wip: zen

4 files changed, 191 insertions, 62 deletions

diff --git a/cloud/app/src/routes/zen/v1/chat/completions.ts b/cloud/app/src/routes/zen/v1/chat/completions.ts
index 0a3b0e759..b68264cc4 100644
--- a/cloud/app/src/routes/zen/v1/chat/completions.ts
+++ b/cloud/app/src/routes/zen/v1/chat/completions.ts
@@ -2,14 +2,19 @@ import type { APIEvent } from "@solidjs/start/server"
 import { handler } from "~/util/zen"
 
 export function POST(input: APIEvent) {
+  let usage: any
   return handler(input, {
-    transformBody: (body: any) => ({
+    modifyBody: (body: any) => ({
       ...body,
       stream_options: {
         include_usage: true,
       },
     }),
-    parseUsageChunk: (chunk: string) => {
+    setAuthHeader: (headers: Headers, apiKey: string) => {
+      headers.set("authorization", `Bearer ${apiKey}`)
+    },
+    parseApiKey: (headers: Headers) => headers.get("authorization")?.split(" ")[1],
+    onStreamPart: (chunk: string) => {
       if (!chunk.startsWith("data: ")) return
 
       let json
@@ -19,15 +24,15 @@ export function POST(input: APIEvent) {
         return
       }
 
-      return json.usage
+      if (!json.usage) return
+      usage = json.usage
     },
-    buildUsage: (usage: any) => ({
+    getStreamUsage: () => usage,
+    normalizeUsage: (usage: any) => ({
       inputTokens: usage.prompt_tokens ?? 0,
       outputTokens: usage.completion_tokens ?? 0,
       reasoningTokens: usage.completion_tokens_details?.reasoning_tokens ?? 0,
       cacheReadTokens: usage.prompt_tokens_details?.cached_tokens ?? 0,
-      //cacheWriteTokens = usage.providerMetadata?.["anthropic"]?.["cacheCreationInputTokens"] ?? 0
-      cacheWriteTokens: 0,
     }),
   })
 }
diff --git a/cloud/app/src/routes/zen/v1/messages.ts b/cloud/app/src/routes/zen/v1/messages.ts
new file mode 100644
index 000000000..ff399b034
--- /dev/null
+++ b/cloud/app/src/routes/zen/v1/messages.ts
@@ -0,0 +1,61 @@
+import type { APIEvent } from "@solidjs/start/server"
+import { handler } from "~/util/zen"
+
+type Usage = {
+  cache_creation?: {
+    ephemeral_5m_input_tokens?: number
+    ephemeral_1h_input_tokens?: number
+  }
+  cache_creation_input_tokens?: number
+  cache_read_input_tokens?: number
+  input_tokens?: number
+  output_tokens?: number
+  server_tool_use?: {
+    web_search_requests?: number
+  }
+}
+
+export function POST(input: APIEvent) {
+  let usage: Usage
+  return handler(input, {
+    modifyBody: (body: any) => ({
+      ...body,
+      service_tier: "standard_only",
+    }),
+    setAuthHeader: (headers: Headers, apiKey: string) => headers.set("x-api-key", apiKey),
+    parseApiKey: (headers: Headers) => headers.get("x-api-key") ?? undefined,
+    onStreamPart: (chunk: string) => {
+      const data = chunk.split("\n")[1]
+      if (!data.startsWith("data: ")) return
+
+      let json
+      try {
+        json = JSON.parse(data.slice(6)) as { usage?: Usage }
+      } catch (e) {
+        return
+      }
+
+      if (!json.usage) return
+      usage = {
+        ...usage,
+        ...json.usage,
+        cache_creation: {
+          ...usage?.cache_creation,
+          ...json.usage.cache_creation,
+        },
+        server_tool_use: {
+          ...usage?.server_tool_use,
+          ...json.usage.server_tool_use,
+        },
+      }
+    },
+    getStreamUsage: () => usage,
+    normalizeUsage: (usage: Usage) => ({
+      inputTokens: usage.input_tokens ?? 0,
+      outputTokens: usage.output_tokens ?? 0,
+      cacheReadTokens: usage.cache_read_input_tokens ?? 0,
+      cacheWrite5mTokens: usage.cache_creation?.ephemeral_5m_input_tokens,
+      cacheWrite1hTokens: usage.cache_creation?.ephemeral_1h_input_tokens,
+    }),
+  })
+}
diff --git a/cloud/app/src/routes/zen/v1/responses.ts b/cloud/app/src/routes/zen/v1/responses.ts
index 844b4fef8..518a431ab 100644
--- a/cloud/app/src/routes/zen/v1/responses.ts
+++ b/cloud/app/src/routes/zen/v1/responses.ts
@@ -2,8 +2,13 @@ import type { APIEvent } from "@solidjs/start/server"
 import { handler } from "~/util/zen"
 
