diff --git a/internal/translator/codex/openai/responses/codex_openai-responses_request.go b/internal/translator/codex/openai/responses/codex_openai-responses_request.go
index 1161c515a0..3757311858 100644
--- a/internal/translator/codex/openai/responses/codex_openai-responses_request.go
+++ b/internal/translator/codex/openai/responses/codex_openai-responses_request.go
@@ -20,12 +20,14 @@ func ConvertOpenAIResponsesRequestToCodex(modelName string, inputRawJSON []byte,
 	rawJSON, _ = sjson.SetBytes(rawJSON, "store", false)
 	rawJSON, _ = sjson.SetBytes(rawJSON, "parallel_tool_calls", true)
 	rawJSON, _ = sjson.SetBytes(rawJSON, "include", []string{"reasoning.encrypted_content"})
-	// Codex Responses rejects token limit fields, so strip them out before forwarding.
+	if v := gjson.GetBytes(rawJSON, "service_tier"); v.Exists() && v.String() == "fast" {
+		rawJSON, _ = sjson.SetBytes(rawJSON, "service_tier", "priority")
+	}
+	// Codex Responses rejects token limit and sampling fields, so strip them out before forwarding.
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "max_output_tokens")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "max_completion_tokens")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "temperature")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "top_p")
-	rawJSON, _ = sjson.DeleteBytes(rawJSON, "service_tier")
 	rawJSON, _ = sjson.DeleteBytes(rawJSON, "truncation")
 	rawJSON = applyResponsesCompactionCompatibility(rawJSON)