diff --git a/Data/ByteString/FastBuilder/Internal.hs b/Data/ByteString/FastBuilder/Internal.hs
--- a/Data/ByteString/FastBuilder/Internal.hs
+++ b/Data/ByteString/FastBuilder/Internal.hs
@@ -615,6 +615,8 @@
       cur <- getCur
       io $ flushQueue h queueRef cur
       io $ S.hPut h bstr
+      io $ modifyIORef' queueRef
+        $ \q -> q { queueTotal = queueTotal q + S.length bstr }
 {-# NOINLINE byteStringInsert_ #-}
 
 -- | Turn a C String into a 'Builder'. The behavior is undefined if the given
diff --git a/fast-builder.cabal b/fast-builder.cabal
--- a/fast-builder.cabal
+++ b/fast-builder.cabal
@@ -1,5 +1,5 @@
 name:                fast-builder
-version:             0.1.1.0
+version:             0.1.2.0
 synopsis:            Fast ByteString Builder
 description:         An efficient implementation of ByteString builder. It should be faster than
                      the standard implementation in most cases.
diff --git a/tests/prop.hs b/tests/prop.hs
--- a/tests/prop.hs
+++ b/tests/prop.hs
@@ -29,7 +29,8 @@
 data BuilderPrim
   = BP_Word8 !Word8
   | BP_Word64le !Word64
-  | BP_ByteString !BS.ByteString
+  | BP_ByteStringC !BS.ByteString
+  | BP_ByteStringI !BS.ByteString
   | BP_Empty
   deriving (Show)
 
@@ -46,7 +47,8 @@
   arbitrary = QC.oneof
     [ BP_Word8 <$> QC.arbitrary
     , BP_Word64le <$> QC.arbitrary
-    , BP_ByteString . BS.pack <$> QC.arbitrary
+    , BP_ByteStringC . BS.pack <$> QC.arbitrary
+    , BP_ByteStringI . BS.pack <$> QC.arbitrary
     , pure BP_Empty
     ]
 
@@ -84,7 +86,8 @@
 
     prim (BP_Word8 w) = word8 w
     prim (BP_Word64le w) = word64LE w
-    prim (BP_ByteString bs) = byteString bs
+    prim (BP_ByteStringI bs) = byteStringInsert bs
+    prim (BP_ByteStringC bs) = byteStringCopy bs
     prim BP_Empty = mempty
 
 buildWithBuilder :: Driver -> BuilderTree -> BS.ByteString
@@ -100,7 +103,8 @@
     prim (BP_Word64le w) = (list++)
       where
         list = take 8 $ map fromIntegral $ iterate (`div` 256) w
-    prim (BP_ByteString bs) = (BS.unpack bs ++)
+    prim (BP_ByteStringI bs) = (BS.unpack bs ++)
+    prim (BP_ByteStringC bs) = (BS.unpack bs ++)
     prim BP_Empty = id
 
 errorBuilder :: TestException -> Builder