 export function POST(input: APIEvent) {
+  let usage: any
   return handler(input, {
-    parseUsageChunk: (chunk: string) => {
+    setAuthHeader: (headers: Headers, apiKey: string) => {
+      headers.set("authorization", `Bearer ${apiKey}`)
+    },
+    parseApiKey: (headers: Headers) => headers.get("authorization")?.split(" ")[1],
+    onStreamPart: (chunk: string) => {
       const [event, data] = chunk.split("\n")
       if (event !== "event: response.completed") return
       if (!data.startsWith("data: ")) return
@@ -15,9 +20,11 @@ export function POST(input: APIEvent) {
         return
       }
 
-      return json.response?.usage
+      if (!json.response?.usage) return
+      usage = json.response.usage
     },
-    buildUsage: (usage: any) => {
+    getStreamUsage: () => usage,
+    normalizeUsage: (usage: any) => {
       const inputTokens = usage.input_tokens ?? 0
       const outputTokens = usage.output_tokens ?? 0
       const reasoningTokens = usage.output_tokens_details?.reasoning_tokens ?? 0
@@ -27,7 +34,6 @@ export function POST(input: APIEvent) {
         outputTokens: outputTokens - reasoningTokens,
         reasoningTokens,
         cacheReadTokens,
-        cacheWriteTokens: 0,
       }
     },
   })
diff --git a/cloud/app/src/util/zen.ts b/cloud/app/src/util/zen.ts
index edaac1441..9ccf9e567 100644
--- a/cloud/app/src/util/zen.ts
+++ b/cloud/app/src/util/zen.ts
@@ -9,14 +9,18 @@ import { Resource } from "@opencode/cloud-resource"
 export async function handler(
   input: APIEvent,
   opts: {
-    transformBody?: (body: any) => any
-    parseUsageChunk: (chunk: string) => string | undefined
-    buildUsage: (body: any) => {
+    modifyBody?: (body: any) => any
+    setAuthHeader: (headers: Headers, apiKey: string) => void
+    parseApiKey: (headers: Headers) => string | undefined
+    onStreamPart: (chunk: string) => void
+    getStreamUsage: () => any
+    normalizeUsage: (body: any) => {
       inputTokens: number
       outputTokens: number
-      reasoningTokens: number
+      reasoningTokens?: number
       cacheReadTokens: number
-      cacheWriteTokens: number
+      cacheWrite5mTokens?: number
+      cacheWrite1hTokens?: number
     }
   },
 ) {
@@ -25,20 +29,63 @@ export async function handler(
   class ModelError extends Error {}
 
