diff --git a/libs/acl/openai/cache_usage_test.go b/libs/acl/openai/cache_usage_test.go
new file mode 100644
index 000000000..987ba7398
--- /dev/null
+++ b/libs/acl/openai/cache_usage_test.go
@@ -0,0 +1,85 @@
+/*
+ * Copyright 2024 CloudWeGo Authors
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package openai
+
+import (
+	"encoding/json"
+	"testing"
+
+	"github.com/cloudwego/eino/schema"
+	openai "github.com/meguminnnnnnnnn/go-openai"
+	"github.com/stretchr/testify/assert"
+)
+
+func TestToEinoTokenUsageCachedTokens(t *testing.T) {
+	tests := []struct {
+		name  string
+		usage *openai.Usage
+		want  int
+	}{
+		{
+			name: "OpenAI prompt token details",
+			usage: &openai.Usage{
+				PromptTokensDetails: &openai.PromptTokensDetails{CachedTokens: 64},
+				ExtraFields: map[string]json.RawMessage{
+					"prompt_cache_hit_tokens": json.RawMessage(`32`),
+				},
+			},
+			want: 64,
+		},
+		{
+			name: "DeepSeek top-level cache hit extension",
+			usage: &openai.Usage{
+				ExtraFields: map[string]json.RawMessage{
+					"prompt_cache_hit_tokens": json.RawMessage(`48`),
+				},
+			},
+			want: 48,
+		},
+		{
+			name: "zero standard detail falls back to DeepSeek extension",
+			usage: &openai.Usage{
+				PromptTokensDetails: &openai.PromptTokensDetails{},
+				ExtraFields: map[string]json.RawMessage{
+					"prompt_cache_hit_tokens": json.RawMessage(`32`),
+				},
+			},
+			want: 32,
+		},
+		{
+			name: "malformed extension",
+			usage: &openai.Usage{
+				ExtraFields: map[string]json.RawMessage{
+					"prompt_cache_hit_tokens": json.RawMessage(`"invalid"`),
+				},
+			},
+			want: 0,
+		},
+		{
+			name:  "missing cache details",
+			usage: &openai.Usage{},
+			want:  0,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := toEinoTokenUsage(tt.usage)
+			assert.Equal(t, schema.PromptTokenDetails{CachedTokens: tt.want}, got.PromptTokenDetails)
+		})
+	}
+}
diff --git a/libs/acl/openai/chat_model.go b/libs/acl/openai/chat_model.go
index 4b3f98f25..4219cfc7e 100644
--- a/libs/acl/openai/chat_model.go
+++ b/libs/acl/openai/chat_model.go
@@ -1335,6 +1335,19 @@ func toEinoTokenUsage(usage *openai.Usage) *schema.TokenUsage {
 	if usage.PromptTokensDetails != nil {
 		promptTokenDetails.CachedTokens = usage.PromptTokensDetails.CachedTokens
 	}
+	if promptTokenDetails.CachedTokens == 0 {
+		// DeepSeek's OpenAI-compatible API reports cache hits as a top-level
+		// usage extension instead of prompt_tokens_details.cached_tokens.
+		// go-openai preserves unknown usage fields in ExtraFields. Falling back
+		// on a zero standard value also covers gateways that synthesize an empty
+		// prompt_tokens_details object while retaining DeepSeek's extension.
+		if raw, ok := usage.ExtraFields["prompt_cache_hit_tokens"]; ok {
+			var cachedTokens int
+			if err := json.Unmarshal(raw, &cachedTokens); err == nil {
+				promptTokenDetails.CachedTokens = cachedTokens
+			}
+		}
+	}
 	completionTokensDetails := schema.CompletionTokensDetails{}
 	if usage.CompletionTokensDetails != nil {
 		completionTokensDetails.ReasoningTokens = usage.CompletionTokensDetails.ReasoningTokens