diff --git a/src/renderer/src/ai/providers/claude.ts b/src/renderer/src/ai/providers/claude.ts
index deed4b0..c9aa54a 100644
--- a/src/renderer/src/ai/providers/claude.ts
+++ b/src/renderer/src/ai/providers/claude.ts
@@ -87,10 +87,26 @@ function splitSystem(messages: ChatMessage[]): { system: AnthropicSystemBlock[];
   return { system, rest }
 }
 
-function parseUsage(usage: { input_tokens?: number; output_tokens?: number; cache_read_input_tokens?: number } | undefined): TokenUsage | undefined {
+/** Anthropic's `input_tokens` is deliberately ONLY the fresh, non-cached
+ *  portion of the prompt — a cache hit/write moves those tokens into
+ *  `cache_read_input_tokens`/`cache_creation_input_tokens` instead, so a
+ *  well-cached request can report `input_tokens: 21` even for a 10K-token
+ *  prompt. `TokenUsage.inputTokens` is meant to be the TOTAL (matching
+ *  OpenAI's `prompt_tokens`, which already includes its cached portion —
+ *  see providers/openai.ts) — folding all three in here is what makes the
+ *  displayed counter track real spend instead of silently undercounting
+ *  every cached request (i.e. nearly every request, once caching is on). */
+function parseUsage(usage: {
+  input_tokens?: number
+  output_tokens?: number
+  cache_read_input_tokens?: number
+  cache_creation_input_tokens?: number
+} | undefined): TokenUsage | undefined {
   if (!usage) return undefined
+  const cacheRead = usage.cache_read_input_tokens ?? 0
+  const cacheCreation = usage.cache_creation_input_tokens ?? 0
   return {
-    inputTokens: usage.input_tokens ?? 0,
+    inputTokens: (usage.input_tokens ?? 0) + cacheRead + cacheCreation,
     outputTokens: usage.output_tokens ?? 0,
     ...(usage.cache_read_input_tokens !== undefined ? { cachedInputTokens: usage.cache_read_input_tokens } : {}),
   }
@@ -141,7 +157,7 @@ async function claudeChat(req: ChatRequest, cfg: ProviderConfig): Promise<ChatRe
     model?: string
     content?: AnthropicBlock[]
     stop_reason?: string
-    usage?: { input_tokens?: number; output_tokens?: number; cache_read_input_tokens?: number }
+    usage?: { input_tokens?: number; output_tokens?: number; cache_read_input_tokens?: number; cache_creation_input_tokens?: number }
   }
   return {
     content: fromAnthropicContent(data?.content ?? []),
diff --git a/src/renderer/src/components/RadicalForgeModal.tsx b/src/renderer/src/components/RadicalForgeModal.tsx
index 8e72562..70acc7c 100644
--- a/src/renderer/src/components/RadicalForgeModal.tsx
+++ b/src/renderer/src/components/RadicalForgeModal.tsx
@@ -563,8 +563,13 @@ export function RadicalForgeModal({ open, onClose }: Props): React.ReactElement
                     </div>
                   ))}
                   <div style={{ display: 'flex', gap: 8, marginTop: 4 }}>
+                    {/* Deliberately NOT labeled "Continue" — that's the
+                        footer button's job (advancing to the next wizard
+                        step once this stage has generated). This one only
+                        confirms the answers above and reveals the Generate
+                        button for the current stage. */}
                     <button type="button" className="forge-btn forge-btn-primary" onClick={() => submitClarify(currentStage.id)}>
-                      Continue
+                      Confirm answers
                     </button>
                     <button type="button" className="forge-btn forge-btn-ghost" onClick={() => skipClarify(currentStage.id)}>
                       Skip
diff --git a/src/renderer/src/index.css b/src/renderer/src/index.css
index e362de0..f732956 100644
--- a/src/renderer/src/index.css
+++ b/src/renderer/src/index.css
@@ -3770,6 +3770,10 @@ body {
   align-items: center;
   gap: 8px;
   margin: 0 0 6px;
+  /* Reserves room for the absolutely-positioned .ai-settings-close button
+     (right: 10px, width: 28px) so a wide token count (e.g. "25.9K tokens")
+     doesn't run underneath it. */
+  padding-right: 32px;
 }
 .forge-title-icon {
   display: inline-flex;
diff --git a/tests/aiProviders.test.ts b/tests/aiProviders.test.ts
index 575f7b7..d27c217 100644
--- a/tests/aiProviders.test.ts
+++ b/tests/aiProviders.test.ts
@@ -269,7 +269,7 @@ describe('AI providers', () => {
       ])
     })
 
