Coinbase砍半AI開銷：GLM 5.2與Kimi路由實戰 | Coinbase Cuts AI Costs 50%: The Model-Routing Playbook

7月 30, 2026

By Kit 小克 | AI Tool Observer | 2026-07-31

🇹🇼 Coinbase砍半AI開銷：GLM 5.2與Kimi路由實戰

Coinbase 砍半 AI 開銷的真相：不是省更少，是用更聰明

Coinbase 執行長 Brian Armstrong 在 X 上公開了一件事：把公司內部預設的 AI 模型從 Anthropic Opus 4.8、OpenAI GPT-5.5 換成中國模型 GLM 5.2（智譜 Zhipu）與 Kimi 2.7（月之暗面 Moonshot AI），結果是 token 用量創新高，但整體 AI 開銷卻砍半。這不是靠少用，而是靠一套完整的模型路由策略。

五個省錢槓桿，哪個最有效？

Armstrong 列出五招：

閘道器預設模型改成 GLM 5.2、Kimi 2.7
依任務難度做模型路由（簡單任務丟便宜模型，複雜任務才叫貴的）
加強快取（cache hit rate 從 5% 衝到 60%，這招最猛）
精簡 context，減少不必要的 token
把花費攤到每個團隊的產出預期上，讓大家有感

真正的重點其實是快取命中率那條——多數團隊的 AI 帳單爆炸，不是因為模型選錯，是因為同樣的 context 被重複塞進 prompt 裡重算。

價格差距有多大？

GLM 5.2 每百萬 input token 約 1.4 美元、output 4.4 美元；相比之下 Opus 4.8 是 5 美元和 25 美元——貴上五倍。在 SWE-bench Pro 這類程式測試上，GLM 5.2 甚至一度贏過 GPT-5.5。但獨立測試顯示，實際跑起來的品質落差因任務而異，別只看跑分。

對開發者的啟示

Snowflake、AI 新創 Lindy 也跟進轉用開源中國模型，代表企業級市場正在對 Anthropic、OpenAI 的定價施壓。如果你的團隊也在燒 AI 預算，模型路由加快取這套組合拳，比單純換便宜模型更值得先做。當然，用中國模型也有資料合規與供應鏈風險要評估，不是無腦抄作業。

好不好用，試了才知道。

🇺🇸 Coinbase Cuts AI Costs 50%: The Model-Routing Playbook

Coinbase Cut AI Costs in Half — Without Cutting Usage

Coinbase CEO Brian Armstrong revealed on X that the company switched its default AI models from Anthropic Opus 4.8 and OpenAI GPT-5.5 to Chinese models GLM 5.2 (Zhipu AI) and Kimi 2.7 (Moonshot AI). Token usage hit record highs, yet total AI spend dropped by nearly 50%. The trick wasn'''t using less — it was smarter model routing.

Five Levers, One Standout

Armstrong outlined five moves:

Switch gateway defaults to GLM 5.2 and Kimi 2.7
Route tasks by complexity — cheap models for simple jobs, expensive ones only when needed
Improve caching (hit rate jumped from 5% to 60% — the single biggest win)
Trim context to cut wasted tokens
Tie spend visibility to expected output per team

The caching number is the real story. Most teams''' AI bills balloon not because they picked the wrong model, but because the same context gets re-sent and re-processed on every call.

How Big Is the Price Gap?

GLM 5.2 runs about $1.40 per million input tokens and $4.40 per million output tokens, versus $5 and $25 for Opus 4.8 — roughly 5x cheaper. On SWE-bench Pro, GLM 5.2 has scored ahead of GPT-5.5 in some benchmarks, though independent testing shows real-world quality varies by task — don'''t trust the leaderboard blindly.

What This Means for Builders

Snowflake and AI startup Lindy have made similar moves toward open-weight Chinese models, adding real pricing pressure on Anthropic and OpenAI at the enterprise level. If your team'''s AI bill is climbing, model routing plus aggressive caching is worth tackling before you swap providers. That said, using Chinese models brings data compliance and supply-chain questions worth weighing — this isn'''t a copy-paste decision.

You won'''t know until you try it.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士