fix(core): refine small model defaults (#33926)

packages/core/src/catalog.ts

@@ -235,6 +235,11 @@ export const layer = Layer.effect(
           if (!record) return
           const provider = record.provider
 
+          // TODO: Remove these provider-specific assumptions once model syncing reliably reports available deployments.
+          if (providerID === ProviderV2.ID.azure || providerID === ProviderV2.ID.make("azure-cognitive-services")) {
+            return
+          }
+
           if (providerID === ProviderV2.ID.opencode) {
             const gpt5Nano = record.models.get(ModelV2.ID.make("gpt-5-nano"))
             if (gpt5Nano?.enabled && gpt5Nano.status === "active") return projectModel(gpt5Nano, provider)

packages/opencode/src/provider/provider.ts

@@ -1868,44 +1868,43 @@ export const layer = Layer.effect(
         }
       }
 
-      const defaultPriority = [
+      // TODO: Remove these provider-specific assumptions once model syncing reliably reports available deployments.
-        "claude-haiku-4-5",
+      if (providerID === ProviderV2.ID.azure || providerID === ProviderV2.ID.make("azure-cognitive-services")) {
-        "claude-haiku-4.5",
+        return undefined
-        "3-5-haiku",
+      }
-        "3.5-haiku",
+
-        "gemini-3-flash",
-        "gemini-2.5-flash",
-        "gpt-5-nano",
-      ]
       const priority = providerID.startsWith("opencode")
-        ? ["gpt-5-nano"]
+        ? ["gpt-nano"]
         : providerID.startsWith("github-copilot")
-          ? ["gpt-5-mini", "claude-haiku-4.5", ...defaultPriority]
+          ? ["gpt-mini", ...smallModelFamilyPriority]
-          : defaultPriority
+          : smallModelFamilyPriority
-      for (const item of priority) {
+      const models = sortBy(
+        Object.values(provider.models),
+        [(model) => model.release_date, "desc"],
+        [(model) => model.id, "desc"],
+      )
+      for (const family of priority) {
+        const candidates = models.filter((model) => model.family === family)
         if (providerID === ProviderV2.ID.amazonBedrock) {
           const crossRegionPrefixes = ["global.", "us.", "eu."]
-          const candidates = Object.keys(provider.models).filter((m) => m.includes(item))
 
-          const globalMatch = candidates.find((m) => m.startsWith("global."))
+          const globalMatch = candidates.find((model) => model.id.startsWith("global."))
-          if (globalMatch) return provider.models[globalMatch]
+          if (globalMatch) return globalMatch
 
           const region = provider.options?.region
           if (region) {
             const regionPrefix = region.split("-")[0]
             if (regionPrefix === "us" || regionPrefix === "eu") {
-              const regionalMatch = candidates.find((m) => m.startsWith(`${regionPrefix}.`))
+              const regionalMatch = candidates.find((model) => model.id.startsWith(`${regionPrefix}.`))
-              if (regionalMatch) return provider.models[regionalMatch]
+              if (regionalMatch) return regionalMatch
             }
           }
 
-          const unprefixed = candidates.find((m) => !crossRegionPrefixes.some((p) => m.startsWith(p)))
+          const unprefixed = candidates.find((model) => !crossRegionPrefixes.some((p) => model.id.startsWith(p)))
-          if (unprefixed) return provider.models[unprefixed]
+          if (unprefixed) return unprefixed
-        } else {
+          continue
-          for (const model of Object.keys(provider.models)) {
-            if (model.includes(item)) return provider.models[model]
-          }
         }
+        if (candidates[0]) return candidates[0]
       }
 
       return undefined
@@ -1962,6 +1961,7 @@ export const defaultLayer = Layer.suspend(() =>
 )
 
 const priority = ["gpt-5", "claude-sonnet-4", "big-pickle", "gemini-3-pro"]
+const smallModelFamilyPriority = ["gemini-flash", "gpt-nano", "claude-haiku"]
 export function sort<T extends { id: string }>(models: T[]) {
   return sortBy(
     models,

packages/opencode/test/provider/provider.test.ts

@@ -652,6 +652,102 @@ it.instance("getSmallModel returns appropriate small model", () =>
   }),
 )
 