-    it('parses usage, including cache_read_input_tokens when present', async () => {
+    it('folds cache_read_input_tokens into the total inputTokens (Anthropic reports it separately from input_tokens, unlike OpenAI\'s prompt_tokens which already includes its cached portion) while still surfacing it separately as cachedInputTokens for display', async () => {
       installFetch({
         model: 'claude-haiku-4-5',
         content: [{ type: 'text', text: 'hi' }],
@@ -277,7 +277,18 @@ describe('AI providers', () => {
         usage: { input_tokens: 500, output_tokens: 40, cache_read_input_tokens: 300 },
       })
       const out = await claudeAdapter.chat({ model: 'm', messages: SAMPLE }, { apiKey: 'k' })
-      expect(out.usage).toEqual({ inputTokens: 500, outputTokens: 40, cachedInputTokens: 300 })
+      expect(out.usage).toEqual({ inputTokens: 800, outputTokens: 40, cachedInputTokens: 300 })
+    })
+
+    it('also folds cache_creation_input_tokens (a cache WRITE, billed at a premium, happens on essentially every first request in a session) into inputTokens', async () => {
+      installFetch({
+        model: 'claude-haiku-4-5',
+        content: [{ type: 'text', text: 'hi' }],
+        stop_reason: 'end_turn',
+        usage: { input_tokens: 21, output_tokens: 40, cache_creation_input_tokens: 1500, cache_read_input_tokens: 0 },
+      })
+      const out = await claudeAdapter.chat({ model: 'm', messages: SAMPLE }, { apiKey: 'k' })
+      expect(out.usage).toEqual({ inputTokens: 1521, outputTokens: 40, cachedInputTokens: 0 })
     })
 
     it('omits usage entirely when the response has none', async () => {
diff --git a/tests/aiRunner.test.ts b/tests/aiRunner.test.ts
index abb7f8a..782f8b8 100644
--- a/tests/aiRunner.test.ts
+++ b/tests/aiRunner.test.ts
@@ -491,12 +491,16 @@ describe('runAIPrompt — onProgress (live feed for Radical Forge)', () => {
       diagram: makeFacade(),
       onProgress: (e) => { if (e.type === 'usage') usageEvents.push(e) },
     })
-    // Running total after each round, not per-round deltas.
+    // Running total after each round, not per-round deltas. Anthropic's
+    // `input_tokens` is only the fresh, non-cached portion — the adapter
+    // folds `cache_read_input_tokens` into `inputTokens` too (see
+    // providers/claude.ts's `parseUsage`) so the total tracks real spend:
+    // round 1 is 1000 + 700 = 1700, round 2 is 1200 + 900 = 2100.
     expect(usageEvents).toEqual([
-      { type: 'usage', usage: { inputTokens: 1000, outputTokens: 50, cachedInputTokens: 700 } },
-      { type: 'usage', usage: { inputTokens: 2200, outputTokens: 70, cachedInputTokens: 1600 } },
+      { type: 'usage', usage: { inputTokens: 1700, outputTokens: 50, cachedInputTokens: 700 } },
+      { type: 'usage', usage: { inputTokens: 3800, outputTokens: 70, cachedInputTokens: 1600 } },
     ])
-    expect(result.usage).toEqual({ inputTokens: 2200, outputTokens: 70, cachedInputTokens: 1600 })
+    expect(result.usage).toEqual({ inputTokens: 3800, outputTokens: 70, cachedInputTokens: 1600 })
   })
 
   it('leaves result.usage undefined when the provider never reports usage', async () => {