Data-backed head-to-heads for the gateway questions people actually search — each grounded in the same reproducible cost benchmark and scorecard.
📊 关键数字 · 省钱的来源不是路由器本身——同一个编码 Agent 会话,换个模型成本就从 $0.021 到 $1.75,差 ~83 倍(实测,单测脚本可复算)。风险的来源也不是路由器的牌子,而是机制:自带 API key(BYOK)是官方支持的用法;把自己订阅的 OAuth…
📊 Key numbers · The savings don't come from the router — the same coding-agent session costs $0.021 to $1.75 depending on the model behind it, a ~83×…
📊 Key numbers · OpenRouter adds ~5.5% on credits; Vercel AI Gateway and Helicone pass through at 0% markup. But the bigger lever is the model, not the…
📊 Key numbers · Writing a 100K-token report costs $0.03 on DeepSeek V4-Flash vs $3.01 on GPT-5.5 — a ~106× spread (reproducible benchmark, computed…
📊 Key numbers · LiteLLM is the most-starred OSS gateway (⭐ 54.9k) but shipped two serious 2026 CVEs — a pre-auth SQL injection and an unauthenticated RCE…
📊 Key numbers · Both gateways are 0% markup, so the deciding cost factor isn't the gateway — it's the model: a 100K-token report runs $0.03 (DeepSeek…
📊 Key numbers · Self-hosting adds $0 per-token markup — but the model still dominates the bill: a 100K-token report ranges $0.03–$3.01 (a ~106× spread…
📊 关键数字 · new-api(⭐ 43.6k)已超越 one-api(⭐ 36k) 成为最活跃的国内中转/分发面板(LiteLLM ⭐ 54.9k)。⚠️ 别用转售「免费 API」中转:2026 年研究在转售中转里测出 45.8% 的指纹测试模型身份不符、输出偏差最高 47%(Real Money…