+it.instance("getSmallModel prefers Gemini for Google Vertex", () =>
+  Effect.gen(function* () {
+    yield* set("GOOGLE_VERTEX_PROJECT", "test-project")
+    const model = yield* Provider.use.getSmallModel(ProviderV2.ID.googleVertex)
+    expect(model).toBeDefined()
+    expect(model?.id).toContain("gemini")
+  }),
+)
+
+it.instance(
+  "getSmallModel selects the latest model in the preferred family",
+  Effect.gen(function* () {
+    const model = yield* Provider.use.getSmallModel(ProviderV2.ID.make("test-provider"))
+    expect(model?.id).toBe(ModelV2.ID.make("new-flash"))
+  }),
+  {
+    config: {
+      provider: {
+        "test-provider": {
+          name: "Test Provider",
+          npm: "@ai-sdk/openai-compatible",
+          models: {
+            "old-flash": { family: "gemini-flash", release_date: "2025-01-01" },
+            "new-flash": { family: "gemini-flash", release_date: "2026-01-01" },
+            "newer-haiku": { family: "claude-haiku", release_date: "2026-06-01" },
+          },
+          options: { apiKey: "test-key" },
+        },
+      },
+    },
+  },
+)
+
+it.instance(
+  "getSmallModel matches exact model families",
+  Effect.gen(function* () {
+    const model = yield* Provider.use.getSmallModel(ProviderV2.ID.make("test-provider"))
+    expect(model?.id).toBe(ModelV2.ID.make("claude-haiku"))
+  }),
+  {
+    config: {
+      provider: {
+        "test-provider": {
+          name: "Test Provider",
+          npm: "@ai-sdk/openai-compatible",
+          models: {
+            "glm-flash": { family: "glm-flash", release_date: "2026-06-01" },
+            "claude-haiku": { family: "claude-haiku", release_date: "2026-01-01" },
+          },
+          options: { apiKey: "test-key" },
+        },
+      },
+    },
+  },
+)
+
+it.instance(
+  "getSmallModel ignores model IDs without family metadata",
+  Effect.gen(function* () {
+    const model = yield* Provider.use.getSmallModel(ProviderV2.ID.make("test-provider"))
+    expect(model).toBeUndefined()
+  }),
+  {
+    config: {
+      provider: {
+        "test-provider": {
+          name: "Test Provider",
+          npm: "@ai-sdk/openai-compatible",
+          models: {
+            "gpt-5-nano": { release_date: "2026-01-01" },
+          },
+          options: { apiKey: "test-key" },
+        },
+      },
+    },
+  },
+)
+
+it.instance("getSmallModel skips inferred models for Azure", () =>
+  Effect.gen(function* () {
+    yield* set("AZURE_RESOURCE_NAME", "test-resource")
+    yield* set("AZURE_API_KEY", "test-key")
+    const model = yield* Provider.use.getSmallModel(ProviderV2.ID.azure)
+    expect(model).toBeUndefined()
+  }),
+)
+
+it.instance("getSmallModel skips inferred models for Azure Cognitive Services", () =>
+  Effect.gen(function* () {
+    yield* set("AZURE_COGNITIVE_SERVICES_RESOURCE_NAME", "test-resource")
+    yield* set("AZURE_COGNITIVE_SERVICES_API_KEY", "test-key")
+    const model = yield* Provider.use.getSmallModel(ProviderV2.ID.make("azure-cognitive-services"))
+    expect(model).toBeUndefined()
+  }),
+)
+
 it.instance(
   "getSmallModel respects config small_model override",
   Effect.gen(function* () {