   const MODELS = {
-    //  "anthropic/claude-sonnet-4": {
-    //    auth: true,
-    //    api: "https://api.anthropic.com",
-    //    apiKey: Resource.ANTHROPIC_API_KEY.value,
-    //    model: "claude-sonnet-4-20250514",
-    //    cost: {
-    //      input: 0.0000015,
-    //      output: 0.000006,
-    //      reasoning: 0.0000015,
-    //      cacheRead: 0.0000001,
-    //      cacheWrite: 0.0000001,
-    //    },
-    //    headerMappings: {},
-    //  },
+    "claude-opus-4-1": {
+      id: "claude-opus-4-1" as const,
+      auth: true,
+      api: "https://api.anthropic.com",
+      apiKey: Resource.ANTHROPIC_API_KEY.value,
+      model: "claude-opus-4-1-20250805",
+      cost: {
+        input: 0.000015,
+        output: 0.000075,
+        cacheRead: 0.0000015,
+        cacheWrite5m: 0.00001875,
+        cacheWrite1h: 0.00003,
+      },
+      headerMappings: {},
+    },
+    "claude-sonnet-4": {
+      id: "claude-sonnet-4" as const,
+      auth: true,
+      api: "https://api.anthropic.com",
+      apiKey: Resource.ANTHROPIC_API_KEY.value,
+      model: "claude-sonnet-4-20250514",
+      cost: (usage: any) => {
+        const totalInputTokens =
+          usage.inputTokens + usage.cacheReadTokens + usage.cacheWrite5mTokens + usage.cacheWrite1hTokens
+        return totalInputTokens <= 200_000
+          ? {
+              input: 0.000003,
+              output: 0.000015,
+              cacheRead: 0.0000003,
+              cacheWrite5m: 0.00000375,
+              cacheWrite1h: 0.000006,
+            }
+          : {
+              input: 0.000006,
+              output: 0.0000225,
+              cacheRead: 0.0000006,
+              cacheWrite5m: 0.0000075,
+              cacheWrite1h: 0.000012,
+            }
+      },
+      headerMappings: {},
+    },
+    "claude-3-5-haiku": {
+      id: "claude-3-5-haiku" as const,
+      auth: true,
+      api: "https://api.anthropic.com",
+      apiKey: Resource.ANTHROPIC_API_KEY.value,
+      model: "claude-3-5-haiku-20241022",
+      cost: {
+        input: 0.0000008,
+        output: 0.000004,
+        cacheRead: 0.00000008,
+        cacheWrite5m: 0.000001,
+        cacheWrite1h: 0.0000016,
+      },
+      headerMappings: {},
+    },
     "gpt-5": {
       id: "gpt-5" as const,
       auth: true,
@@ -48,9 +95,9 @@ export async function handler(
       cost: {
         input: 0.00000125,
         output: 0.00001,
-        reasoning: 0.00001,
         cacheRead: 0.000000125,
-        cacheWrite: 0,
+        cacheWrite5m: 0,
+        cacheWrite1h: 0,
       },
       headerMappings: {},
     },
@@ -63,9 +110,9 @@ export async function handler(
       cost: {
         input: 0.00000038,
         output: 0.00000153,
-        reasoning: 0,
         cacheRead: 0,
-        cacheWrite: 0,
+        cacheWrite5m: 0,
+        cacheWrite1h: 0,
       },
       headerMappings: {},
     },
@@ -78,9 +125,9 @@ export async function handler(
       cost: {
         input: 0.0000006,
         output: 0.0000025,
-        reasoning: 0,
         cacheRead: 0,
-        cacheWrite: 0,
+        cacheWrite5m: 0,
+        cacheWrite1h: 0,
       },
       headerMappings: {},
     },
@@ -93,9 +140,9 @@ export async function handler(
       cost: {
         input: 0,
         output: 0,
-        reasoning: 0,
         cacheRead: 0,
-        cacheWrite: 0,
+        cacheWrite5m: 0,
+        cacheWrite1h: 0,
       },
       headerMappings: {
         "x-grok-conv-id": "x-opencode-session",
@@ -112,9 +159,9 @@ export async function handler(
       cost: {
         input: 0.00000038,
         output: 0.00000153,
-        reasoning: 0,
         cacheRead: 0,
-        cacheWrite: 0,
+        cacheWrite5m: 0,
+        cacheWrite1h: 0,
       },
       headerMappings: {},
     },
@@ -156,14 +203,14 @@ export async function handler(
         const headers = input.request.headers
         headers.delete("host")
         headers.delete("content-length")
-        headers.set("authorization", `Bearer ${MODEL.apiKey}`)
+        opts.setAuthHeader(headers, MODEL.apiKey)
         Object.entries(MODEL.headerMappings ?? {}).forEach(([k, v]) => {
           headers.set(k, headers.get(v)!)
         })
         return headers
       })(),
       body: JSON.stringify({
-        ...(opts.transformBody?.(body) ?? body),
+        ...(opts.modifyBody?.(body) ?? body),
         model: MODEL.model,
       }),
     })
@@ -199,32 +246,31 @@ export async function handler(
         let buffer = ""
         let responseLength = 0
         let startTimestamp = Date.now()
-        let receivedFirstByte = false
 
         function pump(): Promise<void> {
           return (
             reader?.read().then(async ({ done, value }) => {
               if (done) {
                 logger.metric({ response_length: responseLength })
+                const usage = opts.getStreamUsage()
+                if (usage) await trackUsage(usage)
                 c.close()
                 return
               }
 
