diff --git a/service/relayconvert/internal/oai_chat/to_claude_messages_resp.go b/service/relayconvert/internal/oai_chat/to_claude_messages_resp.go
index 7f0c6b8be5f2..07d06a623f6d 100644
--- a/service/relayconvert/internal/oai_chat/to_claude_messages_resp.go
+++ b/service/relayconvert/internal/oai_chat/to_claude_messages_resp.go
@@ -41,11 +41,15 @@ func buildClaudeUsageFromOpenAIUsage(oaiUsage *dto.Usage) *dto.ClaudeUsage {
 	)
 	cacheCreationTokens := oaiUsage.PromptTokensDetails.CacheCreationTokensTotal()
 	inputTokens := oaiUsage.PromptTokens
-	if oaiUsage.PromptTokensDetails.CacheWriteTokens > 0 {
-		// OpenAI native cache-write usage counts cached and cache-write tokens
-		// inside prompt_tokens, while Claude semantics reports input_tokens
-		// excluding both. Both counts are unadjusted prefixes and may overlap,
-		// so clamp a negative remainder at zero.
+	// OpenAI native cache-write / cache-read usage counts both cached and
+	// cache-creation tokens inside prompt_tokens, while Claude semantics
+	// reports input_tokens excluding both.  Deduplicate so that the sum
+	// `input_tokens + cache_creation_input_tokens + cache_read_input_tokens`
+	// equals the true total context size rather than double-counting cached
+	// tokens (§ref HsMirage, issue #4395).
+	//   - CacheWriteTokens > 0: openai cache-write route (original guard)
+	//   - CachedTokens    > 0: pure cache-read (bypassed the old guard)
+	if oaiUsage.PromptTokensDetails.CacheWriteTokens > 0 || oaiUsage.PromptTokensDetails.CachedTokens > 0 {
 		inputTokens = oaiUsage.PromptTokens - oaiUsage.PromptTokensDetails.CachedTokens - cacheCreationTokens
 		if inputTokens < 0 {
 			inputTokens = 0
diff --git a/service/relayconvert/internal/oai_chat/to_claude_messages_resp_test.go b/service/relayconvert/internal/oai_chat/to_claude_messages_resp_test.go
index facd873e709f..9d0aa243ac3d 100644
--- a/service/relayconvert/internal/oai_chat/to_claude_messages_resp_test.go
+++ b/service/relayconvert/internal/oai_chat/to_claude_messages_resp_test.go
@@ -103,6 +103,31 @@ func TestBuildClaudeUsageFromOpenAICacheWriteUsage(t *testing.T) {
 	assert.Equal(t, 3616, usage.BillingUsage.OpenAIUsage.PromptTokensDetails.CacheWriteTokens)
 }
 
+func TestBuildClaudeUsageFromOpenAIPureCacheRead(t *testing.T) {
+	// HsMirage's pure cache-read reproduction (issue #4395):
+	// second identical request hits cache; CacheWriteTokens==0, CachedTokens>0.
+	usage := buildClaudeUsageFromOpenAIUsage(&dto.Usage{
+		PromptTokens:     15305,
+		CompletionTokens: 5,
+		TotalTokens:      15310,
+		PromptTokensDetails: dto.InputTokenDetails{
+			CachedTokens:     15104,
+			CacheWriteTokens: 0,
+		},
+	})
+
+	require.NotNil(t, usage)
+	// input_tokens must exclude cached: 15305 - 15104 = 201
+	assert.Equal(t, 201, usage.InputTokens)
+	assert.Equal(t, 15104, usage.CacheReadInputTokens)
+	assert.Equal(t, 0, usage.CacheCreationInputTokens)
+	assert.Equal(t, 5, usage.OutputTokens)
+	// Total context = 201 + 15104 = 15305 (not double-counted)
+	require.NotNil(t, usage.BillingUsage)
+	require.NotNil(t, usage.BillingUsage.OpenAIUsage)
+	assert.Equal(t, 15305, usage.BillingUsage.OpenAIUsage.PromptTokens)
+}
+
 func TestStreamResponseOpenAI2ClaudeClosesTextThinkingAndToolBlocks(t *testing.T) {
 	info := &relaycommon.RelayInfo{
 		ClaudeConvertInfo: &relaycommon.ClaudeConvertInfo{