 5.5 $4 in / $20 out per MTok), so the allowance is worth exactly its credits. Includes a 500-credit flex portion; code completions are unlimited and not counted.
Pro+$39/月→ $70/月 (~1.8×)高
7,000 AI credits × $0.01 (3,100 flex).
Max$100/月→ $200/月 (~2×)高
20,000 AI credits × $0.01 (10,000 flex).
Business$19/月→ $19/月 (~1×)高
1,900 AI credits per user × $0.01, pooled across the org.
Gemini Code Assist / CLI
Standard$22.80/月→ 未公开
Quotas are model requests per day; Google does not publish tokens per request, and no credible measurement exists. Closed to new subscriptions.
Enterprise$54/月→ 未公开
Quotas are model requests per day; Google does not publish tokens per request, and no credible measurement exists. Closed to new subscriptions.
Free / individual (Google login)免费→ 未公开
Retired.
GLM Coding Plan (Z.ai)
Lite$18/月→ $71–151/月 (~3.9–8.4×)中
10,000 credits/week × API dollars per credit × 4.345 weeks. Z.ai publishes both the credit multipliers (GLM-5.3: 6.9 input / 1.7 cached / 24 output credits per 10k tokens) and GLM-5.3 API prices ($1.40 / $0.26 / $4.40 per MTok); at an agentic mix of 85–95% cache hits a credit is worth $0.00163–0.00174 of API usage at peak. Low end = all peak hours; high end = all off-peak (half the credits per token). Peak is 23:00–03:00 PT on weekdays, so US daytime use sits near the high end.
Pro$80/月→ $425–907/月 (~5.3–11×)中
60,000 credits/week × API dollars per credit × 4.345 weeks. Z.ai publishes both the credit multipliers (GLM-5.3: 6.9 input / 1.7 cached / 24 output credits per 10k tokens) and GLM-5.3 API prices ($1.40 / $0.26 / $4.40 per MTok); at an agentic mix of 85–95% cache hits a credit is worth $0.00163–0.00174 of API usage at peak. Low end = all peak hours; high end = all off-peak (half the credits per token). Peak is 23:00–03:00 PT on weekdays, so US daytime use sits near the high end.
Max$168/月→ $992–2,117/月 (~5.9–13×)中
140,000 credits/week × API dollars per credit × 4.345 weeks. Z.ai publishes both the credit multipliers (GLM-5.3: 6.9 input / 1.7 cached / 24 output credits per 10k tokens) and GLM-5.3 API prices ($1.40 / $0.26 / $4.40 per MTok); at an agentic mix of 85–95% cache hits a credit is worth $0.00163–0.00174 of API usage at peak. Low end = all peak hours; high end = all off-peak (half the credits per token). Peak is 23:00–03:00 PT on weekdays, so US daytime use sits near the high end.
Kimi Code
Plus$19/月→ 未公开
Kimi does not publish the monthly credit pool or 5-hour cap in tokens or dollars. The only public measurement is one subscriber's instrumented data on a legacy plan (Aug 2026: ~19–23M billable tokens/week, down from ~55–90M), which doesn't map to the current tiers. SemiAnalysis tracks Moonshot behind a paywall.
Pro$39/月→ 未公开
Kimi does not publish the monthly credit pool or 5-hour cap in tokens or dollars. The only public measurement is one subscriber's instrumented data on a legacy plan (Aug 2026: ~19–23M billable tokens/week, down from ~55–90M), which doesn't map to the current tiers. SemiAnalysis tracks Moonshot behind a paywall.
Max$99/月→ 未公开
Kimi does not publish the monthly credit pool or 5-hour cap in tokens or dollars. The only public measurement is one subscriber's instrumented data on a legacy plan (Aug 2026: ~19–23M billable tokens/week, down from ~55–90M), which doesn't map to the current tiers. SemiAnalysis tracks Moonshot behind a paywall.
Ultra$199/月→ 未公开
Kimi does not publish the monthly credit pool or 5-hour cap in tokens or dollars. The only public measurement is one subscriber's instrumented data on a legacy plan (Aug 2026: ~19–23M billable tokens/week, down from ~55–90M), which doesn't map to the current tiers. SemiAnalysis tracks Moonshot behind a paywall.
MiniMax M Plan
Go$22/月→ $131–178/月 (~6–8.1×)低
MiniMax's own estimate of ~1.7B tokens/month for this tier (from the Token Plan page; MiniMax says M Plan keeps Token Plan text limits) priced at MiniMax-M3 list prices ($0.30 in / $0.06 cache read / $1.20 out per MTok) under agentic mixes of 95% cache reads/0.6% output (low) to 84% cache reads/0.7% output (high). It is a vendor estimate, not a measurement, and the weekly window may bind first. If MiniMax counted no cache reads in its estimate, the figure would be several times higher.