-              if (!receivedFirstByte) {
-                receivedFirstByte = true
+              if (responseLength === 0) {
                 logger.metric({ time_to_first_byte: Date.now() - startTimestamp })
               }
+              responseLength += value.length
 
               buffer += decoder.decode(value, { stream: true })
-              responseLength += value.length
 
               const parts = buffer.split("\n\n")
               buffer = parts.pop() ?? ""
 
               for (const part of parts) {
                 logger.debug(part)
-                const usage = opts.parseUsageChunk(part.trim())
-                if (usage) await trackUsage(usage)
+                opts.onStreamPart(part.trim())
               }
 
               c.enqueue(value)
@@ -255,10 +301,9 @@ export async function handler(
 
     async function authenticate() {
       try {
-        const authHeader = input.request.headers.get("authorization")
-        if (!authHeader || !authHeader.startsWith("Bearer ")) throw new AuthError("Missing API key.")
+        const apiKey = opts.parseApiKey(input.request.headers)
+        if (!apiKey) throw new AuthError("Missing API key.")
 
-        const apiKey = authHeader.split(" ")[1]
         const key = await Database.use((tx) =>
           tx
             .select({
@@ -300,26 +345,38 @@ export async function handler(
     }
 
     async function trackUsage(usage: any) {
-      const { inputTokens, outputTokens, reasoningTokens, cacheReadTokens, cacheWriteTokens } = opts.buildUsage(usage)
-
-      const inputCost = MODEL.cost.input * inputTokens * 100
-      const outputCost = MODEL.cost.output * outputTokens * 100
-      const reasoningCost = MODEL.cost.reasoning * reasoningTokens * 100
-      const cacheReadCost = MODEL.cost.cacheRead * cacheReadTokens * 100
-      const cacheWriteCost = MODEL.cost.cacheWrite * cacheWriteTokens * 100
-      const totalCostInCent = inputCost + outputCost + reasoningCost + cacheReadCost + cacheWriteCost
+      const { inputTokens, outputTokens, reasoningTokens, cacheReadTokens, cacheWrite5mTokens, cacheWrite1hTokens } =
+        opts.normalizeUsage(usage)
+
+      const modelCost = typeof MODEL.cost === "function" ? MODEL.cost(usage) : MODEL.cost
+
+      const inputCost = modelCost.input * inputTokens * 100
+      const outputCost = modelCost.output * outputTokens * 100
+      const reasoningCost = reasoningTokens ? modelCost.output * reasoningTokens * 100 : undefined
+      const cacheReadCost = modelCost.cacheRead * cacheReadTokens * 100
+      const cacheWrite5mCost = cacheWrite5mTokens ? modelCost.cacheWrite5m * cacheWrite5mTokens * 100 : undefined
+      const cacheWrite1hCost = cacheWrite1hTokens ? modelCost.cacheWrite1h * cacheWrite1hTokens * 100 : undefined
+      const totalCostInCent =
+        inputCost +
+        outputCost +
+        (reasoningCost ?? 0) +
+        cacheReadCost +
+        (cacheWrite5mCost ?? 0) +
+        (cacheWrite1hCost ?? 0)
 
       logger.metric({
         "tokens.input": inputTokens,
         "tokens.output": outputTokens,
         "tokens.reasoning": reasoningTokens,
         "tokens.cache_read": cacheReadTokens,
-        "tokens.cache_write": cacheWriteTokens,
+        "tokens.cache_write_5m": cacheWrite5mTokens,
+        "tokens.cache_write_1h": cacheWrite1hTokens,
         "cost.input": Math.round(inputCost),
         "cost.output": Math.round(outputCost),
-        "cost.reasoning": Math.round(reasoningCost),
+        "cost.reasoning": reasoningCost ? Math.round(reasoningCost) : undefined,
         "cost.cache_read": Math.round(cacheReadCost),
-        "cost.cache_write": Math.round(cacheWriteCost),
+        "cost.cache_write_5m": cacheWrite5mCost ? Math.round(cacheWrite5mCost) : undefined,
+        "cost.cache_write_1h": cacheWrite1hCost ? Math.round(cacheWrite1hCost) : undefined,
         "cost.total": Math.round(totalCostInCent),
       })
 
@@ -335,7 +392,7 @@ export async function handler(
           outputTokens,
           reasoningTokens,
           cacheReadTokens,
-          cacheWriteTokens,
+          cacheWriteTokens: (cacheWrite5mTokens ?? 0) + (cacheWrite1hTokens ?? 0),
           cost,
         })
         await tx