fix(cloudflare): correct overstated context windows

packages/core/src/sync/providers/cloudflare-workers-ai.ts

@@ -18,6 +18,13 @@ const ENABLE_THINKING_MODELS = new Set([
   "@cf/zai-org/glm-4.7-flash",
   "@cf/zai-org/glm-5.2",
 ]);
+// Search and the model pages report 1,310,720 for these, but the serving backend rejects any request over 1,048,576
+// tokens. The sync re-derives context from search, so the served window is pinned here. It only ever lowers a window.
+const SERVED_CONTEXT = new Map([
+  ["@cf/zai-org/glm-5.3", 1_048_576],
+  ["@cf/zai-org/glm-5.3-flash", 1_048_576],
+  ["@cf/deepseek-ai/deepseek-v4-flash-0731", 1_048_576],
+]);
 const ROOT_DIR = path.join(import.meta.dirname, "..", "..", "..", "..", "..");
 const MODELS_DIR = path.join(ROOT_DIR, "models");
 const metadataFilesByPublisher = new Map<string, string[]>();
@@ -172,8 +179,10 @@ export function buildWorkersAiModel(
   const reasoning = reasoningOptions !== undefined
     ? true
     : existing?.reasoning ?? hasReasoning(model);
+  const context = Math.min(model.context_length, SERVED_CONTEXT.get(model.id) ?? model.context_length);
   const source = {
     ...model,
+    context_length: context,
     reasoning: undefined,
     supported_parameters: [
       ...model.supported_parameters.filter((parameter) => !["reasoning", "include_reasoning"].includes(parameter)),
@@ -182,6 +191,7 @@ export function buildWorkersAiModel(
     name: existing?.name ?? model.name,
     top_provider: {
       ...model.top_provider,
+      context_length: context,
       max_completion_tokens: existing?.limit?.output ?? model.top_provider.max_completion_tokens,
     },
   };

packages/core/test/cloudflare-workers-ai.test.ts

@@ -54,3 +54,43 @@ test("preserves OpenRouter pricing overrides as context tiers", () => {
     }],
   });
 });
+
+function searchRow(id: string, context: number) {
+  return {
+    id: `workers-ai/${id}`,
+    name: "Served Window",
+    created: 1_782_777_600,
+    hugging_face_id: null,
+    knowledge_cutoff: null,
+    context_length: context,
+    architecture: {
+      input_modalities: ["text"],
+      output_modalities: ["text"],
+    },
+    pricing: { prompt: "0.000001", completion: "0.000002" },
+    top_provider: {
+      context_length: context,
+      max_completion_tokens: context,
+    },
+    supported_parameters: [],
+  };
+}
+
+function translatedContext(id: string, context: number) {
+  const [source] = cloudflareWorkersAi.parseModels({ result: { data: [searchRow(id, context)] } });
+  return cloudflareWorkersAi.translateModel(source!, {
+    existing: () => undefined,
+    authored: () => undefined,
+  }).model.limit?.context;
+}
+
+test("pins the served context window when search reports a larger one", () => {
+  expect(translatedContext("@cf/zai-org/glm-5.3", 1_310_720)).toBe(1_048_576);
+  expect(translatedContext("@cf/zai-org/glm-5.3-flash", 1_310_720)).toBe(1_048_576);
+  expect(translatedContext("@cf/deepseek-ai/deepseek-v4-flash-0731", 1_310_720)).toBe(1_048_576);
+});
+
+test("never raises a context window or pins other models", () => {
+  expect(translatedContext("@cf/zai-org/glm-5.3", 262_144)).toBe(262_144);
+  expect(translatedContext("@cf/test/unpinned", 1_310_720)).toBe(1_310_720);
+});

providers/cloudflare-ai-gateway/curation.toml

@@ -84,7 +84,6 @@ reasoning_options = [{ type = "effort", values = ["low", "medium", "high", "xhig
 
 [models."anthropic/claude-sonnet-4.5"]
 structured_output = true
-limit = { context = 1000000 }
 reasoning_options = [{ type = "budget_tokens", min = 1024 }]
 
 [models."anthropic/claude-sonnet-4.6"]

providers/cloudflare-ai-gateway/models/anthropic/claude-sonnet-4.5.toml

@@ -12,7 +12,7 @@ cache_read = 0.3
 cache_write = 3.75
 
 [limit]
-context = 1_000_000
+context = 200_000
 
 [provider]
 npm = "@ai-sdk/anthropic"

providers/cloudflare-workers-ai/models/@cf/deepseek-ai/deepseek-v4-flash-0731.toml

@@ -16,5 +16,5 @@ output = 1.32
 cache_read = 0.014
 
 [limit]
-context = 1_310_720
+context = 1_048_576
 output = 1_048_576

providers/cloudflare-workers-ai/models/@cf/zai-org/glm-5.3-flash.toml

@@ -16,7 +16,7 @@ output = 0.5
 cache_read = 0.03
 
 [limit]
-context = 1_310_720
+context = 1_048_576
 output = 1_048_576
 
 [modalities]

providers/cloudflare-workers-ai/models/@cf/zai-org/glm-5.3.toml

@@ -16,5 +16,5 @@ output = 4.4
 cache_read = 0.26
 
 [limit]
-context = 1_310_720
+context = 1_048_576
 output = 1_048_576