diff --git a/NEWS.md b/NEWS.md
index 0b5a84b216..53acf6e84c 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -67,6 +67,8 @@ rowwiseDT(
 
 10. Fixed possible segfault in `setDT(df); attr(df, key) <- value; set(df, ...)`, i.e. adding columns to an object with `set()` that was converted to data.table with `setDT()` and later had attributes add with `attr<-`, [#6410](https://github.com/Rdatatable/data.table/issues/6410). Thanks to @hongyuanjia for the report and @ben-schwen for the PR. Note that `setattr()` should be preferred for adding attributes to a data.table.
 
+11. `fread()` automatically detects timestamps with sub-second accuracy again, [#6440](https://github.com/Rdatatable/data.table/issues/6440). This was a regression due to interference with new `dec='auto'` support. Thanks @kav2k for the concise report and @MichaelChirico for the fix.
+
 ## NOTES
 
 1. Tests run again when some Suggests packages are missing, [#6411](https://github.com/Rdatatable/data.table/issues/6411). Thanks @aadler for the note and @MichaelChirico for the fix.
diff --git a/inst/tests/tests.Rraw b/inst/tests/tests.Rraw
index 3647e61f9a..b8e28c5cf3 100644
--- a/inst/tests/tests.Rraw
+++ b/inst/tests/tests.Rraw
@@ -18549,21 +18549,38 @@ test(2255, as.data.table(DF), output="DF1.V1.*DF1.V2.*DF2.V3.*DF2.V4.*V5")
 DT = data.table(a = letters, b = 1:26/6, c = 1:26)
 ## auto-detect dec=','
 fwrite(DT, f <- tempfile(), dec=',', sep=';')
-test(2256.1, fread(f), DT)
+test(2256.01, fread(f), DT)
 
 fwrite(DT, f, dec=',', sep='|')
-test(2256.2, fread(f), DT)
+test(2256.02, fread(f), DT)
 
 ## auto-detect dec='.'
 fwrite(DT, f)
-test(2256.3, fread(f), DT)
+test(2256.03, fread(f), DT)
 
 ## verbose output
-test(2256.4, fread(f, verbose=TRUE), DT, output="sep=',' so dec set to '.'")
+test(2256.04, fread(f, verbose=TRUE), DT, output="sep=',' so dec set to '.'")
 
 fwrite(DT, f, dec=',', sep=';')
-test(2256.5, fread(f, verbose=TRUE), DT, output="dec=',' detected based on a balance of 18")
-test(2256.6, fread('a;b\n1,14;5', verbose=TRUE), data.table(a=1.14, b=5L), output="dec=',' detected based on a balance of 1 ")
+test(2256.05, fread(f, verbose=TRUE), DT, output="dec=',' detected based on a balance of 18")
+test(2256.06, fread('a;b\n1,14;5', verbose=TRUE), data.table(a=1.14, b=5L), output="dec=',' detected based on a balance of 1 ")
+
+## timestamps with subsecond accuracy thrown off by auto-dec #6440
+test(2256.07, fread(text="t\n2023-10-12T06:53:53.123Z"), data.table(t=as.POSIXct('2023-10-12 06:53:53.123', tz='UTC')))
+### TODO(#6447): sep="\t" shouldn't be needed here, right?
+test(2256.08, fread(text="t\n2023-10-12T06:53:53,123Z", sep="\t"), data.table(t=as.POSIXct('2023-10-12 06:53:53.123', tz='UTC')))
+test(2256.09, fread(text="x,t\n1.0,2023-10-12T06:53:53.123Z"), data.table(x=1.0, t=as.POSIXct('2023-10-12 06:53:53.123', tz='UTC')))
+test(2256.10, fread(text="t,x\n2023-10-12T06:53:53.123Z,1.0"), data.table(t=as.POSIXct('2023-10-12 06:53:53.123', tz='UTC'), x=1.0))
+### from PR comment
+s = 'CoilID;AntennaID;Time;TagID;Pen;Side;Position;Location;Coil_Y;Coil_X
+1;16403160;2023-10-12T10:30:55.270Z;DA2C6411;1;AKB;Litter central;middle;1;6
+3;16403160;2023-10-12T10:30:55.270Z;DA459D86;1;AKB;Litter central;middle;1;4
+15;16402963;2023-10-12T10:31:00.900Z;DA459D86;1;AKB;Litter central;right;2;3
+14;16402963;2023-10-12T10:31:02.240Z;DA2C6411;1;AKB;Litter central;right;2;1
+11;16403160;2023-10-12T10:31:02.650Z;DA2C6411;1;AKB;Litter central;middle;2;6'
+test(2256.11,
+     unname(sapply(fread(s, sep=';'), function(x) class(x)[1L])),
+     c("integer", "integer", "POSIXct", "character", "integer", "character", "character", "character", "integer", "integer"))
 
 # helpful error about deleting during grouping, #1873
 DT = data.table(id = c(1, 1, 2, 2), a = 1:4, b = 5:8)
diff --git a/src/fread.c b/src/fread.c
index 5e86f1ab56..59dfbd23e1 100644
--- a/src/fread.c
+++ b/src/fread.c
@@ -1072,7 +1072,6 @@ static void parse_iso8601_timestamp(FieldParseContext *ctx)
 
   date_only:
 
-  //Rprintf("date=%d\thour=%d\tz_hour=%d\tminute=%d\ttz_minute=%d\tsecond=%.1f\n", date, hour, tz_hour, minute, tz_minute, second);
   // cast upfront needed to prevent silent overflow
   *target = 86400*(double)date + 3600*(hour - tz_hour) + 60*(minute - tz_minute) + second;
 
@@ -1233,9 +1232,13 @@ static int detect_types( const char **pch, int8_t type[], int ncol, bool *bumped
         }
       }
       ch = fieldStart;
-      if (autoDec && IS_DEC_TYPE(tmpType[field]) && dec == '.') { // . didn't parse a double; try ,
-        dec = ',';
-        continue;
+      if (autoDec && IS_DEC_TYPE(tmpType[field])) {
+        if (dec == '.') { // '.' didn't parse a double; try ','
+          dec = ',';
+          continue; // i.e., restart since tmpType not incremented
+        } else if (dec == ',') { // ',' didn't parse a double either; reset
+          dec = '\0';
+        }
       }
       while (++tmpType[field]<CT_STRING && disabled_parsers[tmpType[field]]) {};
       *bumped = true;
diff --git a/src/fread.h b/src/fread.h
index 89dea25926..060b8006f9 100644
--- a/src/fread.h
+++ b/src/fread.h
@@ -36,7 +36,7 @@ typedef enum {
   NUMTYPE          // placeholder for the number of types including drop; used for allocation and loop bounds
 } colType;
 
-#define IS_DEC_TYPE(x) ((x) == CT_FLOAT64 || (x) == CT_FLOAT64_EXT) // types where dec matters
+#define IS_DEC_TYPE(x) ((x) == CT_FLOAT64 || (x) == CT_FLOAT64_EXT || (x) == CT_ISO8601_TIME) // types where dec matters
 
 extern int8_t typeSize[NUMTYPE];
 extern const char typeName[NUMTYPE][10];