diff --git a/packages/console/app/src/routes/zen/util/provider/openai.ts b/packages/console/app/src/routes/zen/util/provider/openai.ts
index cebc78a127a3..e5767b787bd3 100644
--- a/packages/console/app/src/routes/zen/util/provider/openai.ts
+++ b/packages/console/app/src/routes/zen/util/provider/openai.ts
@@ -51,7 +51,12 @@ export const openaiHelper: ProviderHelper = ({ workspaceID }) => ({
     const cacheReadTokens = usage.input_tokens_details?.cached_tokens ?? undefined
     const cacheWriteTokens = usage.input_tokens_details?.cache_write_tokens ?? undefined
     return {
-      inputTokens: inputTokens - (cacheReadTokens ?? 0),
+      // `input_tokens` is the whole prompt: cached_tokens and cache_write_tokens are
+      // both subsets of it, not addends on top of it (OpenAI's own sample: 2600 =
+      // 2000 read + 400 written + 200 neither). Subtract both so the four buckets
+      // downstream are disjoint -- handler.ts sums them for the long-context
+      // threshold and prices each one separately.
+      inputTokens: inputTokens - (cacheReadTokens ?? 0) - (cacheWriteTokens ?? 0),
       outputTokens,
       reasoningTokens,
       cacheReadTokens,
diff --git a/packages/console/app/test/providerUsage.test.ts b/packages/console/app/test/providerUsage.test.ts
index 7be3aacffddb..0f7cb50c0514 100644
--- a/packages/console/app/test/providerUsage.test.ts
+++ b/packages/console/app/test/providerUsage.test.ts
@@ -73,7 +73,7 @@ describe("provider usage extraction", () => {
     )
 
     expect(providers.openai.normalizeUsage(usageParser.retrieve())).toEqual({
-      inputTokens: 6,
+      inputTokens: 3,
       outputTokens: 2,
       reasoningTokens: undefined,
       cacheReadTokens: 4,
@@ -81,4 +81,82 @@ describe("provider usage extraction", () => {
       cacheWrite1hTokens: undefined,
     })
   })
+
+  test("OpenAI cache buckets are subsets of input_tokens, not addends", () => {
+    // The sample from OpenAI's prompt-caching guide: 2600 input tokens made up of
+    // 2000 read from cache, 400 newly written, and 200 that were neither.
+    const normalized = providers.openai.normalizeUsage({
+      input_tokens: 2600,
+      input_tokens_details: { cached_tokens: 2000, cache_write_tokens: 400 },
+      output_tokens: 10,
+    })
+
+    expect(normalized.inputTokens).toBe(200)
+
+    // The four buckets must partition the prompt exactly -- handler.ts adds them
+    // up to pick the long-context tier, and double-counting the written tokens
+    // trips that threshold early.
+    const promptTokens =
+      normalized.inputTokens +
+      (normalized.cacheReadTokens ?? 0) +
+      (normalized.cacheWrite5mTokens ?? 0) +
+      (normalized.cacheWrite1hTokens ?? 0)
+    expect(promptTokens).toBe(2600)
+  })
+
+  // The invariant above is not OpenAI-specific: handler.ts adds the same four
+  // buckets for every provider to pick the long-context tier. The wire shapes
+  // disagree about whether the cache buckets live inside the prompt count --
+  // OpenAI's input_tokens includes them, Anthropic's excludes them, Google has
+  // no cache-write bucket at all -- so the guard belongs on the normalized
+  // side, where all three must agree. Same 2,600-token prompt three ways.
+  test("every helper partitions the prompt into the same four buckets", () => {
+    const cases = [
+      {
+        name: "openai",
+        expectedInput: 200,
+        normalized: providers.openai.normalizeUsage({
+          input_tokens: 2600,
+          input_tokens_details: { cached_tokens: 2000, cache_write_tokens: 400 },
+          output_tokens: 10,
+        }),
+      },
+      {
+        name: "anthropic",
+        expectedInput: 200,
+        normalized: providers.anthropic.normalizeUsage({
+          input_tokens: 200,
+          cache_read_input_tokens: 2000,
+          cache_creation: { ephemeral_5m_input_tokens: 400 },
+          output_tokens: 10,
+        }),
+      },
+      {
+        name: "google",
+        expectedInput: 600,
+        normalized: providers.google.normalizeUsage(
+          providers.google.extractUsage({
+            usageMetadata: {
+              promptTokenCount: 2600,
+              cachedContentTokenCount: 2000,
+              candidatesTokenCount: 10,
+            },
+          }),
+        ),
+      },
+    ]
+
+    for (const { name, expectedInput, normalized } of cases) {
+      const promptTokens =
+        normalized.inputTokens +
+        (normalized.cacheReadTokens ?? 0) +
+        (normalized.cacheWrite5mTokens ?? 0) +
+        (normalized.cacheWrite1hTokens ?? 0)
+      expect({ name, inputTokens: normalized.inputTokens, promptTokens }).toEqual({
+        name,
+        inputTokens: expectedInput,
+        promptTokens: 2600,
+      })
+    }
+  })
 })