Provider Operations
这里展示 Router 的 Provider key、adapter blocker 和 gateway smoke readiness。AgentCosts 仍然坚持 observe-first：先记录真实调用和成本报告，再逐步开放自动路由。
当前接入规则
AgentCosts manages provider integrations as a non-secret catalog plus server-only environment-variable readiness. During beta, the Router observes and reports usage first; proxy routing should graduate provider by provider after cost reports and budget alerts are trusted.
cheap / openai-compatible
Run a small observe-first gateway smoke test for DeepSeek R1, then review D1 event quality before expanding traffic.
已追踪模型
deepseek-v4-flash
Input $0.14/1M, cached input $0.0028/1M, output $0.28/1M
deepseek-v4-pro
Input $0.435/1M, cached input $0.003625/1M, output $0.87/1M
Env: DEEPSEEK_API_KEY
Provider docs ->cheap / openai-compatible
Run a small observe-first gateway smoke test for DeepSeek V3.2, then review D1 event quality before expanding traffic.
已追踪模型
deepseek-v4-flash
Input $0.14/1M, cached input $0.0028/1M, output $0.28/1M
deepseek-v4-pro
Input $0.435/1M, cached input $0.003625/1M, output $0.87/1M
Env: DEEPSEEK_API_KEY
Provider docs ->cheap / openai-compatible
Add MOONSHOT_API_KEY in Vercel Production before running Moonshot Kimi gateway smoke tests.
已追踪模型
kimi-k2.6
Input $0.95/1M, cached input $0.16/1M, output $4/1M
kimi-k2.5
Input $0.6/1M, cached input $0.1/1M, output $3/1M
Env: MOONSHOT_API_KEY
Provider docs ->cheap / openai-compatible
Add ZAI_API_KEY in Vercel Production before running Z.AI GLM gateway smoke tests.
已追踪模型
glm-5.1
Input $1.4/1M, cached input $0.26/1M, output $4.4/1M
glm-5
Input $1/1M, cached input $0.2/1M, output $3.2/1M
glm-4.7-flashx
Input $0.07/1M, cached input $0.01/1M, output $0.4/1M
Env: ZAI_API_KEY
Provider docs ->fast / openai-compatible
Add GROQ_API_KEY in Vercel Production before running Groq gateway smoke tests.
已追踪模型
gpt-oss-20b
Input $0.075/1M, cached input $0.0375/1M, output $0.3/1M
llama-4-scout
Input $0.11/1M, output $0.34/1M
Env: GROQ_API_KEY
Provider docs ->fast / openai-compatible
Add MINIMAX_API_KEY in Vercel Production before running MiniMax gateway smoke tests.
已追踪模型
MiniMax-M2.7
Input $0.3/1M, cached input $0.06/1M, output $1.2/1M
MiniMax-M2.7-highspeed
Input $0.6/1M, cached input $0.06/1M, output $2.4/1M
MiniMax-M2.5
Input $0.3/1M, cached input $0.03/1M, output $1.2/1M
Env: MINIMAX_API_KEY
Provider docs ->premium / openai-compatible
Add OPENAI_API_KEY in Vercel Production before running DALL-E 3 gateway smoke tests.
已追踪模型
gpt-5.5
Input $5/1M, cached input $0.5/1M, output $30/1M
gpt-5.5-pro
Input $30/1M, output $180/1M
gpt-5.4
Input $2.5/1M, cached input $0.25/1M, output $15/1M
另有 3 个模型见 provider JSON
Env: OPENAI_API_KEY
Provider docs ->premium / openai-compatible
Add OPENAI_API_KEY in Vercel Production before running OpenAI o1 gateway smoke tests.
已追踪模型
gpt-5.5
Input $5/1M, cached input $0.5/1M, output $30/1M
gpt-5.5-pro
Input $30/1M, output $180/1M
gpt-5.4
Input $2.5/1M, cached input $0.25/1M, output $15/1M
另有 3 个模型见 provider JSON
Env: OPENAI_API_KEY
Provider docs ->fallback / aggregator
Add OPENROUTER_API_KEY in Vercel Production before running OpenRouter gateway smoke tests.
已追踪模型
deepseek/deepseek-chat
Provider pricing
Env: OPENROUTER_API_KEY
Provider docs ->premium / native-adapter
Defer adapter work until observe-first OpenAI-compatible providers prove the savings report loop.
已追踪模型
claude-opus-4.7
Input $5/1M, cached input $0.5/1M, output $25/1M
claude-opus-4.6
Input $5/1M, cached input $0.5/1M, output $25/1M
claude-sonnet-4.6
Input $3/1M, cached input $0.3/1M, output $15/1M
另有 1 个模型见 provider JSON
Env: ANTHROPIC_API_KEY
Provider docs ->