Explore$55/月→ $393–534/月 (~7.1–9.7×)低
MiniMax's own estimate of ~5.1B tokens/month for this tier (from the Token Plan page; MiniMax says M Plan keeps Token Plan text limits) priced at MiniMax-M3 list prices ($0.30 in / $0.06 cache read / $1.20 out per MTok) under agentic mixes of 95% cache reads/0.6% output (low) to 84% cache reads/0.7% output (high). It is a vendor estimate, not a measurement, and the weekly window may bind first. If MiniMax counted no cache reads in its estimate, the figure would be several times higher.
Build$132/月→ $963–1,309/月 (~7.3–9.9×)低
MiniMax's own estimate of ~12.5B tokens/month for this tier (from the Token Plan page; MiniMax says M Plan keeps Token Plan text limits) priced at MiniMax-M3 list prices ($0.30 in / $0.06 cache read / $1.20 out per MTok) under agentic mixes of 95% cache reads/0.6% output (low) to 84% cache reads/0.7% output (high). It is a vendor estimate, not a measurement, and the weekly window may bind first. If MiniMax counted no cache reads in its estimate, the figure would be several times higher.
Qwen (Alibaba Model Studio)
Pro$50/月→ 未公开
The cap is 90,000 requests/month, but Alibaba does not publish tokens per request. Closed to new buyers.
Lite$8/月→ 未公开
Alibaba does not publish how many credits a token costs, so credits can't be converted to API value. (Extra bundles sell credits at $0.00075 each; that is a price, not API value.)
Essential$16/月→ 未公开
Alibaba does not publish how many credits a token costs, so credits can't be converted to API value. (Extra bundles sell credits at $0.00075 each; that is a price, not API value.)
Standard$25/月→ 未公开
Alibaba does not publish how many credits a token costs, so credits can't be converted to API value. (Extra bundles sell credits at $0.00075 each; that is a price, not API value.)
Pro$80/月→ 未公开
Alibaba does not publish how many credits a token costs, so credits can't be converted to API value. (Extra bundles sell credits at $0.00075 each; that is a price, not API value.)
采用的 API 标价
- Anthropic Opus 5.5 $4 in / $0.20 cache read / $20 out; Sonnet 5.5 $2 / $0.10 / $10; Fable 5.1 $10 / $0.25 / $50 每 100 万 token。 来源 ↗
- GitHub Copilot AI credits at $0.01, drawn at provider list rates (e.g. GPT-6.1 Sol $2 in / $0.10 cached / $10 out) 每 100 万 token。 来源 ↗
- Z.ai GLM-5.3 $1.40 in / $0.26 cached / $4.40 out; GLM-5.3-Flash $0.15 / $0.03 / $0.50 每 100 万 token。 来源 ↗
- MiniMax M3 (≤512k context) $0.30 in / $0.06 cache read / $1.20 out; M3.1 Flash $0.15 / $0.015 / $0.60 每 100 万 token。 来源 ↗
- Moonshot kimi-k3 $3.00 in / $0.30 cached / $15 out; kimi-k2.7-code $0.95 / $0.19 / $4.00 (reference only, no plan figure) 每 100 万 token。 来源 ↗
- Alibaba qwen3.7-plus (≤256K) $0.40 in / $1.60 out (reference only, no plan figure) 每 100 万 token。 来源 ↗
工作负载组合
根据 token 数或积分计算价值时，我们采用以缓存读取为主的智能体编程组合：
- 95% 缓存读取，约 4.4% 新输入，约 0.6% 输出，可复现 Z.ai 自己的表格（Lite ≈ 每周 48M token，缓存命中率 95%，高峰时段）。
- 84–85% 缓存读取，约 0.7% 输出，来自 Wokey 实测的 Claude Pro 流量。
第三方实测来自 Wokey（其网关后的真实账号，每周一重新测量，5 个以上账号时公布）和 SemiAnalysis（受控的单模型测试；付费内容，因此我们采用免费摘录和公开引用）。
无订阅：仅 API
DeepSeek API (no subscription)
DeepSeek sells no coding plan, only pay-as-you-go. It works in Claude Code through its Anthropic-compatible endpoint. Per 1M tokens, peak/off-peak: deepseek-flash $0.30/$0.15 in, $1.20/$0.60 out; deepseek-v4-pro $1.32/$0.66 in, $3.96/$1.98 out. Off-peak is outside 01:00–04:00 and 06:00–10:00 UTC on weekdays.
DeepSeek 价格 ↗邮件通知
额度一变，就通知你
套餐有变动时发一封简短邮件。不发垃圾邮件，我们也被限流。